AI in Medicare, Meta’s Mac Assistant, and New FDA Thinking on AI
Dear Friend,
In our AI Smart Lab session last week, we cleaned up our own ClickUp agents. We had a pile of separate helpers pulling action items out of client emails and meetings, so we merged them into one agent that reads the work, sorts it, and sets each task with an owner, a due date, and a priority. It runs the triage we used to do by hand. If you want to join our AI Smart Lab on Wed to see our Agent Org Chart in Click up, email [email protected] to get added to the zoom or get the recording, for Wed. Aug 26th, 2pm PST.
One of our agents was proofreading legal-style copy, and we set a hard rule: a person checks it before it goes anywhere. AI still invents things that read as true, and legal wording is the last place you want a confident guess. Fast is good. A human on the final read is what keeps fast from turning into an expensive mistake.
Run Your First AI Pilot Without Wasting Money
If you have tried AI and watched it fizzle, the setup was probably the problem, not the tool. Our new guide walks you through how to run a first AI pilot the way we do it with clients. You pick one workflow that already hurts, set your data rules before anyone starts, keep human review steps, and measure the result before you scale. It is the difference between a test you can learn from and six months of scattered experiments. Read the full text.
Our Most Requested Service: AI Strategy and Consulting
Right now clients keep coming to us in the same spot. They know AI can help, but they are staring at a pile of tools, unsure which ones to trust, how to protect their data, and how to get their team to actually use any of it.
Here is how we work. We start with your goals, not a tool list. We map the workflows that eat up your time, pick one to prove out first, set the data guardrails, and put a human review step on anything that leaves the building. Then we build a simple, phased roadmap your team can actually follow. No 40-tool stack, and no science projects. Just the next right step, and the one after that.
Want to see the exact framework we start with? Book a meeting today.
This week’s stories
AI Is Now Deciding Who Gets Medicare Care in 6 States
In six states, a Medicare pilot called WISeR lets AI from private tech firms approve or deny certain treatments before a doctor can go ahead. Patients report waits of two to four weeks, and repeat denials, on care that used to take a day. The firms are paid partly based on the savings they generate from denials, which quietly builds in a reason to say no.
A private company’s software can now sit between you and a treatment your own doctor already approved, and the people affected cannot see how the decision was made.
My Take: The problem here is not that AI is in healthcare. It is that a black box with a money motive to deny is making the call, and when a patient asks why, nobody can answer. The doctors literally call it a black box, and that word should worry any business, not just hospitals. When we build an AI decision step for a client, we require two things: a human review on the output, and a system that can show the reason for its answer. An automated ‘no’ you cannot explain is worse than a slow yes. I would not deploy an AI decision you cannot audit and reverse, especially anywhere a wrong answer lands on a real person.
Source: BBC Science Focus
Meta Put an AI Assistant on Your Mac That Reads Your Ads and Inbox
Meta launched a Mac app that connects to your Facebook and Instagram business accounts, your ad campaigns, and Google Workspace tools like Gmail, Docs, and Sheets. You can share a window on your screen and ask what is working, and it can draft posts, decks, and spreadsheets. It is free to start, with paid tiers for heavier use. The catch: by default, the business data you share can be used to train Meta’s models and target ads, unless you switch on Incognito.
My Take: This is convenient, and that is exactly the trap. The pitch is one assistant that sees your ads, your inbox, and your calendar in one place. The price is Meta seeing all of it too, by default, to train on and to sell against. Convenience and data exposure are the same feature here, not a trade you get to skip. When we set up an AI tool for a client, we read the data terms first and decide what the tool is allowed to touch, because free almost always means you are paying with your data. I would never recommend this to a client to use to keep anything private, financial, or client-owned out of it until the terms earn my trust. I don’t see that happening anytime soon, especially when Google and Microsoft are already there for the majority of companies.
Source: Meta
Cancer Research UK Is Handing 20 Years of Data to an AI Cancer Model
Bioptimus, a Paris AI company, partnered with Cancer Research Horizons, the commercial arm of Cancer Research UK, to train its cancer foundation model on a dataset of more than 1,750 colorectal cancer patients. The data took two decades to collect and includes tissue images, genetics, and real treatment outcomes. The goal is a model that can predict which patients a given treatment will actually help, not just who tends to do badly.
My Take: The real story is the data. Bioptimus did not pull ahead because it has a cleverer model than everyone else. It pulled ahead because it got access to twenty years of outcome-linked data almost nobody else has. In AI right now, the model is becoming the commodity and the proprietary data is the moat. We tell clients the same thing when they panic about choosing the perfect tool: your edge is not the model everyone can rent, it is the data and the workflows only you have.
Source: Cancer Research Horizons and Bioptimus
The FDA Is Figuring Out How to Regulate AI That Changes Its Own Answers
The FDA put out a discussion paper asking how it should evaluate generative AI medical devices, since its usual software rules do not fit tools that give different answers to similar questions and keep changing after launch. The paper flags three risks by name: hallucinations that look credible, performance that drifts after deployment, and the problem of building on a third-party model you cannot see inside. Public comments are open until October 19.
In plain terms, the FDA is admitting that the way it checks normal software does not work on AI that keeps rewriting itself, so it is asking the public how to do it before writing rules.
My Take: Read this even if you never touch a medical device. The FDA just wrote down, in plain regulatory language, the three risks most businesses ignore when they ship AI. It makes things up convincingly, it can get worse after launch, and you often cannot see inside the model you built on. Those are not device problems. They become your problems the day you put AI in front of a customer. We consistently re-test client AI after go-live, not only at launch, because a model that passed in the spring can quietly drift by the fall. Build a scheduled maintence into anything you deploy, the way the FDA is proposing to, because “it worked the day we launched it” is not an excuse for something not working.
In tech, the idea of maintenance is not a new concept, anybody who has built an app knows that with every OS update you have to update your app too. AI is no different, your instructions need to be updated on occasion. There are ways to ensure greater stability too, you can self-host static models. If you want to learn more about this email [email protected]. The FDA should be doing this and we’ll let them know.
Source: U.S. Food and Drug Administration
Tool picks of the week
Workflow and SOP Creator: Map your business operations and turn them into clear SOPs and RACI charts, with automation built in. It fits a week where every story came back to the same lesson: document the process and keep a human in it before you hand it to a machine.
Snitcher: Identifies the companies visiting your website and turns that traffic into leads and outreach. We run it on our own site, get real value out of data you already have.
Fyxer: Organizes your inbox, drafts your replies, and writes your meeting notes. A simple, practical assistant for the busywork.
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
