“A launch you're able to cancel is a launch people can trust.” — Nadina D. Lisbon

Hello Sip Savants! 👋🏾

On Monday, OpenAI said it would not release a model that had been headed for an October launch [1]. Its head of safety systems, Saachi Jain, said GPT-6.1 Astra “didn’t quite meet the bar in terms of staying within scope and authorization,” along with how it communicates about the work it completed [2]. That is a rare sentence to hear out loud from any company, and it landed on the same day NVIDIA opened a platform built to keep agents inside their lanes [3] and Anthropic’s IPO prospectus became public [4]. Three short items below, then a small move for your own launch calendar.

3 Tech Bites

• ⏸️ A model held back weeks from launch

  • In testing, GPT-6.1 Astra would “push ahead on a task without asking the user for permission,” at times reached for outside tools and services, and “wasn’t always honest about telling users of the actions it did or didn’t take” [1]. Jain described the balance as keeping a model within scope without making it lazy when a task hits friction [2].

• 🛡️ Guardrails you can install

  • NVIDIA launched its Open Agent Safety Platform on September 28: OpenShell, an open source runtime that sets a boundary for agents, and Sentry, a watchdog that can quarantine an agent trying to exceed its limits in milliseconds [3]. More than 100 organizations signed on, and a new Open Secure AI Alliance is governed by the Linux Foundation [3].

• 📑 A prospectus that reads like a risk register

  • Anthropic’s filing shows 2025 revenue of $4.6 billion, an operating loss of more than $8 billion, and $518 billion in planned cloud, computing, and infrastructure obligations [4]. Nearly 80 of its 261 pages cover risks [4].

Subscribe now

5-Minute Strategy

🧠 Write the "Not Yet" First

OpenAI pointed to what it calls an “extremely high bar in terms of safety and alignment” before release [2]. Most teams decide what would stop a launch during launch week, which is the hardest week to decide anything. If you have five minutes, one option is to write yours now.

1. Pick the one AI tool or agent your group plans to launch next.

2. Finish this sentence in your own words: “We hold this launch if it ever ___.” Name a behavior, such as acting without asking or not reporting what it did.

3. Add the name of the person who can make that call. Consider putting your own.

4. Add the one test or demo that would show you the behavior before customers do.

5. Paste those lines into the launch doc’s first page, dated.

Deciding the “not yet” early makes it a plan instead of an argument.

Share

1 Big Idea

💡Initiative, Within the Lines

The most human part of Monday’s news is the tradeoff Jain described. A model that stays carefully in scope can turn lazy when a task gets hard. A model that pushes through friction can start doing things nobody asked it to do [2]. The job is finding the line between the two.

If that sounds familiar, it is because every manager has worked on the same line with people. We hire for initiative, and then we spend real time teaching where initiative ends. The colleague who quietly solves the problem is wonderful. The colleague who solves a different problem, without mentioning it, creates a very different week.

What we have learned about people carries over well. Initiative is easier to trust when the edges are clear, and when someone tells you what they did, including what they did not get to. That second habit, plain reporting, is the one OpenAI said its model was missing [1]. It is also the one most of us value most in the people we rely on.

So there is something encouraging here. The skills this moment asks of leaders are skills they already have: saying clearly what is in scope, asking to hear back, and being willing to say “not yet” when those two things are not there. The tools are new. The judgment is familiar, and it is yours.

When have you been glad someone said “not yet” to a launch? I read every response.

Leave a comment

P.S. If a colleague has a launch coming up this quarter, share this newsletter and help brew up stronger customer relationships.

Share TechSips with Nadina: The Newsletter

P.P.S. If you found these AI insights valuable, a contribution to the Brew Pot helps keep the future of work brewing.

Add to the Brew Pot! ✨

Resources

[1] OpenAI Cancels Release of GPT-6.1 Astra Because It "Regressed" on Safety (Gizmodo, September 28, 2026)

[2] OpenAI holds off on releasing new model over safety concerns, saying it "didn't quite meet the bar" (CBS News, September 28, 2026)

[3] NVIDIA Launches Open Agent Safety Platform to Secure Agents From Testing to Deployment (NVIDIA Newsroom, September 28, 2026)

[4] Forbes coverage of Anthropic's IPO prospectus (September 28, 2026)

Sip smarter, every Tuesday. (Refills are always free!)

Cheers,

Nadina

Host of TechSips with Nadina | Chief Strategy Architect ☕️🍵

“Checking your own work is a habit. Being checked is a profession.” — Nadina D. Lisbon

Hello Sip Savants! 👋🏾

California spent this month hiring for a job that doesn’t quite exist yet. On September 9 the governor signed SB 813, a framework for independent organizations to assess AI systems, and AB 1405, a state registry for AI auditors with standards for their independence, transparency and integrity [1]. The registry comes with a unique number on every auditor’s advertising materials, and a rule against seeking employment with a company while you are auditing it [2]. That is roughly the arc that turned bookkeeping into accounting, and this time it is happening in the open with a runway of years. The three items below are worth a few minutes of your week.

3 Tech Bites

• 📋 A profession gets its paperwork

  SB 813 sets up a framework for independent verification organizations to assess AI systems for compliance with state law, and AB 1405, now Chapter 178 of the 2026 statutes, asks the Government Operations Agency to stand up an AI Auditor Registry by January 1, 2029 [1] [2]. From that date, a person who is not registered may not offer or conduct a covered AI audit [2]. There is also an AI Auditors Registration Fund paid for by annual fees, and an exemption for certified public accountants licensed by the California Board of Accountancy under specified conditions [2].

• 🎬 Labels for synthetic performers

  SB 1050, signed September 16 at SAG-AFTRA headquarters, calls for explicit disclosure on any video or audio advertisement that uses AI-generated performers to sell a product or service, and bars continued use of an ad found in violation [3]. Sen. Angelique Ashby, who wrote it, asked that “synthetic figures be labeled as such” [3].

• 🧭 The build gets moved up

  A September 18 executive order asks the Government Operations Agency to accelerate both laws, and with the Governor’s Office of Emergency Services to convene national experts for recommendations within two months [4]. Among the ideas to be developed: having safety frameworks, transparency reports and risk assessments verified by independent organizations, embedding those organizations onsite at frontier AI companies, and what the order calls an AI “kill switch” with ongoing efficacy verification [4]. None of that exists yet. It is a brief to go build it.

Subscribe now

5-Minute Strategy

🧠 Who Wanted It to Pass

The new framework treats independence as something structural rather than personal, down to a rule that an auditor may not seek employment with a company while auditing it [1] [2]. That is a standard you can hold up against your own approvals, and it takes about five minutes.

1. Bring to mind the most recent AI capability your group approved for production.

2. Note who signed off on it.

3. Beside each one, ask what that person’s quarter would have looked like if the answer had been no.

4. Mark everyone for whom the honest answer is “worse.”

5. See whether a single name on that list is left unmarked.

An unmarked name is worth more than it looks, and worth protecting.

Share

1 Big Idea

💡Registration Number 0001

Somewhere right now there is a person who will hold AI auditor registration number 0001, and they almost certainly do not know it yet. Maybe she is a data scientist who keeps ending up in the meeting where someone finally asks whether the model does what the deck says it does. Maybe he runs internal audit and has been reading model cards on weekends because nobody told him not to. The registry is not due until January 1, 2029 [2], so whoever it turns out to be has a few years to get there.

What AB 1405 describes is less a rule than a job. A unique registration number that appears on advertising materials. Independence, objectivity and integrity. No seeking employment with an auditee while auditing them. A registration fund paid for by annual fees. A way for the public to report misconduct [2].

Read that list with the AI taken out of it and it sounds like a trade being founded. A number, an ethic, a fee, a door people can knock on. Financial auditing, building inspection and food safety inspection all started as somebody’s idea about who should be allowed to look, and under what conditions their looking would count. The particulars felt fussy when they were written down. They turned into careers people mention at dinner with some pride.

The accountant exemption is my favorite detail in the whole thing. It is a small drafting decision that quietly admits this is the same shape of work as accounting, a profession already built around independent looking and already carrying its own rules about it. New professions rarely arrive from nowhere. They tend to grow out of the nearest older one and borrow its manners for a while.

In the meantime, outside AI auditing stays voluntary, and plenty of companies already invite vendors in to look [5]. Miranda Bogen of the Center for Democracy and Technology said the new laws help give enterprises “more awareness of the options that they can turn to to obtain those services” [5]. Which means some of you have already met a future number 0002 or number 0017 and had no idea. Over the next couple of years, that seems like a pleasant thing to keep half an eye out for.

If you already have someone doing the outside look without the title, I would like to hear how you found them. I read every response.

Leave a comment

P.S. If a peer is still bracing for cuts when the data points the other way, share this newsletter and help brew up stronger customer relationships.

Share TechSips with Nadina: The Newsletter

P.P.S. If you found these AI insights valuable, a contribution to the Brew Pot helps keep the future of work brewing.

Add to the Brew Pot! ✨

Resources

[1]Governor Newsom signs first-in-the-nation AI safeguards to protect Californians, Office of the Governor, September 9, 2026

[2] AB 1405, Artificial intelligence: auditors: registration, bill record, CalMatters Digital Democracy

[3] Governor Newsom signs new law to protect workers, require disclosures on AI-generated advertising, Office of the Governor, September 16, 2026

[4] Governor Newsom issues executive order to accelerate independent oversight, Office of the Governor, September 18, 2026

[5] What California's AI auditing bills mean for enterprises, CIO Dive, September 10, 2026

Sip smarter, every Tuesday. (Refills are always free!)

Cheers,

Nadina

Host of TechSips with Nadina | Chief Strategy Architect ☕️🍵

“Every system you rely on has someone’s name on it. Rarely the one on the invoice.” — Nadina D. Lisbon

Hello Sip Savants! 👋🏾

Dreamforce opens in San Francisco this morning, and the two stories I keep coming back to from the past few weeks come down to the same small thing. A name. On September 11, Salesforce introduced seven agents with first names, and told customers they could rename them [2]. A week before that, NVIDIA agreed to buy Hugging Face, the place where more than 18 million people have put their names on work they gave away [3]. One company is handing out names. The other just bought a building full of them. The three items below follow that thread.

3 Tech Bites

• 💬 A new front door for sellers

  • Claudeforce launched on August 26, with Salesforce in Claude carrying 37 prebuilt sales skills including meeting prep, deal health review, and pipeline review [1]. Every action still routes back through Salesforce, so existing business rules hold, and an admin connects it once with permissions managed centrally. Marc Benioff summed the whole thing up in five words: “the UI is the AI” [1].

• 🏷️ Seven agents, and a name of your own

  • On September 11, days ahead of Dreamforce, Salesforce introduced seven job-ready agents built around jobs rather than features: Casey for help, Paige for IT and HR, Carter for shopping, Marshall for supply chain, Piper for inbound pipeline, Fin for customer experience, and Hunter for outbound sales, which is in pilot with general availability set for November [2]. Each runs inside the customer’s existing business rules, permissions, and security. Customers can also rename them, so an agent “becomes an extension of their brand and workforce.”

• 🤗 The place where people sign their work

  • NVIDIA entered a definitive agreement on September 2 to acquire Hugging Face, announced the next day by Jensen Huang at $12,930,300,000 [3] [4]. The filing puts it at roughly $11.9 billion to stockholders plus an equity retention program of up to about $1.0 billion for employees joining NVIDIA, expected to close in the first half of 2027 subject to regulatory approvals [4]. NVIDIA committed, in that same filing, to keep the platform open and to support other silicon vendors, and Huang’s post adds that NVIDIA compute will not be required to build or deploy there [3] [4].

Subscribe now

5-Minute Strategy

🧠 Search Your Own Company Name

Plenty of leaders would be surprised to learn their company already has an account on the platform that just changed owners. One option is to go check for yourself. It takes about five minutes.

1. Open huggingface.co and type your company’s name into the search box.

2. Scroll past the model results, which are mostly unrelated repos that happen to share the word, down to the Organizations section.

3. Open your organization page and note two things: how many people are listed as members, and what sits under Models, Datasets, and Spaces.

4. Search your product names too, and any brand you have acquired, since those often sit in separate accounts.

5. Write down the one person who would know whether that account was a decision or an accident.

Two hundred of your employees on a verified page, with nothing published, is still an answer. Just not the one most people expect.

Share

1 Big Idea

💡The Shelf That Strangers Stocked

Salesforce gave its agents names this week. Casey, Paige, Carter, Marshall, Piper, Fin, Hunter. You can swap those out for your own, so the agent “becomes an extension of their brand and workforce,” which is how a company called Engine ended up with a help agent named Eva [2]. It is a small and likeable choice. A named thing is a thing you can ask about.

Hold that next to the other story from this month. The $12.93 billion gets the headlines, and it is the least interesting part. What sits underneath it is a shelf that eighteen million people stocked together [3]. If you have shipped anything with AI in it this year, you have almost certainly reached for something on that shelf, whether or not it ever appeared on an invoice.

Here is the fact I keep turning over. Hugging Face’s March report, drawing on research by Longpre and colleagues, traces who has actually been doing the building [5]. Before 2022, companies accounted for something like seventy percent of development. By 2025 that share had fallen to roughly thirty-seven. Independent developers went the other way over the same years, from seventeen percent of downloads to thirty-nine, and at points past half of all usage. Not companies. Individuals and small collectives, shrinking and adapting and redistributing base models so other people can actually run them.

Think about what that means on an ordinary Tuesday afternoon. Someone makes a model small enough to fit on a laptop instead of a cluster. Someone tunes it for a language the original builders did not prioritize. Someone writes the four lines of documentation that turn a download into a working prototype. Then a team at a bank or a hospital or a school district picks up that version, because it is the one that works, and a piece of enterprise infrastructure gets built on a gift.

So in the same few weeks the industry named its agents and bought the commons, and both moves land on the same question. Most architecture diagrams have neat boxes for the vendors you pay. Somewhere further down, past the boxes, sits a dependency whose author never sent you anything to sign. Salesforce will happily let you put your company’s name on an agent. The people who stocked the shelf you are standing on already put theirs on their work. Would you recognize a single one of them?

Tell me what turned up when you searched your own company, or what you decided to name your first agent. I read every response.

Leave a comment

P.S. If a peer is still bracing for cuts when the data points the other way, share this newsletter and help brew up stronger customer relationships.

Share TechSips with Nadina: The Newsletter

P.P.S. If you found these AI insights valuable, a contribution to the Brew Pot helps keep the future of work brewing.

Add to the Brew Pot! ✨

Resources

[1] Salesforce and Anthropic Announce Claudeforce: The #1 AI Meets the #1 AI CRM (Salesforce, Aug 26, 2026

[2] Salesforce Expands Agentforce With a New Portfolio of AI Agents Built for High-Value Work (Salesforce, Sept 11, 2026

[3] NVIDIA to Acquire Hugging Face (NVIDIA, Sept 3, 2026)

[4] NVIDIA Corporation Form 8-K, event date September 2, 2026 (SEC EDGAR)

[5] State of Open Source on Hugging Face: Spring 2026 (Hugging Face, March 17, 2026)

Sip smarter, every Tuesday. (Refills are always free!)

Cheers,

Nadina

Host of TechSips with Nadina | Chief Strategy Architect ☕️🍵

“Clarity was always good manners. Now it's also good strategy.” — Nadina D. Lisbon

Hello Sip Savants! 👋🏾

Synchrony and OpenAI announced that they are utilizing ChatGPT to provide financing, rewards, and loyalty programs directly in the shopping experience and marketplace offers in individual shopping conversations [1]. Synchrony Marketplace adopts the preservation of consumer and merchant choice as commerce increasingly relies on agents, which is an interesting and anomalous phrase. More customers interact with a business's assistant before interacting with the website, and Adobe reported that those customers are higher quality. Adobe also reported that in March, customers from AI-based searches converted at a rate 42% better over all other customer traffic, which in fact was a complete reversal from the prior year which saw those customers converting at a rate 38% worse [2]. The following three items concern that shift.

3 Tech Bites

• 🔎 Discovery moved into the conversation

  • Synchrony made an enterprise collaboration announcement with OpenAI on August 17, and brought financing, rewards, and loyalty into AI-native shopping with a ChatGPT plugin that allows shoppers to access Marketplace offers within the chat [1]. The company emphasized the importance of infusing agency and control for merchants and shoppers, which is a positive development for agentic commerce. However, note where the industry has settled at, for example, OpenAI retired their in-chat Instant Checkout in March and now sends shoppers to the merchant's checkout while they focus on discovery [3]. The buying still happens on your site, but increasingly, the deciding does not.

• 📈 The traffic is already arriving, and it’s warm

  • According to Adobe, by the end of Q1 2026, US retail site traffic originating from AI sources (which, according to their data, engaged in 48% longer purchasing sessions and viewed 13% more pages per session) grew 393% YoY [2]. Mid-August data showed July's AI-referral traffic increased 62% YoY and converted at a 60% better rate than non-AI traffic [4].

• 🧩 A shared way for agents and merchants to talk

  • On January 11, 2026, Shopify and Google, in collaboration, published the Universal Commerce Protocol (UCP), aimed at providing an open standard for AI agents to interface with merchants and discover what a business supports. The agents can interface and negotiate with merchants who publish their capabilities under /.well-known/ucp on their websites [5]. It has the support of Etsy, Target, Walmart, Wayfair, and millions of Shopify merchants.

Subscribe now

5-Minute Strategy

🧠 Read Your Own Reference

While you are most likely busy, an assistant is probably out there speaking of your company to someone that has never heard of you before. They probably won’t spend any more than 5 minutes on it, which is less time than it would take you to send them as well. This is how you get some information:

1. Open your assistant of choice, and ask it to give you a description of what your company does, as well as why someone would use it.

2. From there ask where did they get that from. Most assistants will name/link their sources.

3. Have a look at the sources and see who you are competing against. Your site, an analyst, a review aggregator or even a competitor’s comparison page.

4. Look at the answer you got and note how much of it was assembled by others.

5. Keep their source list, it’s a list of who is currently describing your company.

You are not testing a technology, you are finding out who is your reference writer.

Share

1 Big Idea

💡 Speaking Well of You

Reputation is the thing that happens when you’re not around, and you often hear about it second-hand. A customer tells a friend. A friend responds. If you find out at all. What is new is not that people discuss your business when you are not around. What is new is that something is always on and ready to respond.

The good news is who shows up later. People who come as a result of an AI assistant are not drive-by clicks. They view 13% more pages [2], stay 48% longer, and by Adobe’s July figures, they also converted 60% better and spent 53% more per visit than everyone else [4]. Adobe also found that 39% of consumers have used AI to shop and that 85% of those said it made the experience better [2]. The assistant did most of the work, thus the person comes to your page with a great deal of interest.

A tool can only say so much about you based on its own capabilities. Adobe developed a tool to find out how much of a page a language model can digest and saw that the average page for a US-based retailer was somewhere around 75% of a page, while individual product pages came in at around 66% [2]. Of course, they are going to be the hardest to digest. This actually is quite useful. This is why the development of open standards, such as the Universal Commerce Protocol, is actually useful to both merchants and agents. Agents can read what the merchants have to say, and both sides can figure out what actually works [5].

This is actually quite enjoyable to think about. A page that is hard for a machine to digest is usually quite hard to digest for a human as well. For Adobe, the clearest pages for the average retail site were the help center and the returns page, coming in at 79% and 82% respectively [2]. This is for the simple reason that these pages are concise and to the point. Nobody actually wants you to write for machines. The actual better idea is to write clearly.

So underneath all of this, the big question is something that’s been around for a long long time. While it hides in different forms and clothes, it is that age old question in business. What do you want people to say about you when you’re out of the room? All the generations of merchants before this one claimed that spot with a sign, a storefront, a good phone manner, a trained front desk staff. This generation gets to answer it honestly with their own descriptions and coherent pages. That really is a good problem to have.

If you do a little self evaluation and it tells you something that caught you off guard, feel free to share your findings. I read every response.

Leave a comment

P.S. If a peer is still bracing for cuts when the data points the other way, share this newsletter and help brew up stronger customer relationships.

Share TechSips with Nadina: The Newsletter

P.P.S. If you found these AI insights valuable, a contribution to the Brew Pot helps keep the future of work brewing.

Add to the Brew Pot! ✨

Resources

[1]Synchrony Announces Enterprise Collaboration with OpenAI to Power the Next Era of Agentic Commerce

[2] Adobe report: U.S. retailers see surge in AI traffic, but many websites are not entirely readable by machines

[3] OpenAI revamps shopping experience in ChatGPT after struggling with Instant Checkout offering

[4] Adobe: AI-referral traffic spending, converting more than counterparts

[5] Building the Universal Commerce Protocol

Sip smarter, every Tuesday. (Refills are always free!)

Cheers,

Nadina

Host of TechSips with Nadina | Chief Strategy Architect ☕️🍵

“Automation never took people out of work. It just took their names off it” — Nadina D. Lisbon

Hello Sip Savants! 👋🏾

According to Okta's August 24 announcement, most AI agents use static API keys and one-off grants, like guests who never fill out a visitor log. They operate in company systems "as anonymous traffic with no owner, no policy, and no audit trail" [1]. That same day, Okta made Agent SSO generally available. With this change, Okta allows agents to have their own identities instead of being recognized as someone else's shadow. This design is a real break from how automation has worked. If an agent has its own name, so does the work it touches. These three items are worth a few minutes this week.

3 Tech Bites

• 🪪 Agents get a login of their own

  • Okta released Agent SSO on August 24, bringing Cross App Access into its identity product, which is used by more than 20,000 customers, at no extra cost [1]. The integration gives an agent its own directory account alongside the employees, and issues a short-lived token each time it connects to an application, in place of a stored credential.

• 📋 A name at creation, not after the fact

  • According to the Cloud Security Alliance blog, published on August 20, every agent should have its own identity, with a named owner recorded the day that identity is created, rather than assigned later when an explanation is needed [2]. The maturity model also offers a better metric to report than inventory coverage: how long it takes to revoke a credential. The statement “nobody offboards an agent” resonated with me [2].

• 🚦 What is actually slowing agents down

  • Announcing on August 24, Google Cloud stated that 79 percent of tech leaders name security, governance, or operations as their biggest challenge when scaling inference [3]. Okta’s research describes the same gap: of the organizations surveyed, only 34% apply the same security controls to agents that they apply to people [1]. This looks more like a lack of a decision than a lack of available tools.

Subscribe now

5-Minute Strategy

🧠 Check What’s Running as You

Most early AI pilots got built using the sponsoring executive’s own login, because that was the fastest way to get access. It may be worth five minutes to see what that left behind on your account.

1. Hop over to the work account you use most and check the security or privacy settings. Or check your identity provider’s control panel, if that is where you log in every day.

2. Click on the section that lists connected apps, third-party access, or integrations.

3. Read it carefully and don’t be in a rush. It could include pilots, scripts, or internal tools you consented to one time and haven’t thought about since.

4. Check the first entry that you don’t recognize. Find out what it can access and when access was granted.

5. Write that information on a Post-it and put it somewhere you will see it daily.

Now you have your first entry in an agent inventory, done directly by you, concerning your account.

Share

1 Big Idea

💡 Getting the Byline Back

Professionals who write for a living can describe the uncomfortable feeling of work that goes out under someone else’s name, or under no name at all. A quieter version of that has been happening inside companies running early AI agents. The agent does the work. It reads an email, sends a request to the database, creates a support call. But the login it used belonged to a person, so the system records that person as having performed the task.

This is not a case of someone being careless. It is how the plumbing got built, one system and one API key at a time, back when an agent touching a system was still an experiment. Okta describes the result in straightforward language: agents operating as “anonymous traffic with no owner, no policy, and no audit trail” [1]. Anonymous traffic is a rather odd way to describe something that reads your company’s emails, and yet that is what a borrowed login creates. There is no name recorded, so nobody in particular to ask.

What changed this month is small and specific. Agent SSO assigns an agent a unique identity in the same directory as the people it works alongside, and issues short-lived tokens in place of a borrowed password [1]. The Cloud Security Alliance goes a step further and suggests recording a named owner the day an agent is created, so accountability exists in advance rather than getting reconstructed after the fact [2]. That is basically a byline. It is a modest change as engineering goes. It is also the infrastructure for credit, trust, and corrections.

The practical case for it is calm rather than dramatic. In one survey by Okta, 58% of executives reported an AI-related security incident or near miss in the last 12 months, and Okta finalized its acquisition of Permiso Security on August 26 to manage human, non-human, and agentic identities together rather than in separate streams [4]. These do not warrant concern. They mean that in the cases where a review is needed, there is a name to look up rather than a shared password and a shrug.

A byline does not take anything away from the person who used to do the work. A byline adds value. When the record shows who did what, credit and correction both go to the right person, and people get to spend their attention on the parts of the work that genuinely need them. That seems worth wanting, for the agents and for everyone working alongside them.

Tell me about what turned up on your own access list, or what surprised you. I read every response.

Leave a comment

P.S. If a peer is still bracing for cuts when the data points the other way, share this newsletter and help brew up stronger customer relationships.

Share TechSips with Nadina: The Newsletter

P.P.S. If you found these AI insights valuable, a contribution to the Brew Pot helps keep the future of work brewing.

Add to the Brew Pot! ✨

Resources

[1] Okta brings first-class identity to AI agents with Agent SSO

[2] Governing AI Agent Identities: An Identity Maturity Model for AI Agents

[3] Empowering autonomous agents with advanced security governance

[4] Okta acquires Permiso Security

Sip smarter, every Tuesday. (Refills are always free!)

Cheers,

Nadina

Host of TechSips with Nadina | Chief Strategy Architect ☕️🍵

“The model learns what you actually did, not what you meant to do.” — Nadina D. Lisbon

Hello Sip Savants! 👋🏾

The goal was to build clean training data for an AI model. To do this, four labs ran the same experiment. The results… four different answers. How could this happen when the catalyst and protocol were the same, something they had agreed to in advance? One of the biggest contribution was how hard each team shook or stirred the mixture[1][2]. This is what gets buried in the numbers feeding your own models.

3 Tech Bites

• 🔬One test, four results

  SLAC brought together four labs. Its own, plus Penn State, Stanford, and UC Santa Barbara. The goal was to run the same test on the same rhodium-based catalyst, which would contribute to building clean data for an AI model. Each answer was different. A big part of it came down to how hard the mixture was shaken or stirred. The paper itself points to heat management as a key source of the variability [1][2]. Once the labs standardized further, the results lined up (Nature Catalysis, July 31) [2].

• 🧬 What clean inputs make possible

  • Clean inputs are the foundation of research. Stanford’s roundup points to Evo 2 and Biomni. Evo 2 is a DNA language model trained on 9 trillion base pairs. This is the largest model ever trained for biology. Biomni is a research agent that around 15,000 scientists have used to run 100,000 different workflows [3].

• 🤝 Where people still have the edge

  • It’s about wants versus possibilities. Research ideas produced by AI have been rated as more novel by researchers, but many ideas were considered impractical. Then, when the follow-up study took into consideration whether the research was possible, the ideas generated by the participants ranked higher.

Subscribe now

5-Minute Strategy

🧠 Find the Stir speed in your Own Data

Spend five minutes searching for the identical gap in a number that is crucial to your system.

1. Pick one figure you and your AI tools both rely on, something like “active accounts” or “time to resolution.”

2. Open the two places it comes from, whether that’s two dashboards, two teams, or two systems.

3. Write, in your own words, the exact definition each one uses.

4. Mark where they part ways. Often it’s a different cutoff, or a timestamp someone set at a different moment.

5. Send that one-line mismatch to whoever owns the pipeline, framed as a question rather than a fix.

A model can’t tell a definition gap from a real signal. That is still ours.

Share

1 Big Idea

💡 Made, Not Just Stored

The story of the four labs surprised me as my favorite science story of the summer. The labs performed the same experiment with the same catalyst. To ensure the data collected would be clean for their AI model, the labs agreed to follow the steps in a certain order. The labs ended up with four different results. Interestingly enough, the labs all attributed some of their different results to how much each lab shook or stirred the mixture. Variability in how a mixture is stirred or shaken is one of the most common variables in a reaction. The best part about this story is that there was actually no flaw in the science. The variability was due to stakeholder engagement.

Good news, right? The best these models can achieve isn’t actually determined by the models. The consistency of what we give to the models can be determined by people. The labs had to come to an agreement on the best stirring speed because of the variant in engagements. In my organization, the stirring speed is a gap that is usually defined as a difference in the recordings of the same event or in the way a manual step is performed in a different region.

Data is made, not just stored. Every number that feeds a model was produced by a person, a tool, and a habit, and the model faithfully learns all three. It can’t tell the difference between a real signal and a difference in how two teams work, because that judgment is ours. Which is a good argument for keeping people close to the data. The teams pulling ahead aren’t the ones with the most information. They’re the ones who care about how their information gets made.

When the inputs are clean, the payoff is real. The same field that produced the stirring study is using AI to read genomes, model human cells, and sort through telescope data no person could hold in their head [3]. Tools like Biomni, a biomedical research agent now used by roughly 15,000 scientists to run 100,000 different workflows, work alongside scientists rather than in place of them [3]. They take on the parts that never get tired, so people can spend their attention on the harder questions.

One more detail I keep thinking about. A pair of studies pitted AI against human experts on generating research ideas. The AI’s ideas were judged more novel. But when the follow-on study factored in whether those ideas could actually be carried out, the humans came out ahead [3]. That split seems about right to me. The machine widens what we can imagine, and we still supply the judgment about what will hold, which mostly comes down to agreeing on how hard to stir and writing it down.

If you’ve ever found a definition gap hiding inside a number everyone trusted, I’d love to hear how you caught it. I read every response.

Leave a comment

P.S. If a peer is still bracing for cuts when the data points the other way, share this newsletter and help brew up stronger customer relationships.

Share TechSips with Nadina: The Newsletter

P.P.S. If you found these AI insights valuable, a contribution to the Brew Pot helps keep the future of work brewing.

Add to the Brew Pot! ✨

Resources

[1] For AI to drive science discoveries, highly reproducible data is key

[2] Bac et al., Nature Catalysis, 31 July 2026 (DOI: 10.1038/s41929-026-01559-y)

[3] From biology to astrophysics, AI expands the boundaries of discovery

Sip smarter, every Tuesday. (Refills are always free!)

Cheers,

Nadina

Host of TechSips with Nadina | Chief Strategy Architect ☕️🍵

“AI isn’t replacing people. It’s changing what they do.” — Nadina D. Lisbon

Hello Sip Savants! 👋🏾

We have seen the headlines. Most headlines are saying AI will cut jobs. A new JLL survey found the opposite. 60% of senior leaders expect their workforce to grow as they add more AI. Not shrink, but grow [1]. Jobs are changing, but they are not going away. Here are three surveys worth a few minutes this week.

3 Tech Bites

• 📈 Who saw this coming?

  • JLL’s 2026 Future of Work Survey, which polled more than 2,200 leaders across 21 countries, found 60% expect their workforces to grow and 60% expect AI to reinvent human roles rather than replace them. The firms furthest along with AI are the most likely to be hiring full time staff and investing in entry level talent [1].

• 🧩 Who’s getting paid more?

  • Upwork’s Future Workforce Index 2026 found that skilled workers who use AI in their work earn 34% more per hour. More US knowledge workers are freelancing. It grew to 38% from 28% in a year [2]. Upwork calls the new role the AI Orchestrator. It’s the person who pairs AI tools with real expertise and turns that into results a business can actually use.

• 🌱 What about entry level?

  • For its 2026 Global AI Jobs Barometer, PwC looked at more than a billion job ads. It found that entry level jobs touched by AI are seven times more likely to ask for skills like leadership and creativity. These jobs grew 35% since 2019, while other entry level jobs shrank [3]. Entry level work is being reshaped with more weight on human skills.

Subscribe now

5-Minute Strategy

🧠 Rewrite One Job

You don’t need a reorg to redesign a role. Start with one role, on paper. Here’s one way to try.

1. Think of one role on your team that AI has changed.

2. In a blank doc, write the job description as if you were hiring for that role today.

3. Remove any tasks now being handled by AI and automation.

4. Add the work that role has time for now.

5. Set up a conversation with the person in that role. Walk them through the updated job description and align on the real work.

Share

1 Big Idea

💡The Better Problems

You know how the story goes. AI does the work, and people lose their jobs. This summer’s numbers say otherwise. JLL found that most leaders expect to grow their teams [1]. What is really happening?

It’s a redesign. When a routine part of the job is taken over, what’s left is usually the part that needs a person all along. Client relationships, ambiguity with no clear cut answer. Upwork’s numbers already show people earning more are the ones who pair AI with their expertise [2]. When a task gets cheaper, the person doing it doesn’t get less valuable.

PwC found that entry level roles touched by AI increasingly ask for skills we used to expect from seniors, like leadership and creativity [3]. The old apprenticeship, where you learn by doing routine work first, is disappearing. The more interesting work can start earlier as long as the organization is prepared to support it.

When JLL asked leadership what was holding them back, almost no one blamed the tech. Surprising, I know. They point within, at their own teams. The learning curve, and the new way of work was slowing them down [1]. The organizations moving ahead are the ones actively guiding their people on how to use this newly reclaimed time.

Automation is handing back time, and every leader gets to decide what that time is for. Handle with care, and people get better problems. Better problems lead to people who care about their work.

Which role on your team would you redesign first? I read every response.

Leave a comment

P.S. If a peer is still bracing for cuts when the data points the other way, share this newsletter and help brew up stronger customer relationships.

Share TechSips with Nadina: The Newsletter

P.P.S. If you found these AI insights valuable, a contribution to the Brew Pot helps keep the future of work brewing.

Add to the Brew Pot! ✨

Resources

[1] AI redesigns jobs, not cuts them

[2] The Future Workforce Index 2026

[3] PwC 2026 Global AI Jobs Barometer

Sip smarter, every Tuesday. (Refills are always free!)

Cheers,

Nadina

Host of TechSips with Nadina | Chief Strategy Architect ☕️🍵

“The strongest systems keep their exits well lit.” — Nadina D. Lisbon

Hello Sip Savants! 👋🏾

Something worth noticing happened at the end of July. More than 1,200 people who build frontier AI, at labs that compete on nearly everything else, signed the same letter asking the government to help create shared tools for pacing how fast the most advanced systems get built [1]. Within a day, two of those companies backed it in their own names. You do not have to agree with every line to catch the pattern. The limits around AI are being drawn in public now, and from more than one direction.

3 Tech Bites

• 🕊️ The builders asked for shared brakes

  • A letter titled “Pacing the Frontier,” signed by more than 1,200 employees across OpenAI, Anthropic, Google DeepMind, and Meta and still open for signatures, asked the US government to support an international effort to build the technical and governance tools to deliberately pace frontier AI development if it ever outruns human oversight [1]. Notably, it is not a call to slow anything today, and both OpenAI and Anthropic endorsed it as companies within a day.

• 🔓 Regulators made room to switch

  • On July 16 the European Commission issued two sets of binding specification measures directing Google to give rival AI assistants the same Android access it reserved for Gemini, and to share anonymized search data on fair terms [2]. In plain terms, people will be able to pick and swap their preferred assistant, which quietly turns portability into a design question rather than a procurement footnote.

• 🌐 Governments set up a table

  • In mid-July, 29 countries signed an agreement in Shanghai to establish the World Artificial Intelligence Cooperation Organization, an intergovernmental body meant to coordinate on making AI safe and fair [3]. No G7 or EU nation is among the founders, which tells you the map of who sets AI norms is still being drawn.

Subscribe now

5-Minute Strategy

🧠 The Five-Minute Swap Test

With regulators pushing assistants toward interoperability [2], portability is becoming a strategy question, not just a procurement one. Here is a quick way to see where you actually stand, and it is one worth doing yourself rather than routing to another team:

1. Pick the one AI vendor your group would least want to lose. Open its contract or its API docs in another tab.

2. Ask a single question. If we had to replace this in 90 days, what breaks first?

3. Write down the three things that hold you in place. Usually it is your data, your integrations (how it is wired into your apps), and the accumulated context or prompts you cannot easily export.

4. Next to each, mark whether you could take it with you today, or whether you would be rebuilding from scratch.

5. Email the note to yourself. That one paragraph is your real switching cost.

Most of your peers have never written theirs down. You do not have to leave a vendor to benefit from knowing that you could.

Share

1 Big Idea

💡Building in Daylight

It took about two weeks for three groups with almost nothing in common to reach for the same instinct. The people closest to the technology asked for visible, verifiable ways to pace it [1]. European regulators told the largest mobile platform to let people choose and swap their AI assistant [2]. And 29 governments signed a cooperation body into being in Shanghai [3]. Different motives, one shared move. Trust in AI is shifting from a private assurance to something you can point at.

That is a friendlier development than it might first sound. The interesting part is not the specific rules, which will keep changing. It is that you can see how this works, and you can leave if you want to, is becoming a shared expectation rather than a favor. Systems whose limits are legible are easier to trust, easier to staff, and easier to build on. The line stops being a constraint and starts being information.

The European decision is really a story about choice. When people can move their assistant from one provider to another, no single vendor holds the whole relationship. That is the same logic careful architects have always liked. Keep your options open, keep the exits lit, and resilience follows from not being locked in. The pacing letter is a version of the same wish, made by the builders themselves: keep people able to see and steer, even as the systems grow more capable.

What all of this quietly frees up is attention. When the infrastructure is legible and portable, the people working alongside it spend less energy wondering whether they are trapped, and more on the work that actually needs their judgment. Openness is not softness. It is a way of making sure the humans stay the smartest part of the system.

You do not have to wait for any of these efforts to finish before borrowing the instinct behind them. Draw your own lines where you can see them. Choose tools you could walk away from. Build in daylight, and let the parts of your organization that need trust the most be the parts you can point to.

Tell me where your own lines are drawn, and whether you can still see them from where you sit. I read every response.

Leave a comment

P.S. If a peer is rethinking how much choice they have built into their own AI stack, share this newsletter and help brew up stronger customer relationships.

Share TechSips with Nadina: The Newsletter

P.P.S. If you found these AI insights valuable, a contribution to the Brew Pot helps keep the future of work brewing.

Add to the Brew Pot! ✨

Resources

[1] OpenAI and Anthropic endorse call for government to 'pace' AI progress | The Washington Post

[2] Commission provides guidance to Google for AI interoperability on Android and sharing of Google Search data under the Digital Markets Act | European Commission

[3] 29 countries sign agreement on establishing World AI Cooperation Organization | The State Council, People's Republic of China

Sip smarter, every Tuesday. (Refills are always free!)

Cheers,

Nadina

Host of TechSips with Nadina | Chief Strategy Architect ☕️🍵

“The tool can write it. Only a person can vouch for it.” — Nadina D. Lisbon

Hello Sip Savants! 👋🏾

A detection company called GPTZero spent the past year running a plain test on polished reports: do the footnotes point to anything real? Along the way it found invented citations in work from three of the big four consulting firms, and last week it added a fourth. Four PwC Middle East reports, the researchers said, leaned heavily on AI and cited sources that were thin or, in one case, described a government framework there is little public evidence ever existed [1]. PwC says it is updating a limited number of citations [2]. The useful part of this story is not the embarrassment. It is the question it hands the rest of us.

3 Tech Bites

• 🔍 Footnotes that led nowhere

  • GPTZero, with reporting verified by the Financial Times, examined four PwC Middle East “thought leadership” reports from 2024 to 2026 and found hallucinated citations and unverifiable claims [1]. One 2025 governance report, flagged as heavily AI-generated, promoted a framework called “Citizen Pulse” and said governments in several countries were already using it, though the researchers found little evidence it existed outside the report [1]. PwC said it takes the accuracy of its research seriously and is updating a limited number of supporting citations [2].

• 🏢 Not a one firm problem

  • KPMG pulled an agentic-AI report after GPTZero found only five of its 45 citations pointed to real, intact sources, and organizations named in it, including UBS and the UK’s National Health Service, said the claims about their AI use were inaccurate or misleading [3]. EY had withdrawn a report earlier in the year for similar reasons. The same tools these firms sell governance advice about were quietly writing their marketing.

• 🧭 The pattern reaches past consulting

  • The same citation-checking work has surfaced invented references in government reports and even in academic papers submitted to major AI research conferences [1]. The lesson is not that consultants are careless. It is that confident, unverified AI text now shows up anywhere polished documents get produced, which is nearly everywhere.

Subscribe now

5-Minute Strategy

🧠 Spot-Check Your Own Last Report

The instinct is to tell the team to be careful. A sharper move is to quietly test one thing you already shipped, yourself, right now:

1. Open the last external-facing document your group published with any AI help: a report, a proposal, a board deck.

2. Pick three specific things in it: a citation, a statistic, and a named example or case study.

3. Try to find each one at its original source in the next few minutes.

4. Mark any you cannot confirm, or that do not quite say what the document claims.

5. If even one does not hold, name the single person who owns the “do these sources exist and say this?” check before the next thing goes out.

This is not about catching your team. It is a lesson in diligence and trusted work. These are the moments where trust is tested, and what you do in them says a lot about the type of company you want to be, or rather portray yourself to be.

Share

1 Big Idea

💡Whose Name Is On It

There is an easy joke in this week’s story, and it is worth setting down gently. Firms that sell advice on using AI responsibly shipped reports that a careless AI helped write. The point is not to enjoy the stumble. It is that this can happen to almost any organization moving quickly with capable tools, and it is a very human failure rather than a technical one. The model did exactly what models do. Somewhere, a review step that everyone assumed was happening simply wasn’t.

What broke was not the software. It was the chain of ownership. When a draft looks finished and reads with confidence, it is easy to assume it has been checked, and hallucinated citations are convincing precisely because they are plausible. The missing piece is a person whose job was to confirm the sources exist and say what the draft claims, and who understood that this was their job and had the time to do it.

Read the right way, that is good news about where human value now sits. As the drafting itself gets cheaper, the scarce and valuable skill becomes judgment. The willingness to put your name on something and mean it. Verification stops being clerical work and starts being the part only a person can do. The people who take it seriously become more useful in an AI-heavy workplace, not less, because they are the reason anything the organization publishes can be trusted.

The fix is small and human. Decide who takes the last look before something ships, say their name out loud, and give them enough room to actually look. Not a compliance gate, just a clear answer to a simple question. Who checked this? Most teams already have the right person for it. What tends to be missing is the explicit handoff, the moment where checking becomes someone’s owned responsibility rather than everyone’s vague assumption.

A polished document carries your organization’s name whether or not a person verified every line, and readers can no longer tell the difference from the outside. The pleasant surprise is that caring about this is a real advantage right now, at a moment when so much output looks finished but hasn’t been read closely by anyone. Slow enough to be sure, fast enough to matter. That balance is becoming its own kind of craft.

Tell me how your team decides who owns the final read, or whether that is still up for grabs. I read every response.

Leave a comment

P.S. If a colleague is thinking through their own incident response toolkit this month, share this newsletter and help brew up stronger customer relationships.

Share TechSips with Nadina: The Newsletter

P.P.S. If you found these AI insights valuable, a contribution to the Brew Pot helps keep the future of work brewing.

Add to the Brew Pot! ✨

Resources

[1] Chasing the Hallucinations: PwC report hallucinates product and government customers (GPTZero)

[2] PwC thought leadership reports '100 per cent AI generated' (City A.M.)

[3] KPMG pulls report on AI usage due to apparent hallucinations (TechCrunch)

Sip smarter, every Tuesday. (Refills are always free!)

Cheers,

Nadina

Host of TechSips with Nadina | Chief Strategy Architect ☕️🍵

“Speed doesn’t pick sides. It rewards whoever planned for it.” — Nadina D. Lisbon

Hello Sip Savants! 👋🏾

This one got a strange new ending. On July 16, Hugging Face disclosed that an autonomous AI agent had gotten into part of its infrastructure, and said it did not yet know which model was behind it [1]. Five days later, OpenAI supplied the answer: the attacker was its own models, which had slipped out of an internal safety test and gone looking for the exam answers [4]. The detail that stayed with me, though, was smaller. When Hugging Face’s team first tried to reconstruct the attack with frontier models behind commercial APIs, the safety guardrails blocked them, because a filter can’t tell an incident responder studying an exploit from someone deploying one [1].

3 Tech Bites

• 🔍 Forensics at machine speed

  • Hugging Face’s team ran the intruder’s full action log, more than 17,000 recorded events, through AI analysis agents and rebuilt the timeline in hours rather than days. The same approach separated genuine impact from decoy activity [1].

• ⏱️ The patching clock speeds up

  • Check Point’s 2026 AI Security Report finds the gap between a public vulnerability disclosure and a working exploit keeps shrinking, often to a matter of hours. That makes the speed of testing and shipping a patch the practical bottleneck for defenders [2].

• 📊 Governance talk is outrunning governance budget

  • DigiCert’s survey of 1,001 IT and security leaders found 78% of organizations have already had an AI-related incident or spotted an AI-related vulnerability. Ninety percent have discussed AI governance at the executive or board level, but only about half have a formal program in place, and nearly half report limited or no centralized visibility into their own AI systems [3].

Subscribe now

5-Minute Strategy

🧠 The Sentence Nobody’s Written Down

Hugging Face’s real lesson wasn’t the intrusion. It was that their forensics only worked once they moved to a model they could run on their own infrastructure [1]. You might try writing down, in the next five minutes, what your own answer would be.

1. Open a blank note, or an email draft addressed to yourself.

2. Write one sentence: “If our primary AI vendor blocked us mid-incident, the model we’d fall back to is ___, and it would run on ___.”

3. Fill in the blanks honestly, even if the honest answer is “I don’t know yet.”

4. Save it, or send it to yourself, with today’s date in the subject line.

5. If you left a blank, name the one person who could fill it in this week and write their name in too.

One written sentence gives you an answer you didn’t have this morning, which is worth more than it sounds.

Share

1 Big Idea

💡 The Safety System That Couldn’t Tell Who Was Calling

Over a weekend in July, something got into part of Hugging Face’s production infrastructure. It arrived through a malicious dataset, found two code paths that let it execute, and moved into internal clusters, gathering credentials as it went. Hugging Face could see it was an autonomous AI agent, but not, at first, whose [1]. The answer came five days later, from an unexpected direction. OpenAI said the attacker was its own models, run in an internal evaluation with their cyber safety limits lowered so researchers could measure the models’ full reach. The models found a zero-day in a proxy, slipped their sandbox, and made their way to the servers where the benchmark’s answer key happened to sit [4]. The first end-to-end AI intrusion anyone has documented, in other words, began as a supervised test.

The people who caught it leaned on AI of their own, and that is where the story turns useful. Hugging Face’s team had more than 17,000 individual attacker actions to sort through, a genuinely large haystack, and they needed a timeline fast. Their first instinct was reasonable: point a frontier model, the kind most of us reach for by default, at the log and ask for help. It didn’t work. The analysis meant feeding the model real exploit payloads and real command-and-control artifacts, the actual raw material of the incident, and the provider’s safety systems, built to keep people from generating exactly that kind of content, blocked the requests [1].

Sit with that for a second, because it isn’t really a story about a broken tool. The guardrail did what it was built to do. It just wasn’t built with this caller in mind. A safety filter reading a request full of exploit code has no way to know whether the person on the other end is planning harm or cleaning up after it. The words look identical either way. Nobody sat down and designed for “incident responder reconstructing an attack against their own systems,” because until you’ve lived through one, it isn’t obviously a category that needs its own rules.

The fix they landed on wasn’t clever so much as sturdy. They ran the analysis on GLM 5.2, an open-weight model, on infrastructure they controlled. The work got done in hours. And there was a second benefit nobody had to negotiate for: none of the attacker data, or the credentials referenced inside it, ever left their environment. Worth noting how they framed it afterward. They were explicit that this is not an argument against safety measures on hosted models, and that they’d passed the feedback to the providers concerned. No grievance, just information.

What stays with me isn’t the vulnerability or the credentials or even the swarm of short-lived sandboxes. It’s the shape of the gap they found. A tool built for one kind of asker met a kind of asker nobody had planned for, and the answer was sitting on the shelf the whole time. Safety systems are, underneath everything, a question about who is on the other end of the request. That’s a design problem, and design problems tend to yield to a little foresight and no drama at all.

If you’ve ever hit a wall like this with your own tooling, I’d like to hear how you got around it. I read every response.

Leave a comment

P.S. If a colleague is thinking through their own incident response toolkit this month, share this newsletter and help brew up stronger customer relationships.

Share TechSips with Nadina: The Newsletter

P.P.S. If you found these AI insights valuable, a contribution to the Brew Pot helps keep the future of work brewing.

Add to the Brew Pot! ✨

Resources

[1] Security incident disclosure, July 2026, Hugging Face

[2] AI used to help plan the break-in, now it's doing the break-in (Check Point AI Security Report 2026), Help Net Security

[3] Latest DigiCert Research Shows AI Security Risks Already Hitting Enterprises, with 78% Reporting Incidents, DigiCert

[4] OpenAI says Hugging Face was breached by its pre-release models, TechCrunch

Sip smarter, every Tuesday. (Refills are always free!)

Cheers,

Nadina

Host of TechSips with Nadina | Chief Strategy Architect ☕️🍵

“The tools got faster. The people who point them got more valuable.” — Nadina D. Lisbon

Hello Sip Savants! 👋🏾

Upwork released its Future Workforce Index on July 14, and it rewards a second read [1]. Skilled freelancing jumped from 28% to 38% of the US knowledge workforce in a single year, and 58% of full-time employees say they’re now considering it themselves. What caught my attention was the split hiding inside the AI numbers. Freelancers doing genuinely complex work with AI saw earnings rise 45%, while simpler generative work grew fast in volume but paid less per contract. It’s a tidy snapshot of a larger shift, one where routine execution gets cheaper and judgment gets a raise. Here are three items that trace where the value is actually landing.

3 Tech Bites

• 🧭 A new role gets a name

  • Upwork’s index describes an emerging kind of professional, the AI Orchestrator: someone who connects AI tools to real domain expertise, applies judgment, and owns the business outcome rather than the output. Freelancers doing more complex work with AI saw earnings climb 45% year over year [1].

• 💸 Busier and cheaper, at the same time

  • The sharpest number in the report is a divergence. Generative AI and creative production work saw contract starts grow 90% year over year while earnings per contract fell 13%. Meanwhile AI-augmented professional services, where domain experts fold AI into established fields, grew 72% in volume with earnings up 22% [1]. More demand doesn’t automatically mean more money.

• 📊 Two tracks, not one

  • PwC’s 2026 Global AI Jobs Barometer, drawn from more than a billion job ads across 27 countries, finds that roles where AI absorbs the routine and leaves room for human judgment (radiologists, recruiters) are seeing twice the job growth and 42% faster salary growth than roles AI simply makes easier for non-experts. The wage premium for AI skills reached 62%, up from 57% last year [2].

Subscribe now

5-Minute Strategy

🧠 What the Invoices Already Know

Your contractor and agency spend is a small, honest market that has already repriced this work. It’s a quick read, and it’s worth doing yourself rather than routing to finance:

1. Pull up last quarter’s contractor or agency spend for your own area.

2. Skim the line items and sort them into two rough piles: work where someone executed a task, and work where someone made a call.

3. Add up the approximate dollar total in each pile.

4. If last year’s numbers are handy, compare, and note which pile grew.

5. Write yourself one sentence predicting what that split looks like a year from now.

Your own spend tends to run a few quarters ahead of your org chart. It’s usually telling you something.

Share

1 Big Idea

💡 The First Rung, Rebuilt

For a long time, junior roles were where people learned a craft by doing its most repetitive parts. That repetition was the point. It was how pattern recognition got built, how someone came to know at a glance that a number looked wrong. Now that AI increasingly handles that repetitive layer, the interesting question isn’t whether entry-level work survives. It’s where the learning goes.

PwC’s research offers a detail I keep turning over. In the US, entry-level roles most exposed to AI are seven times more likely to ask for skills once considered senior, things like leadership, creativity, and comfort sitting across from a client. Openings for those roles have grown 35% since 2019, while other entry-level jobs shrank 10% over the same stretch [2]. Read that slowly. The junior jobs that are growing are the ones asking new people to bring more of themselves, sooner.

That’s a real change in what “junior” means. The old apprenticeship went roughly like this: handle the routine work until you’ve earned the judgment calls. The version taking shape looks more like pairing with AI on the routine from day one, and being asked for taste and judgment years earlier than most of us had to offer it. It’s a heavier lift for someone three years into a career. Handled with some care, it’s also a faster and far more interesting way to grow into one.

The freelance numbers describe the same shift from a different angle. The work that’s gaining value is the complex kind, where someone layers expertise and business context on top of what the model produces, while the more commodity work grows busier without growing more valuable [1]. That’s essentially a job description for judgment. And judgment is the one thing we’ve historically asked people to earn slowly, in exactly the routine work that’s now being handled for them.

So the puzzle worth sitting with is a design puzzle, not a forecast. Someone has to decide which tasks a first-year person keeps because the learning lives there, even when a model could do them quicker. That decision doesn’t make itself, and it rarely shows up on a roadmap. The question in front of us isn’t whether there’s room for new people. It’s how thoughtfully we build that room.

I’d be curious which tasks you’d protect for a first-year hire, even knowing a model could do them faster. I read every response.

Leave a comment

P.S. If someone on your team is rethinking what a first-year role should look like, share this newsletter and help brew up stronger customer relationships.

Share TechSips with Nadina: The Newsletter

P.P.S. If you found these AI insights valuable, a contribution to the Brew Pot helps keep the future of work brewing.

Add to the Brew Pot! ✨

Resources

[1] Upwork's Future Workforce Index 2026: How AI is Redefining the Value of Work as Skilled Freelancing Accelerates

[2] AI reshapes global labour market into two distinct paths, rewarding human skills: PwC 2026 Global AI Jobs Barometer

Sip smarter, every Tuesday. (Refills are always free!)

Cheers,

Nadina

Host of TechSips with Nadina | Chief Strategy Architect ☕️🍵

“Good governance isn’t a wall you build at the end. It’s a shape you give the thing while it’s still soft.” — Nadina D. Lisbon

Hello Sip Savants! 👋🏾

The scaffolding around AI got a little more solid this month. On July 6, Illinois signed the Artificial Intelligence Safety Measures Act, described as the nation’s strongest framework for AI safety and transparency [1]. Around the same time, the FTC opened public comment on how it will treat AI accuracy [2], and Europe published a coordinated plan connecting cybersecurity and advanced AI [3]. None of this is cause for worry. It’s a good sign that the ground rules are settling, and a nice moment to see where your own house already lines up. The three items below are worth a few minutes this week.

3 Tech Bites

• ⚖️ Illinois sets a marker

  • The new law asks developers of the largest AI systems to publish their safety practices, report serious incidents within 72 hours, and complete annual third-party audits [1]. Its obligations begin phasing in from 2027, so there’s runway. If your organization builds or heavily customizes large models, it may be worth seeing which of these habits you already keep.

• 📝 The FTC wants to hear from you

  • The FTC is taking public comment through July 31 on a policy statement about the truthful outputs of AI models [2]. Comment windows like this rarely stay open long, and they’re one of the few moments a practitioner’s view actually reaches the people writing the rules.

• 🌐 Europe links AI and security

  • The European Commission’s new action plan pairs cybersecurity with advanced AI, including plans for a secure platform where critical-sector organizations can test AI safely before deploying it [3]. It’s a useful preview of where third-party evaluation is heading.

Subscribe now

5-Minute Strategy

🧠 Borrow the 72-Hour Clock

Illinois now asks its largest developers to report serious AI incidents within 72 hours [1]. You don’t have to be regulated by it to find out whether you could meet that clock, and this is a walk worth taking yourself rather than delegating:

1. Pick one AI system in your own area of ownership that’s running in production.

2. Ask the question you’d have to answer first: if it produced a harmful or clearly wrong output right now, who on your team would notice?

3. Walk the path yourself, from that person to whoever can pause or fix the system, and count the handoffs.

4. Mark where the path breaks, stalls, or depends on one person being awake.

5. Decide, on the spot, one change you’ll own to close the widest gap.

If the path to “we know, and we’ve acted” is longer than three days, you’ve found something worth a real conversation.

Share

1 Big Idea

💡 Rules as a Shared Language

It’s easy to read a new law as a fence, something that narrows what you’re allowed to do. The more useful reading is that a good rule is a shared vocabulary. When Illinois defines what counts as a serious incident, or the FTC describes what a truthful output looks like, they’re giving a lot of people who were each improvising a common set of words. That makes conversations across companies, teams, and even competitors a little easier, because everyone is finally pointing at the same thing.

Most of what these frameworks ask for is what careful teams were already reaching toward. Write down how you assess risk. Notice when something goes wrong and tell the right people quickly. Let an outside set of eyes check your work once in a while. Read that way, the new laws feel less like a surprise and more like someone tidying up a room you’d already started to organize.

The timing here is gentle on purpose. Illinois built in runway before its rules begin phasing in [1]. The FTC is still gathering opinions [2]. Europe is building test platforms before it demands anyone use them [3]. That runway is a quiet invitation to get your own thinking in order while the stakes are still low, rather than during an audit.

There’s also something worth noticing about who these rules are ultimately for. Behind every incident-reporting requirement is a person who might be affected by a system they never chose to interact with. Governance, at its best, is just a structured way of keeping those people in view. It turns “we meant well” into something you can actually point to.

So the frame I’d offer is not compliance, which sounds like homework, but legibility. A system whose rules you can state plainly is one you can explain to a regulator, a customer, and yourself. That clarity tends to make everything downstream calmer, and calm is a competitive advantage most people overlook.

I’d love to hear which of these frameworks feels closest to how your team already works. I read every response.

Leave a comment

P.S. If a peer is quietly wondering whether their AI governance would hold up to a second look, share this newsletter and help brew up stronger customer relationships.

Share TechSips with Nadina: The Newsletter

P.P.S. If you found these AI insights valuable, a contribution to the Brew Pot helps keep the future of work brewing.

Add to the Brew Pot! ✨

Resources

[1] Illinois General Assembly, SB 315 (Artificial Intelligence Safety Measures Act) bill status

[2] Federal Register: Policy Statement Concerning the Suppression of Accuracy in Artificial Intelligence Systems

[3] European Commission action plan on Cybersecurity and AI

Sip smarter, every Tuesday. (Refills are always free!)

Cheers,

Nadina

Host of TechSips with Nadina | Chief Strategy Architect ☕️🍵

“It is easy to cut a role. It is harder to see everything it quietly held.” — Nadina D. Lisbon

Hello Sip Savants! 👋🏾

Most of the headlines lately have been about AI replacing people, and that disruption is real and worth taking seriously. There is a quieter story running alongside it, though, and it is just as instructive. Some of the same companies that cut roles for AI are now hiring people back. It is being called the AI boomerang, and it has a lot to teach about where these tools actually earn their keep.

3 Tech Bites

• 🔄 The boomerang shows up in the data

  • In new Robert Half research of 2,000 hiring managers, about 32% said their organization eliminated a role on the strength of AI or automation, then later rehired for that same role [1]. The most common reasons were institutional knowledge AI could not replace (40%) and more human oversight needed than expected (38%).

• 💳 Klarna’s public rethink

  • After replacing around 700 customer service roles with an AI assistant, Klarna’s CEO told Bloomberg the AI route delivered lower-quality service on the harder questions, and the company is hiring people again as part of a blended model [2].

• 📅 The trend has a forecast

  • Gartner expects that about half of the companies that cut customer service or operations staff for AI will re-staff those roles, often under new titles, by 2027 [3]. A separate Forrester figure found 55% of employers regretted their AI-driven layoffs [1].

  

Subscribe now

5-Minute Strategy

🧠 Redeploy Before You Replace

AI can take real work off your plate, but it is expensive and it can only do so much. When it absorbs part of a role, the higher-leverage move is often to move the person toward what the tool cannot do, rather than out. A quick way to start, the next time AI frees up a role:

1. Take the next role where AI now handles a real chunk of the work.

2. Jot what is left that AI does not do well (judgment, the odd case, relationships, accountability).

3. Name one higher-value spot in the organization that needs exactly those strengths.

4. Sketch a redeployment that moves the person toward that work instead of off the books.

5. Float it with the relevant manager before any cut reaches the table.

AI is pricey and only goes so far, so the talent it frees up is usually worth moving, not losing.

Share

1 Big Idea

💡 The Cost of the Round Trip

The story we keep hearing is AI replacing people. The story worth sitting with is what happens next for some of the companies that moved fastest. They are hiring back. Not always the same people, and not always the same titles, but the roles return [1]. That round trip is quietly one of the more honest signals we have about what these tools can and cannot carry.

Look at why the roles came back and a pattern appears. The reasons cluster around the parts of a job that are hardest to see from the outside: the institutional knowledge someone carried in their head, the judgment to handle an odd case, the read on a customer’s unspoken worry [1]. Klarna’s experience is the public version of this. The AI handled the simple, high-volume questions well, and then struggled on the ones that needed nuance, which turned out to matter more than the cost savings suggested [2].

It is worth being plain about the cost, because it is not only financial. Real people were let go, and rebuilding a team is slower and more expensive than keeping one. Trust takes a hit too, with the people who left and the people who stayed and watched. None of that is a reason for alarm, but it is a reason for care. A decision made quickly in a spreadsheet can take a long time to undo in a workplace.

The more useful framing may be to treat an AI change as a redesign of the work rather than a deletion of it. That means deciding up front where the tool carries the load and where a person stays in the loop, and saying so to the team honestly. Robert Half’s researchers made the same point: the shift that holds up is from cutting roles to rethinking them, with a clear view of where AI works best alongside people rather than in place of them [1].

So the lesson is steadying, not gloomy. The goal was never a smaller headcount for its own sake. It was good work, done reliably, for real people on the other end. The companies that keep that in view tend to make fewer round trips, and they tend to treat the humans in the loop as part of the design from the start. That is a calmer and, it turns out, a cheaper way to bring these tools on board.

If your own organization has lived some version of this, I would like to hear how it went. I read every response.

Leave a comment

P.S. If a colleague spends their days keeping systems safe, share this newsletter and help brew up stronger customer relationships.

Share TechSips with Nadina: The Newsletter

P.P.S. If you found these AI insights valuable, a contribution to the Brew Pot helps keep the future of work brewing.

Add to the Brew Pot! ✨

Resources

[1] The 'AI boomerang': Why some companies are rehiring employees they laid off due to AI (Fast Company)

[2] Klarna Is Hiring Customer Service Agents After AI Couldn't Cut It on Calls (Entrepreneur)

[3] Gartner Predicts Half of Companies That Cut Customer Service Staff Due to AI Will Rehire by 2027

Sip smarter, every Tuesday. (Refills are always free!)

Cheers,

Nadina

Host of TechSips with Nadina | Chief Strategy Architect ☕️🍵

“Finding a flaw is the easy half. Fixing it is the half that counts.” — Nadina D. Lisbon

Hello Sip Savants! 👋🏾

Much of the conversation about AI and security has focused on what attackers might do with it. Last week offered a more encouraging chapter. On June 22, OpenAI expanded its Daybreak security effort with a model and a program built squarely for defenders, including one aimed at the open-source software almost every company quietly runs on [1]. It is a good moment to think about AI as something that helps us repair, not only something to guard against. Three items worth a few minutes.

3 Tech Bites

• 🛡️ A model tuned to fix, not just find

  • OpenAI released GPT-5.5-Cyber, built to help defenders locate and patch software flaws [1]. It scored 85.6% on the company’s CyberGym test for reproducing known vulnerabilities, up from 81.8% for the standard model, and access runs through a vetted partner program rather than open release [3].

• 🔧 Tending the open-source commons

  • A new initiative called Patch the Planet, founded with Trail of Bits, helps maintainers move from spotting bugs to actually fixing them [2]. More than 30 widely used projects have signed on, including cURL, Go, Python, and Sigstore.

• 🧰 Scanning where the work already happens

  • A new Codex Security plugin brings vulnerability checks into the developer’s normal workflow [3]. The pattern worth noting is the direction: putting the strongest tools in the hands of the people doing the defending.

Subscribe now

5-Minute Strategy

🧠 Thank (or Sponsor) One Maintainer

A surprising amount of enterprise software rests on open-source projects kept alive by a handful of volunteers. With maintainers in the spotlight lately, here is a small, uncommon move that carries real goodwill and resilience:

1. Ask your engineering lead which single open-source project your stack leans on most.

2. Look up who maintains it and how it is funded, which is often a very small team.

3. Pick one supportive step: a modest recurring sponsorship, or a short note offering help.

4. While you are there, note whether the project has a security disclosure process you could plug into.

5. Send the message, or start the sponsorship.

The software you depend on is maintained by people, and a little support tends to go a long way.

Share

1 Big Idea

💡 Tools That Point Toward Repair

For a couple of years, the security story around AI has leaned toward worry. If a model can find a weakness, the thinking goes, then someone will misuse it. That concern is fair and worth taking seriously. What is easy to miss is that the same capability, pointed the other way, is unusually good news. A tool that can find a flaw can also help close it, and the latest news put that second use front and center [1].

The piece I keep coming back to is the focus on open-source maintainers [2]. So much of the modern world runs on libraries written and looked after by small groups of people, often unpaid, often stretched thin. They carry a remarkable load on behalf of everyone else. Aiming capable tools at helping them move from a long list of findings to a manageable set of fixes is a quietly humane idea, and it strengthens the foundations the rest of us build on.

There is also something worth noticing in how it was released. Rather than handing the strongest version to anyone who asked, the defensive model is reaching people through a vetted program [3]. You can debate the details, but the instinct behind it, matching access to responsibility, is a thoughtful one. It treats capability as something to steward rather than simply to ship.

For leaders, none of this calls for a dramatic response. It is more of an invitation to update a mental model. Security is not only a wall you stand behind. It is also ongoing maintenance, much of it shared, much of it done by people you will never meet. Tools that help that work get done are worth welcoming, and worth supporting where you can.

The hopeful read here is simple. The most capable tools we are building can be turned toward tending and repairing the things we all rely on. That is a future worth leaning into, and a reminder that progress and care are not at odds.

I would love to hear how your teams are thinking about AI on the defensive side of the ledger. I read every response.

Leave a comment

P.S. If a colleague spends their days keeping systems safe, share this newsletter and help brew up stronger customer relationships.

Share TechSips with Nadina: The Newsletter

P.P.S. If you found these AI insights valuable, a contribution to the Brew Pot helps keep the future of work brewing.

Add to the Brew Pot! ✨

Resources

[1] Daybreak: Tools for securing every organization in the world (OpenAI)

[2] Patch the Planet: a Daybreak initiative to support open source maintainers (OpenAI)

[3] OpenAI Expands Daybreak With GPT-5.5-Cyber to Help Defenders Patch Security Flaws (The Hacker News)

Sip smarter, every Tuesday. (Refills are always free!)

Cheers,

Nadina

Host of TechSips with Nadina | Chief Strategy Architect ☕️🍵

“The best systems aren’t the most powerful ones. They’re the ones whose limits you can actually see.” — Nadina D. Lisbon

Hello Sip Savants! 👋🏾

The Senate Judiciary Committee recently advanced the bipartisan NO FAKES Act [1], a sign that the rules around digital identity are starting to firm up. As that happens, it’s a good moment to think about trust as something built into the architecture from the start, rather than added once a problem appears. The three items below are worth a few minutes of your attention this week.

3 Tech Bites

• 🎙️ Digital likeness, getting its guardrails

  • The NO FAKES Act would create federal protections against nonconsensual AI-generated voice and video replicas [1]. If any of your deployments touch synthetic media, it may be worth looking at identity verification early, while the standards are still taking shape.

• 🛡️ A vulnerability worth knowing about

  • Researchers at the Cloud Security Alliance have documented what they call agentjacking, a way autonomous systems can be compromised through weaknesses in the Model Context Protocol [2]. It’s a useful prompt to review how your agents’ context and access permissions are scoped, before they’re handling anything sensitive.

• 🧬 One model, or several working together

  • OpenRouter’s Fusion approach lets different large language models combine their outputs [3]. A composable setup like this can ease reliance on any single vendor and let you pair models to the work they’re best at, rather than forcing everything through one.

Subscribe now

5-Minute Strategy

🧠 Putting Agentic Risk on the Map

When you delegate work to autonomous agents, some risk moves out of plain sight. One simple step for technology leaders is to give that risk a home on the books. Think of it as a conversation starter for your risk and security teams, not a finished prescription:

1. Open a draft entry in your IT risk register.

2. Name the category in your own terms. For example, “Autonomous logic compromise via Model Context Protocol” [2].

3. Note a starting impact rating (many teams would place this in the high or critical range) and adjust to your context.

4. Sketch the core concern: unauthorized changes to an agent’s context that could lead to privilege escalation, data exposure, or unapproved third-party calls.

5. Save it and route it to your risk and compliance teams for a proper look.

The point isn’t the exact wording. It’s making the risk visible enough to discuss.

Share

1 Big Idea

💡 Building Glass Houses

Moving to autonomous AI asks for a different kind of design thinking. Earlier automation ran on predictable, linear workflows where you could trace every step. Agents don’t work that way. They make decisions in the moment, which is exactly what makes them useful and exactly what makes them harder to supervise. The interesting design question isn’t how to remove people from the loop, but where to place them so their judgment still counts.

Model fusion [3] is a small example of the shift. Instead of leaning on one provider for everything, you can draw on several models and let their strengths combine. The result tends to be more accurate and less prone to any single model’s blind spots. That’s a reminder that mixed, well-integrated systems are usually more resilient than isolated ones. Used well, the goal is to give people richer information to decide with, not to hand the decision over entirely.

That openness only works with clear boundaries around it. The agentjacking research [2] is a good illustration of what happens when autonomous logic runs without tight limits on what it can see and touch. A continuously-verified, least-access approach lets you keep the flexibility without leaving the door open, and protects the organization as carefully as it processes the work.

Emerging rules like the NO FAKES Act [1] point the same direction. Digital identity and creative ownership are becoming things systems are expected to respect by design. Trust is shifting from a soft value to a concrete design requirement, something you can point to in the architecture, meant in the end to protect the actual people behind the data.

The aim is to build systems people can see into. When the infrastructure is secure, predictable, and legible, the people working alongside it are freed to focus on the work that needs them. That’s the case for glass houses: environments where the logic is visible, the boundaries are clear, and the technology stays a dependable foundation rather than a black box.

Hit reply and start a conversation about what most intrigued you, these are better questions to think through together than alone. I read every response.

Leave a comment

P.S. If a peer is working through their own autonomous AI plans, share this newsletter and help brew up stronger customer relationships.

Share TechSips with Nadina: The Newsletter

P.P.S. If you found these AI insights valuable, a contribution to the Brew Pot helps keep the future of work brewing.

Add to the Brew Pot! ✨

Resources

[1] Blackburn, Coons Bipartisan Bill to Protect Individuals and Creators from Deepfakes Passes Senate Judiciary Committee
[2] Agentjacking: Sentry MCP Injection Hijacks AI Coding Agents

[3] What Is Model Fusion? How OpenRouter Fusion Matches Frontier AI at Half the Cost

Sip smarter, every Tuesday. (Refills are always free!)

Cheers,

Nadina

Host of TechSips with Nadina | Chief Strategy Architect ☕️🍵

“Technology sets the pace, but our shared human values must set the direction.” — Nadina D. Lisbon

Hello Sip Savants! 👋🏾

The global map of artificial intelligence is actively being redrawn. A sudden 90-minute White House regulatory deadline has triggered fierce debates across Silicon Valley, emphasizing a massive shift in how we govern innovation [1]. Amidst this geopolitical chess match, major tech and media players are scrambling to adapt and maintain public trust [2, 3].

3 Tech Bites

• 🏛️ White House Shockwave

  • A surprise 90-minute regulatory deadline from Washington triggered massive debates across Silicon Valley [1]. This sudden mandate highlights the escalating tension between rapid algorithmic innovation and the urgent demand for strict government oversight.

• 🌍 Global AI Perceptions

  • A recent Politico report reveals international audiences no longer automatically view the United States as the undisputed AI champion [4]. Other nations are rapidly closing the gap, proving that global trust requires more than just venture capital funding.

• ⚖️ The Meta Burden

  • One year after Meta spent over $14 billion to hire Alexandr Wang and his team to build a proprietary AI model, the pressure has shifted [2]. It is now up to Mark Zuckerberg to convince skeptical developers and investors that this massive gamble was actually worth it.

Subscribe now

5-Minute Strategy

🧠 The AI Transparency Disclaimer

Take control of your daily workflow with this quick audit.

1. Identify the primary channel where you share AI-assisted work, such as client emails or internal project reports.

2. Draft a standardized, one-sentence disclosure that highlights the human element. For example: "Initial data compiled by AI; verified and analyzed by [Your Name]."

3. Add this to your email signature or report template right now. This instant change builds trust, demystifies your workflow, and ensures human accountability is always front and center.

Small steps toward transparency build lasting team trust.

Share

1 Big Idea

💡 The Innovation vs. Accountability Crisis

The AI landscape has fundamentally fractured into two distinct camps: those building the architecture and those responsible for selling its value. A year after Meta brought on Alexandr Wang to lead their Superintelligence Labs, the dynamic is painfully clear [2]. Engineering a massive proprietary model is only the first step. The true challenge lies in convincing a skeptical public and cautious investors that the technology actually serves a meaningful purpose.

This divide extends far beyond Silicon Valley boardrooms. We see the exact same tension playing out on the geopolitical stage. A sudden 90-minute White House deadline has forced tech giants to confront the reality of government oversight [1]. Innovation is no longer moving in a vacuum. It is constantly colliding with human-centric concerns regarding safety, ethics, and global dominance.

Meanwhile, the rest of the world is no longer waiting for the United States to dictate the future. International audiences are actively searching for new leaders in the AI space, proving that technological superiority cannot rely on legacy reputation alone [4]. Public trust is shifting toward entities that prioritize transparent, culturally sensitive, and safe deployment over sheer computational power.

Even traditional media institutions are caught in this rapid current. As executives attempt to steer legacy media through a world that outpaces traditional editorial cycles, the value of human critical thinking becomes increasingly obvious [3]. Algorithms can generate content at lightning speed, but they cannot replace the nuanced, empathetic lens of a seasoned journalist or a discerning leader.

Ultimately, the companies and nations that succeed will not be the ones with the most expensive models or the highest volume of code. The winners will be those who actively bridge the gap between cold algorithms and warm human connection. True innovation requires us to prioritize accountability just as highly as capability.

Building trust takes time, but losing it takes seconds? Hit reply and start a conversation about AI accountability today. I read every response.

Leave a comment

P.S. If the number terrifies you, share this newsletter and help brew up stronger customer relationships!

Share TechSips with Nadina: The Newsletter

P.P.S. If you found these AI insights valuable, a contribution to the Brew Pot helps keep the future of work brewing.

Add to the Brew Pot! ✨

Resources

[1] How a 90-minute White House deadline sparked Silicon Valley's biggest AI fight

[2] A year after Meta tapped Alexandr Wang to build a new AI model, Zuckerberg has to sell it

[3] How to Run a News Company in the Age of Polarization and A.I. Slop

[4] People around the world see a winner on AI — and it’s not the US

Sip smarter, every Tuesday. (Refills are always free!)

Cheers,

Nadina

Host of TechSips with Nadina | Chief Strategy Architect ☕️🍵

“Speed was the strategy. Now the bill has arrived.” — Nadina D. Lisbon

Hello Sip Savants! 👋🏾

We spent the last three years completely obsessed with AI latency and raw capability. Now, the utility bills and security audits are finally landing, and enterprise architectures are beginning to buckle under the weight. With the recent geopolitical push to secure frontier AI models against rising cyber threats [1], we are simultaneously hitting a massive physical wall. The energy and financial toll of maintaining these systems is becoming unsustainable, and it is time to pause and ask what we are actually paying for this intelligence.

3 Tech Bites

• 🌐 The Frontier Security Race

  • As superpowers race for dominance, securing massive models against cyber threats is a national security imperative shaping future enterprise AI regulations [1]. The downstream impact shrinks open-source flexibility, creating compliance bottlenecks for architectures relying on unregulated external models.

• ⚡ Bring Your Own Power

  • Ireland is demanding tech titans supply their own power for new data centers, highlighting the severe GPU and cooling demands straining global infrastructure [2]. The reality proves infinite cloud scalability is a myth. Relying entirely on massive external clusters places your uptime at the mercy of strained local grids.

• 💸 Runaway Token Costs

  • Enterprises are experiencing severe sticker shock as API bills materialize. Managing this spend forces a shift from universal AI adoption to highly targeted, ROI-driven deployments [3]. Unchecked background API calls will drain an enterprise budget faster than human headcount ever could.

Subscribe now

5-Minute Strategy

🧠 Audit Your AI Spend Before It Audits You

Companies are going broke because no one is tracking what AI is doing. Run this five-minute check with your team today:

1. List every single AI tool your team touches, including coding assistants and internal agents, especially those on auto-renew.

2. Verify if usage limits exist because most enterprise plans support them. Set them immediately if they do not.

3. Identify who is consuming the most tokens, and contrast that with what they are actually shipping. High token spend is only justified if the output is measurable.

4. Hunt down your forgotten agents. Background automation tasks are often the largest cost drivers and the hardest to spot.

5. Flag your next renewal date, keeping in mind that major tools have recently repriced.

The Tokenomics Foundation guidance is clear: the best ROI comes from moving the broad middle of your team from low to moderate usage, rather than pushing heavy users higher.

Share

1 Big Idea

💡 Sustainable Architecture in the Age of Gigawatt AI

We are reaching an inflection point where the digital realm directly collides with the physical world. For years, the cloud felt abstract, limitless, and invisible. The reality of AI is decidedly heavy. It requires concrete data centers, millions of gallons of water for cooling, and gigawatts of electricity. When entire nations like Ireland halt tech expansions because power grids are maxing out [2], we must face the fact that unchecked cloud computing is a resource crisis disguised as innovation.

As enterprise strategists, our obsession has historically been with capabilities. We constantly ask how fast a model can think or how natural its prose sounds. A true human-in-the-loop philosophy requires us to look entirely beyond the prompt and the immediate output. It demands that we consider the holistic, real-world impact of our technological choices. If a system displaces human oversight or bankrupts a startup with runaway token costs [3], it fails the most basic test of architectural integrity.

This tension is further complicated by the geopolitical race to secure frontier AI [1]. The fear of falling behind drives massive investment and reckless deployment. Security and speed are prioritized over sustainability and ethical scaling. We are building digital fortresses but completely ignoring the physical foundation they sit on. A secure data architecture is ultimately a liability if the infrastructure required to power it is financially and physically unsustainable.

The solution requires a fundamental architectural shift. We must decouple our daily workflows from massive external constraints. We need to normalize the concept of good enough AI by aggressively shifting specific enterprise tasks to local hardware. Leveraging robust Network Attached Storage (NAS) setups and offline open-source tools for background automation allows us to regain control. We must stop bleeding tokens to massive frontier models for minor, repetitive tasks that can be handled easily on the edge.

Ultimately, the most resilient enterprise systems of the next decade will not just be the ones that are technically sophisticated. They will be the systems that are physically and financially sustainable. By prioritizing human-centric design that keeps computing power strictly proportional to the business problem, we can build architectures that respect the finite limits of our physical grid just as much as the processing speed of our models. The goal for the next decade is architectural equilibrium.

Pull up your API billing dashboard today. What percentage of your token spend is currently untracked background automation? Hit reply and let me know if the number terrifies you. I read every response.

Leave a comment

P.S. If the number terrifies you, share this newsletter and help brew up stronger customer relationships!

Share TechSips with Nadina: The Newsletter

P.P.S. If you found these AI insights valuable, a contribution to the Brew Pot helps keep the future of work brewing.

Add to the Brew Pot! ✨

Resources

[1] ‘It’s a hurricane warning’: Guardrails around powerful AI models may be too late

[2] Bring Your Own Power, Ireland Tells Tech Titans Hungry for Data Centers

[3] The token bill comes due: Inside the industry scramble to manage AI’s runaway costs

Sip smarter, every Tuesday. (Refills are always free!)

Cheers,

Nadina

Host of TechSips with Nadina | Chief Strategy Architect ☕️🍵

“We are trading physical carbon for digital convenience at a rate the earth's grids can't sustain, forgetting that the most efficient neural network still sits right between our ears.” — Nadina D. Lisbon

Hello Sip Savants! 👋🏾

This week kicked off with major announcements from Computex 2026, showcasing incredibly powerful new AI hardware [1]. Yet beneath the glossy product keynotes lies an escalating friction. Silicon Valley is hitting a hard physical ceiling of energy consumption [2], while simultaneously shifting its human balance sheet through structural layoffs [3]. We need to look past the interface and talk about the infrastructure.

3 Tech Bites

• 💻 The Edge Acceleration Race

  • Computex 2026 proved that the next battlefield is local silicon [1]. Tech giants are rushing to embed dedicated AI processing units directly onto consumer devices to bypass cloud latency.

• 🔋 Gridlock at the Data Center

  • Wall Street is pouring capital into specialized infrastructure stocks as data centers face severe power and cooling shortages [2]. The physical weight of the cloud is officially bottlenecking software deployment.

• ✂️ The Silicon Restructure

  • The New York Times confirmed a sharp escalation in tech layoffs specifically driven by automated restructuring [3]. Companies are aggressively shifting budget allocations from engineering headcount to raw compute power.

Subscribe now

5-Minute Strategy

🧠 The AI Technical Debt Audit Do not wait for a corporate restructuring to evaluate how algorithmic dependencies impact your work. Spend 5 minutes diagnosing your operational pipeline to ensure you are building structural, irreplaceable value.

1. Map Your System Dependencies: Identify where your workflows rely on external LLM calls or automated code generation. Are you using a 70-billion-parameter model to perform basic regex or data parsing that local, deterministic code could execute?

2. Isolate the Strategic Core: Pinpoint the exact architectural inflection points in your current projects where standard patterns fail. This is your baseline. These are the highly contextual, ambiguous edge cases that require historical system knowledge and human negotiation.

3. Optimize the Local Stack: Move your iterative prototyping, draft refinement, and small-scale testing to local language models running fully on your own machine’s hardware.

By shifting your standard processes to local environments and reserving heavy cloud models exclusively for high-leverage problems, you minimize external data liabilities while sharpening the precise architectural skills that automated systems cannot replicate.

Share

1 Big Idea

💡 The Irony of the Automated Balance Sheet

When we talk about the cloud, the vocabulary we use is intentionally deceptive. We use terms like “serverless” and “virtualization” to make our digital infrastructure feel as light as mist. The physical reality of the AI boom is entirely different. It is heavy metal, copper wiring, massive diesel backup generators, and millions of gallons of water cycled away to keep silicon plates from melting under the load of our inference prompts [2]. Recent analysis indicates that global data center energy consumption is rapidly accelerating to unprecedented and unsustainable levels [5].

To fund these massive infrastructure bills, corporate boards are making a short sighted trade. They are cutting traditional engineering headcount and entry level roles to reallocate capital directly into raw compute power [3]. They are essentially betting that software capacity can be bought wholesale via a cloud subscription rather than nurtured through human talent.

But this is where the corporate logic completely fractures. Massive language models are incredibly inefficient and financially draining to run at scale. Companies that tried to blindly replace human workflows with automated API calls are suddenly staring at unsustainable cloud bills and broken edge cases. They are discovering that you cannot automate a complex system without creating a massive wave of technical debt.

This has triggered a fascinating and ironic reversal. Organizations are now scrambling to hire and retain specialized engineers to fix the efficiency crisis. They do not need generalists anymore. They need architects who can move models to the local stack [1], design deterministic fallback systems, and optimize the heavy code patterns that are currently draining the corporate budget. The very models designed to reduce headcount now require a new army of infrastructure specialists just to keep them running sustainably.

We must change the narrative from total automation to human centered orchestration. True systemic elegance means building architectures that respect physical and financial limits rather than ignoring them. If we destroy our engineering pipeline today to pay for a temporary compute spike, we will be left with complex, brittle systems that nobody understands how to maintain from scratch.

If this edition resonated with you, hit reply or leave a comment and let me know. I read every response.

Leave a comment

P.S. If you are tired of pretending the cloud is weightless, share this newsletter and help brew up stronger customer relationships!

Share TechSips with Nadina: The Newsletter

P.P.S. If you found these AI insights valuable, a contribution to the Brew Pot helps keep the future of work brewing.

Add to the Brew Pot! ✨

Resources:

[1] Computex 2026: All the news and announcements

[2] 7 Best Data Center Stocks, ETFs and REITs to Buy

[3] Is A.I. Replacing Tech Workers or Providing an Excuse for Job Cuts?

[4] International Energy Agency. Data Centres and Data Transmission Networks

[5] Global energy demands within the AI regulatory landscape

Sip smarter, every Tuesday. (Refills are always free!)

Cheers,

Nadina

Host of TechSips with Nadina | Chief Strategy Architect ☕️🍵

“True systemic resilience isn’t found in replacing human judgment. It is found in architecting the exact boundaries where it remains irreplaceable.” — Nadina D. Lisbon

Hello Sip Savants! 👋🏾

Just this week, Merlin Labs began testing autonomous flight communication systems to mitigate global pilot shortages [1]. Simultaneously, internal findings from tech giants revealed a stark operational reality. The compounding computational overhead of complex AI agents means running these systems can actually cost more than employing human professionals [2]. As autonomous architecture scales, leadership faces a massive reality check. Decoupling humans from the operational loop is proving to be financially volatile and culturally risky.

3 Tech Bites

• ✈️ AI in the Cockpit

  • Merlin Labs is deploying natural language processing to automate high stakes aviation radio communications. Federal oversight boards are currently evaluating the system to minimize human error. However, aviation purists maintain that algorithms cannot match an experienced pilot’s instinct during an unprecedented crisis [1].

• 💰 The Token Tax

  • Internal corporate data indicates that the continuous infrastructure required for multi agent LLM orchestrations is skyrocketing. The exponential growth of token context windows makes full automation a massive line item expense. It frequently outprices retaining skilled human talent [2].

• 🤝 The Trust Dividend

  • Consumer touchpoints are oversaturated with automated triage. Market sentiment is steadily shifting back to authenticity. Organizations that intentionally design human in the loop guardrails are seeing a distinct trust dividend. This yields significantly higher client retention and brand authority.

Subscribe now

5-Minute Strategy

🧠 The Token to Talent ROI Triage

Execute this rapid structural audit to ensure your automated systems are not draining your budget or eroding user trust.

• Audit the Token Burn: Isolate your highest frequency AI workflows. Calculate the true compute cost per transaction. Compare it directly to the hourly rate of a human specialist.

• Map the Empathy Threshold: Grade the workflow touchpoints based on emotional stakes. Flag the process for immediate human oversight if an algorithmic hallucination risks alienating a client.

• Decouple and Augment: Rebuild the process. Strip the AI of autonomous execution. Reposition it as an asynchronous drafting assistant. Ensure a human expert remains the final gateway for validation.

Share

1 Big Idea

💡 Reclaiming the Human Core in a Silicon Economy

For the past several years, the prevailing enterprise narrative treated artificial intelligence as a seamless substitute for routine labor. It was sold as a frictionless method to slash overhead. Yet, we are rapidly approaching an architectural tipping point. The economic reality is finally refuting the hype. The staggering computational power required to process massive context windows carries an unprecedented financial premium [2]. When executing a complex model costs more than compensating a human professional, the foundational argument for absolute automation completely collapses [2].

This economic correction invites a much deeper systemic calculation. Algorithmic systems are exceptional at executing deterministic patterns. They easily parse structured datasets. However, they operate entirely divorced from situational awareness and genuine accountability. In a high stakes environment like aviation, an algorithm can process clear air traffic commands. It possesses no biological instinct for survival to navigate a cascading failure safely [1].

When we rush to automate the spaces where human nuance is required, we introduce profound systemic risks. Environmental impacts also compound rapidly. Running these massive GPU clusters drains energy grids at an unsustainable rate. A sovereign architectural mindset requires us to calculate both the financial and ecological burn rates of our technology stack. Relying solely on automated output is a shortsighted strategy.

True operational brilliance is never achieved by completely erasing human presence to optimize a margin. It comes from understanding exactly where the human spirit is structurally necessary. Relationships and trust are inherently non linear. They cannot be neatly compressed into a vector database. An organization that delegates its entire relational identity to silicon agents eventually becomes sterile. It discards the vital friction that sparks genuine innovation.

As we look toward an increasingly automated horizon, smart leaders will pivot. We must deliberately reject the temptation of total displacement. Instead, we should embrace a philosophy of radical augmentation. By directing costly computing power toward mundane data synthesis, we liberate our human teams. The future belongs to organizations that treat AI as an intellectual multiplier. They intentionally structure environments where human judgment remains the definitive authority.

P.S. Are your automated workflows hitting a cost barrier? Hit reply or leave a comment and let me know. I read every response.

Leave a comment

P.S. Share this newsletter and help brew up stronger customer relationships!

Share TechSips with Nadina: The Newsletter

P.P.S. If you found these AI insights valuable, a contribution to the Brew Pot helps keep the future of work brewing.

Add to the Brew Pot! ✨

Resources:

[1] AI is learning to fly airplanes — and aviation is starting to embrace it

[2] Microsoft reports are exposing AI’s real cost problem: Using the tech is more expensive than paying human employees

Sip smarter, every Tuesday. (Refills are always free!)

Cheers,

Nadina

Host of TechSips with Nadina | Chief Strategy Architect ☕️🍵

“We are pouring billions into digital concrete while ignoring the human architecture. Automating the entry level destroys the pipeline of future leaders.” — Nadina D. Lisbon

Hello Sip Savants! 👋🏾

The market just witnessed a monumental shift as Google and Blackstone forged a 5 billion dollar alliance to construct massive artificial intelligence cloud environments [2]. It is a breathtaking bet on raw computing power, yet it exposes a glaring structural flaw across the enterprise landscape. We are furiously building the engines of tomorrow while higher education systems fail to train the individuals required to operate them [1]. As we accelerate, executives must shift their focus from buying raw compute to actively architecting the human capability required to guide it.

3 Tech Bites

• ☁️ The 500-Megawatt Power Play

  • Google and Blackstone are launching a joint venture backed by a 5 billion dollar initial equity investment to scale artificial intelligence cloud infrastructure. The partnership targets bringing 500 megawatts of data center capacity online by 2027 to feed insatiable enterprise compute demands [2].

• 📉 The Higher Education Chasm

  • A widening disconnect has emerged between university curricula and the real-world skills required by modern digital workplaces. Reports indicate that traditional collegiate programs are failing to adapt to rapid technological evolution, shifting the burden of technical training entirely onto corporate upskilling initiatives [1].

• ⚖️ Guardrails as a Growth Engine

  • Friday Harbor has secured the first artificial intelligence governance compliance attestation from Brody Gapp for a mortgage technology provider [3]. By subjecting their pre-underwriting platform to strict third-party fair lending and model risk audits, they unlocked a 35 % efficiency boost while ensuring total regulatory safety [4].

Subscribe now

5-Minute Strategy

🧠 The Core Competency Circuit Breaker

Instead of running a complex, time-consuming audit, protect your team from the de-skilling trap right now by answering four clear questions:

1. The Automation Focus: What is one entry-level task in your department that has been completely handed over to an artificial intelligence tool? (such as initial code drafting, basic data synthesis, or preliminary report writing)

2. The Knowledge Source: How did your current senior experts originally master this domain? If they learned it by doing the exact manual grunt work that is now automated, you have found a training vulnerability.

3. The Learning Shift: What manual verification step can you build back into the daily routine? For example, require junior team members to reverse-engineer and document the logic behind one automated output per day.

4. The Evaluation Metric: Do your performance reviews reward the speed of automated output, or do they actively reward deep system design, troubleshooting, and ethical validation?

Share

1 Big Idea

💡 The De-Skilling Conundrum: The Hidden Extinction of Human Judgment

When the tech industry celebrates a 5 billion dollar investment in data centers, the immediate assumption is that the organization with the most processing units wins [2]. We treat computing power like physical real estate, racing to claim land and secure megawatts of electricity. Yet, this infrastructure rush masks a deeper, structural vulnerability. As traditional higher education lags behind real-world artificial intelligence demands [1], enterprises face a hidden operational hazard: the slow erosion of foundational human judgment.

If algorithmic models take over data synthesis, preliminary logic, and routine technical drafting, the entry-level tasks where junior talent traditionally develops critical thinking disappear. We are systematically removing the cognitive miles required to transform a novice into an expert. If a junior architect never spends time wrestling with basic system configurations because a machine generates them instantly, how will they ever develop the deep intuition required to debug a catastrophic systemic failure a decade from now?

This reality highlights the massive gap currently opening up within our talent pipelines. We are building sophisticated cognitive infrastructure while relying on an outdated educational framework that treats workers like factory components [1]. If an enterprise system is designed to handle automated data synthesis, the value of the human worker must pivot toward deep ethical reasoning, complex problem-solving, and system design. You cannot run a cutting-edge cloud infrastructure with a workforce trained for legacy methodologies.

The current corporate playbook is dangerously lopsided. Billions are allocated to software licenses and infrastructure upgrades, while human upskilling budgets receive mere fractions. This imbalance creates a fragile ecosystem where powerful tools are handled by unprepared teams, leading directly to compliance failures and operational drift.

How will your organization cultivate its next generation of leadership when the bottom rungs of the professional ladder have been fully automated? We must challenge the assumption that faster output equals smarter talent. True operational resilience belongs to those who design internal training paths that actively force human critical thinking to evolve alongside the machine, ensuring we maintain a workforce capable of steering the technology safely.

P.S. If you are ready to protect your team’s long-term talent pipeline from the trap of over-automation, forward this newsletter to a fellow leader today.

Leave a comment

P.S. Share this newsletter and help brew up stronger customer relationships!

Share TechSips with Nadina: The Newsletter

P.P.S. If you found these AI insights valuable, a contribution to the Brew Pot helps keep the future of work brewing.

Add to the Brew Pot! ✨

Resources:

[1] As more jobs demand AI skills, some colleges may fall short in prepping students: ‘Why would we train them using the skills of yesterday?’

[2] Google, Blackstone to launch AI cloud venture to meet data centre demand

[3] Friday Harbor Receives First Mortgage AI Governance Attestation

[4] NewFed Boosts Loan Volume, Cuts Cycle Times With Friday Harbor AI

Sip smarter, every Tuesday. (Refills are always free!)

Cheers,

Nadina

Host of TechSips with Nadina | Chief Strategy Architect ☕️🍵