AI Rules, Big Settlements, and Smarter Materials
Dear Friend,
In our AI Smart Lab session this week, we pulled up our own AI usage and found we had burned through tens of thousands of credits in ClickUp. One agent was responsible for most of it. Our task intake triage agent.
Nobody had flagged it, because it was working. It was just firing far more often than anyone intended, and doing the same job as other triage workflows we had already built. We are tracking down how often it actually triggers before we add a single new agent on top of it.
Did you have any automations going off the rails? The bill is rarely where you think it is.
Why Most SMB AI Rollouts Quietly Die
If you have ever bought a tool your team never opened, this one is for you. The software is almost never the problem. What breaks is buy-in, training, and the human side of the rollout, and that is the part most companies skip entirely.
Andrea walks through what actually happens when a growing business tries to adopt AI, keep operations running, and upskill staff at the same time with the people it already has. It covers how to prepare your team before a consultant ever walks in, what real hands-on training looks like, and how to measure the return once you start.
Our Most Requested Service: AI Strategy and Consulting
The question clients bring us right now is almost always the same. They know AI can help. They do not know which tools to trust, how to protect their data, or how to get the team to actually use all of it.
We start with your workflows, not with software. We map where the time is spent, then build a phased roadmap with secure tools and clear owners, and we train your team to run it after we leave.
One client, a medical training company, worried the whole way through that AI meant cuts. Nobody was laid off. Every person was upskilled and moved into higher-leverage work, and the hands-on training led to automating 99% of their customer service operations, worth $177,000 in combined savings and repurposed labor. We have trained more than 20,217 professionals this way.
This week’s stories
Alabama Subpoenas OpenAI Over a Model That Hacked Another Company
Alabama’s attorney general has sent OpenAI a subpoena. In July, one of OpenAI’s test models escaped its sandbox and spent days hacking Hugging Face, a platform where AI models and data sets live. The state wants to know whether OpenAI’s handling of it broke Alabama’s consumer protection laws. Fourteen other state attorneys general had already written to OpenAI asking it to preserve every record tied to the incident. In plain terms, a state consumer protection law is being used to hold an AI company responsible for something its software did without a human telling it to. Well sort of It went outside of the sandbox doing the thing it was told to do only in the sandbox.
My Take: The real rulebook in the US is being written by 50 state attorneys general who do not need a new AI law to act, because deceptive trade practices statutes already cover you.
That is a harder compliance problem than one federal standard, not an easier one. We give every client our own AI use policy to adapt rather than starting from a blank page, because the first question in any of these investigations is what your written rules said, and “we did not have any” is the worst possible answer.
Need help with your AI policies or protection protocol, we can help, [email protected]
Source: Alabama Attorney General’s Office
Meta Will Pay Up to $16.68 Billion Over How It Designed Its Apps
Twenty-nine states claimed Meta built Facebook and Instagram to be addictive to children and misled people about how safe they were. Meta settled mid-trial for up to $16.68 billion and denied any wrongdoing. And if you know anything about Meta and their efforts, they absolutely design it to be addictive and are well aware of the Compulsion Loop that keeps people scrolling and coming back.
The money is only part of it. Meta also has to add daily time limits and nighttime blocks for teens, strengthen age checks, and accept an independent auditor who reports to the states.
My Take: Meta can absorb $16.68 billion. What it cannot undo is a court telling it how to design its own product and installing someone to watch.
The precedent here is not about children or social media. It is that the choices inside your system are now the evidence. In AI Smart Labs this week we went through every agent we run and asked what each one was doing and why, and for a couple of them the honest answer was that somebody built it months ago and nobody had revisited it. That answer is fine in a report. It is not fine in a deposition.
Source: Bloomberg
Oracle’s AI Now Writes the Medical Bill, Not Just the Notes
Oracle Health added three things to its Clinical AI Agent, now live for US customers: automated billing codes, dictation, and chart review. The agent listens to the visit and suggests the billing codes, and a clinician confirms each one before it goes out.
Oracle says note generation alone has saved doctors more than 400,000 hours in about two years. That figure doubled in roughly six and a half months.
My Take: Note-taking AI was safe to pilot because a clumsy summary is embarrassing and nothing more. Coding is different. Get it wrong and you have either left money on the table or billed for something that did not happen, and one of those has a name in federal law.
Watch how fast “the clinician confirms each recommendation” turns into clicking approve 40 times a day. That is the actual risk, and it is a workflow problem, not a model problem. When we build systems for clients, we put a human review step on every AI output before it leaves the building, and we design that step so it is possible to fail it. Right now I’m the human reviewing this newsletter because I sign my name to it and I do make edits AND using AI is still a massive timesaver. We cannot get lazy in our reviews as I hear from many founders is happening.
Source: Oracle
AI Can Now Design Materials That Actually Work
Anyone with a big enough model can generate millions of new material designs in minutes. Almost none of them are chemically stable, so they are useless in the real world.
Researchers built a tool called CrysVCD that applies the chemistry rules first, during generation, instead of screening out the failures afterward. Nearly 70% of what it produces now passes a strict stability test, and it plugs into models that already exist.
My Take: For two years the entire AI pitch has been volume: more drafts, more designs, more variations, faster. Volume was never the constraint. Usable output was.
Constraints applied at the start filtering at the end, every time. That is exactly what our credit problem in Labs turned out to be. We were not generating too little, we were generating plenty that nobody used, and the fix was narrowing what the agent was allowed to do rather than adding a review layer behind it. I would take a tool that produces 10 things I can ship over one that produces 1,000 I have to sort.
Source: Nature Computational Science
Amazon Is Closing the Human Work Platform Behind Early AI
Amazon will shut down Mechanical Turk on September 30, 2026, after 21 years. It matched people with tiny digital jobs, labeling data, transcribing audio, answering surveys, usually for a few cents each. At its peak it served more than 500,000 workers.
SageMaker Ground Truth and Amazon Augmented AI close on the same day. Jeff Bezos once called the service artificial artificial intelligence.
My Take: The human work did not go away. It got repriced and hidden. Data labeling moved to companies that sell it as a platform, so the people are still there, they are just behind someone else’s invoice now instead of visible on a marketplace you could browse.
“Our AI does this automatically” is doing a lot of quiet work in vendor demos right now. When we scope an AI build for a client, we say out loud where the human is in the loop, ours or theirs, and what that person costs. Every AI system has a human in it somewhere, or should. The only question is whether you know where.
Source: CNBC
Tool picks of the week
- AI Use Policy: Our own AI use policy, free for you to reference or adapt. With state attorneys general now investigating AI companies under consumer protection law, having your rules written down stopped being optional.
- Gamma: Creates presentations, documents, and websites quickly. I use it for every presentation that I need to do, it’s so good, fast, and good looking!
- Yoodli: Private, real-time AI coaching for your communication skills.
Got a burning question, a fresh take, or just want to share your latest AI wins? Hit us up at [email protected]. Your insights keep this community growing and thriving!
Have a great day!
-Nicole A. Donnelly
Founder, AI Smart Ventures
AI Strategy – AI Training – AI Consulting – AI Implementation
