AI Prices, Health Costs, and a Big Copyright Call
Dear Friend,
In our AI Smart Lab session last week, we pulled up our Snitcher numbers and read them line by line. Snitcher is a tool that sits on your site and lets you know who’s visiting and what their contact info is. If properly tagged you can put them into an email funnel based on their browsing actions. 249 contacts had finished a sequence and 59 were still moving through one. The higher education and consulting sequences were opening emails at around 50%, with clicks to match. On paper, that looks like a win.
. People were opening, they were clicking, and they were not booking a call. So, now it’s time to review our CTAs and offerings on the pages where all of the “interested” people are landing and send more traffic there to test again. ..
Numbers at the top can hide a broken step underneath it. Look at where people stop, not just where they start.
Why Your Team Is Pushing Back on AI
If your AI rollout has stalled, the tool is probably fine. The job design is not. This piece digs into why people resist AI at work, and the answer is almost never a skills gap.
It is identity. The person asked to approve the rollout is often the person whose role shrinks because of it. The article walks through the four fears that drive most pushback, why silence from leadership makes all of them worse, and the two questions every announcement should answer before the tool ever shows up. Read the full text.
Our Most Requested Service: AI Strategy and Consulting
Almost everyone coming to us right now has already bought the tools. They have ChatGPT seats, a note taker, and maybe an automation or two running in the background. What they do not have is a plan for which work should run through AI, who checks the output, and what happens to the roles that change.
So we start with the work, not the software. We map how a team actually spends its week and find the steps where AI saves real hours. We put a named person on every output that reaches a customer. Then we phase the rollout so your team learns one thing at a time instead of ten at once.
We also lay out the running cost before anyone commits. When we build a client’s tools budget, every subscription goes in a month-by-month grid so nobody gets surprised in month seven by a bill they did not plan for.
This week’s stories
US health plans are getting more expensive, and AI billing is part of the reason
Marsh released early results from its survey of more than 1,800 US employers. Health benefit cost per employee is expected to rise 8.2% in 2027, the biggest jump since 2003. Employers said their current plans would cost 11% more if they did nothing at all.
Marsh named AI-powered claims software as one of the drivers. Doctors’ offices are using it to submit more claims, and higher-level claims, than they did before. 59% of employers plan to cut costs somewhere in their benefits next year, and about two thirds of employers with 500 or more workers expect to raise the employee share of the premium.
My Take: The first corner of healthcare to make real money from AI was not diagnosis. It was billing.
Every AI success story assumes the savings stay with whoever bought the software. This is the proof that they do not. A clinic bought a tool, and your premium went up. I would stop reading AI efficiency claims as good news for everyone and start asking who is sitting on the other side of that transaction, because any AI that helps one party argue harder over money shows up as a cost to the other one.
Source: Marsh
The US government told a court that AI training is fair use
On September 1, the Justice Department filed a statement of interest in the copyright cases against OpenAI. It told a New York federal judge that training large language models on copyrighted writing counts as fair use. The government argued that ruling the other way would hurt American innovation and national security.
This is the first time the federal government has taken a side on the merits in any of the copyright suits against AI companies. A statement of interest lets the government state its position without joining the case, and it does not bind the judge. Judge Sidney Stein still decides.
My Take: The part that impacts you is smaller and potentially permanent.
Whatever you publish is training data, and the federal government has now said so in writing. Suing was never a realistic option for a business your size anyway, so the useful question flips. It stops being how do I keep my work out of these models, and becomes how do I get credited inside their answers.
That is why we build client content to answer real questions directly rather than chase rankings. A page that plainly answers the question a person asked is the page that gets quoted back. Time to move the budget you were mentally holding for legal protection into being the source these systems cite.
Source: U.S. District Court, Southern District of New York
Anthropic cut the cost of running AI agents by 75%
Anthropic released Claude Fable 5.1 on September 1. Regular pricing did not move, at $10 per million input tokens and $50 per million output tokens. What changed is the price of reading cached input, which fell from $1.00 to $0.25 per million tokens.
Anthropic says that makes typical work about 25% cheaper, and work that runs long agent tasks up to 45% cheaper. Cached input is what a tool re-reads over and over: your documents, your instructions, your running conversation.
My Take: Look at what they cut and what they left alone. Base prices held. Cached reads dropped by three quarters.
That is not a discount, it is a bet. Anthropic is pricing to pull you toward software that runs for hours re-reading the same material, because that is where it thinks the next few years live. The headline everyone will repeat is that AI got cheaper. The accurate version is that one specific way of using AI got much cheaper, and it happens to be the way that embeds deepest in your operations and is hardest to walk away from later.
erun your own AI cost math this month. Something you priced out as too expensive in July may not be anymore.
Source: Anthropic
OpenAI says its next model can find and use security holes on its own
OpenAI published a post saying its coming model, Astra, is the first to reach the “Critical” level for cybersecurity in the company’s own safety framework. Astra scored 100% on ExploitBench, a test of building working attacks from known flaws.
In expert testing, it found flaws nobody had reported and chained them together. It built a browser attack that escaped its sandbox and ran commands on the machine, and it combined several flaws in a hardened operating system into full control of that system. OpenAI plans to release Astra soon, with the advanced security features going to a small group of testers first.
My Take: OpenAI published a careful account of how dangerous its own product is, and is shipping it anyway. Read that twice.
The disclosure is real and I give them credit for making it. But a safety framework that has never once stopped a launch is a labeling system, not a brake. The gap between a flaw existing and somebody using it used to be however long a skilled human needed to find it. That number is heading toward zero on both the attacking side and the defending side, and the attackers get there first because they do not have a safety review. We keep an AI use policy that names which tools are approved and who checks their output, and this is exactly the week it earns its keep. I would move patching off the monthly list and onto the weekly one now, well before the defensive versions of these tools reach you.
Source: OpenAI
MIT built AI that explains its decisions while it makes them
Researchers at MIT and the self-driving company Motional built a method called CW-Net. It makes a car’s decision-making model explain itself in plain words, like “approaching stopped vehicle” or “close to cyclist,” at the moment it acts.
Drivers who saw those explanations got better at predicting what the car would do next. The team ran it on a real car on real roads, not only in simulation, and it did not slow the driving down.
My Take:
The real prize is in the researchers’ own framing: real-time reasoning lets you test the system while it is running. You find out the model is wrong before it costs you something. Most AI projects I see do not fail on accuracy. They fail because nobody can say why the thing produced what it produced, so nobody trusts it enough to actually use it, and it quietly dies in a pilot.
Source: arXiv
Tool picks of the week
- Pictory.AI: Turns a script or an article into a finished video in minutes.
- Gamma: Builds presentations, documents, and simple web pages fast.
- Workflow & SOP Creator: Maps your operations, builds clear RACI charts and SOPs, and finds the manual bottlenecks worth automating.
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
