> ## Content Index > Fetch the complete content index at: https://www.mindovermoney.ai/llms.txt > Use this file to discover other available public pages before exploring further. # Volume 52: A Year In and Still Not an AI Expert - URL: https://www.mindovermoney.ai/what-should-a-company-ai-policy-include/ - Published: 2026-09-22T12:00:40.000Z - Updated: 2026-09-22T12:00:41.000Z - Description: A rule against putting confidential data into public AI tools is a prohibition, not a policy. A real AI policy answers five questions about disclosure, editorial control, data leaving the building, training, and what happens when the tool changes. - Author: Santosh Savel - Tags: Newsletter Hey everyone! This is the fifty-second straight Tuesday of Neural Gains Weekly, which makes today one year. Thank you for reading, for the honest feedback, and for every conversation these issues started. I have appreciated every one. I have one ask as year two begins. If any of these issues gave you something you used, forward this one to someone in your network who is still waiting to feel qualified. 🧭 **Founder's Corner:** A year of weekly issues did not produce an AI expert, only an outsider whose work became useful to others. 🧠 **AI Education:** Five questions that separate a real AI policy from a list of prohibitions, no legal training required. ✅ **10-Minute Win:** A trial built on your real work, the case against upgrading already written, and a decision date on your calendar. Let's get into it. ## Signals Over Noise ****We scan the noise so you don’t have to — top 5 stories to keep you sharp** #### **1)**[ **Medicare AI Program Results in 'Lengthy Treatment Delays' — New Data Shows**](https://www.newsweek.com/medicare-ai-program-results-in-lengthy-treatment-delays-new-data-shows-12441120?ref=mindovermoney.ai) **Summary:** Roughly 1,000 pages of CMS records, released through an Electronic Frontier Foundation FOIA lawsuit, show Medicare's WISeR pilot producing widespread delays and inappropriate denials since its January launch. CMS tells vendors to respond to prior authorization requests within about 72 hours. One request sat unanswered for 83 days. **Why it matters:** The safeguard everyone points to with AI-assisted approvals is that a licensed clinician reviews every denial. These records show the failure happened on the clock, before any clinician saw anything. If a vendor pitches you an AI approval tool, ask for the turnaround-time distribution, not just the accuracy rate. #### **2)**[ **Healthcare's agentic AI boom is outpacing governance: report**](https://www.healthcaredive.com/news/healthcares-agentic-ai-boom-is-outpacing-security-governance-report/830697/?ref=mindovermoney.ai) **Summary:** In a Vanson Bourne survey commissioned by security vendor Imprivata, more than 85% of healthcare AI leaders said they have visibility into AI agent activity and can fully control and govern an autonomous agent's actions. In the same survey, 72% admitted AI tools get deployed without IT approval at least occasionally. **Why it matters:** Those two numbers cannot both be true, and the gap between them is the actual governance problem. Before your next AI oversight meeting, ask whether anyone can produce a current inventory of what is running, because confidence is not the same thing as visibility. #### **3)**[ **Lawsuit claims Anthropic, OpenAI, SpaceXAI and Google violated antitrust laws when they coordinated AI slowdown**](https://fortune.com/2026/09/19/lawsuit-anthropic-openai-spacexai-google-antitrust-laws-ai-slowdown-subscription-value/?ref=mindovermoney.ai) **Summary:** A suit filed Friday in the Northern District of California alleges the four leading labs illegally coordinated when Dario Amodei's September 12 essay calling to slow frontier development drew same-day public agreement from Sam Altman, Elon Musk, and Demis Hassabis. Four paying subscribers brought it on behalf of a proposed nationwide class. **Why it matters:** Amodei flagged this risk in the original essay and asked the government for a narrow antitrust waiver for safety conversations. The unresolved question is whether competitors can agree with each other to be more careful, which matters to anyone whose company is drafting shared AI commitments with peers or an industry consortium. #### **4)**[ **China's open-weight AI models are now just 4 months behind frontier US offerings, Mozilla report claims**](https://www.tomshardware.com/tech-industry/artificial-intelligence/chinas-open-weight-ai-models-are-now-just-4-months-behind-frontier-us-offerings-mozilla-report-claims-models-still-lag-in-some-benchmarks-but-are-drastically-cheaper-to-use?ref=mindovermoney.ai) **Summary:** Mozilla's State of Open Source AI report puts the capability gap between open-weight and closed frontier models at about 4.4 months, close to Epoch AI's independent four-month estimate. The best open model scored three points below the closed leader at 60% of the price. **Why it matters:** Two caveats change how you should read that number. Mozilla advocates for open models and its own CTO called the report partly advocacy, and the price comparison is list price on hosted endpoints, not what it costs you to actually run the model. The useful question for your team is which specific workloads need the premium model, not whether the premium is worth it in general. #### **5)**[ **AI makes the right to repair more tempting than ever**](https://www.fastcompany.com/91608009/ai-makes-the-right-to-repair-more-tempting-than-ever?ref=mindovermoney.ai) **Summary:** A reporter facing a $500 dishwasher repair quote found the replacement part cost about $100, cancelled the technician, and worked through the repair herself by sending photos to ChatGPT over a few hours. Manufacturers including Apple and John Deere have fought to keep repair information in-house, and state right-to-repair laws now cover roughly a third of Americans. **Why it matters:** The honest part of this story is that the bot kept guessing from manuals for the wrong model and apologizing, and the repair still worked. That is the realistic shape of useful AI right now, where it gives you a starting point that you verify as you go, and it is worth trying on something low-stakes before you trust it on something that is not. ## Make AI useful in your everyday work. Get Neural Gains Weekly in your inbox, with practical AI explanations, prompts, and lessons from building. Subscribe for free Email sent! Check your inbox to complete your signup. One email every Tuesday, alternating NGW issues and podcast takeaways. **Missed a previous newsletter? No worries, you can find them on the* [**Archive page*](https://www.mindovermoney.ai/archive/)**.* ## Founder's Corner ****What a Weekly Deadline Made Possible** In August 2025, I attended AI4 in Las Vegas for the first time, a three-day conference bringing together business leaders and technical practitioners to explore how to apply artificial intelligence across industries. I felt like a complete outsider. I was interested in what was happening, but I did not have the experience to feel comfortable in the conversations around me. Weeks later, on September 30, I published the[ first issue of Neural Gains Weekly](https://www.mindovermoney.ai/how-to-start-learning-ai-as-a-non-technical-professional/), still with a great deal to learn. I had given myself a weekly responsibility to make something useful out of that learning and share it with anyone who wanted to follow along. Fifty-two issues and one year later, that responsibility has changed how I work, what I can build, and what I am willing to try. The weekly deadline meant that learning always had somewhere to go. I had to take an idea far enough to explain it, a tool far enough to understand how someone might use it, or a problem far enough to have something honest to say about the work. Education gave me a foundation, but executing, writing, and building every week made me use it. Trying something gave me more to write about, and writing about it exposed the parts I still did not understand. I learned the most when I held myself accountable to finishing that work, even while my understanding was still developing. Over time, what I was learning began showing up in conversations that had nothing to do with publishing a newsletter. ## **When the Learning Became Useful to Someone Else** One of those conversations happened while I was working with colleagues on a new project to solve a problem in revenue cycle management. Multiple teams were involved in the planning, and we were discussing how best to move the project forward. I brought up Copilot Notebooks, a feature the colleagues in that conversation were not aware of and had not received formal training on. The approach was familiar from how I used Gemini Notebook, Claude Projects, and ChatGPT Projects for the newsletter, gathering documents, notes, and background information so the assistant had context for the work. I had written about that practice in[ Volume 24](https://www.mindovermoney.ai/how-to-use-ai-projects-mode-save-context-professionals/), and now I could see how it might help us organize the information for this project and give the people involved a common starting point. [Copilot Notebooks](https://support.microsoft.com/en-us/microsoft-365-copilot/how-microsoft-365-copilot-notebooks-works?ref=mindovermoney.ai) lets you gather reference material into a workspace and use it to inform your conversations with Copilot. I showed how project documents and notes could supply that background, giving the assistant information to work from as we explored the project. The idea was basic project and context management, with AI available to help work through and build the plan. Colleagues could begin with work they already understood and practice asking questions using those documents and notes. I wanted to give them a useful way to start, developing their own skills while contributing to the project we were planning together. What stayed with me was how simple the approach was. Something I had learned through the weekly work of producing the newsletter gave me a practical suggestion to bring to colleagues working on a different problem. I would not have brought that perspective to the conversation a year earlier, and I gradually began to appreciate how much the work here was changing what I could contribute elsewhere. I still had plenty to learn, but I could already share something useful from what I had tried. That gave me more reason to keep pushing into new features, new workflows, and tools I had not yet become comfortable using. ## **Returning With Something to Contribute** I had been using Claude Code and Codex at home, and that experience became useful when I gained access to coding tools at work. Lessons from working with my engineering partners then changed how I organized my personal projects, a progression I explored in[ Volume 43](https://www.mindovermoney.ai/how-ai-search-tools-actually-work/). Over time, I was building more than individual tools. I was developing knowledge about how to plan the work, organize its context, and judge the results. That raised a question I am still working through. How do I make that knowledge useful beyond the project where I learned it? That question is behind the second brain I am building now. I want a way to preserve the context, decisions, and lessons from my projects so I can use them in future work with AI. The documents and notes I gather at the beginning are only part of that knowledge. What I learn while working through a problem matters too, and I want to make it easier to carry that understanding forward. The goal is to give the tools relevant background without spending each session reconstructing it, making better use of both my time and the resources the tools consume. It is still in progress, and I am learning how to organize it as I go. By the time I returned to AI4 in August 2026, the work I had been doing gave me much more to contribute to conversations with other people building with AI. I could clearly explain what my team was building with AI agents in customer service, get into the architecture, and discuss what had made our projects successful and what we had learned from failures. Most importantly, I could articulate a vision for where that work could go next, beyond what I was seeing in the market at the time. That vision resonated with vendors working in the space, and I could explain it clearly enough for other industry professionals to appreciate it and get excited about the possibilities. The art of the possible had become something I could explore with them, drawing on experience and ideas of my own. I came away energized by their perspectives and by the realization that I could contribute meaningfully to conversations that had intimidated me a year earlier. Fifty-two issues have given me a clearer understanding of where I am, how to apply what I know, and what I want to learn next. Being able to contribute to those conversations did not mean I had everything figured out. The second brain is still a work in progress, and the projects that helped me develop that confidence continue to give me things to learn. What I want to carry into the next year is the commitment that kept me working through those questions, along with a better understanding of what makes that work useful. ## **What I Am Still Learning** Working with AI has made me pay closer attention to the information behind the task. I need to understand what context the tool has, how that information is organized, and what is missing before I can expect useful work from it. I am also learning to be more deliberate about what happens after an answer arrives, how to inspect the result and trace factual claims back to sources I can verify. Data, architecture, and governance have become practical parts of work I am trying to finish. Education helps me understand those concepts, and experimentation gives me a place to apply them, discover gaps, and improve. I did not have that understanding when the newsletter began. The next part I am working through is how to measure whether the effort is paying off. AI tools can consume a great deal of time and money, and I want to get better at defining what I expect back from that investment. That responsibility becomes especially clear when a company is paying for the tools and other people depend on the result. I am learning to look beyond whether something can be built and ask whether it improves the work enough to justify what it takes to build and use it. My understanding of that return is still developing, and it will matter more as the projects become more ambitious. When I say there is no AI expert, I mean there is no point at which I expect to be finished learning. There are people with far more knowledge and experience than I have, and being able to learn from them is part of what makes this exciting. I can contribute what I know while remaining open about what I still need to understand. If you are waiting to feel qualified before trying something yourself, I hope these[ 52 issues](https://www.mindovermoney.ai/archive/) give you a reason to reconsider. I am living proof that you can begin feeling like an outsider and develop skills that become useful to people around you. The work gives you a way to find out what you are capable of. To everyone reading the newsletter, watching or listening to Explore AI Out Loud, or talking through these ideas with me, thank you. Knowing that someone is following along gives this work meaning beyond my own education. My hope is that something you read here gives you a reason to try a new approach in your own work or life. Start with a problem you understand, see what happens, and come back to it with what you learned. You have to start and you have to try, then keep making time to learn from the experience. I will keep doing that work and sharing what I learn through the newsletter and[ Explore AI Out Loud](https://www.youtube.com/@exploreAIoutloud?ref=mindovermoney.ai). ****Explore AI Out Loud Podcast** Find practical ways to use AI in your work and everyday life. Learn from real experiments, including the mistakes, so you can spend less time guessing and more time trying something useful. New episodes every other Tuesday. [Watch on YouTube](https://www.youtube.com/@exploreAIoutloud?ref=mindovermoney.ai) [Listen on Spotify](https://open.spotify.com/show/0348nx4ylmWesSj02ECRjY?ref=mindovermoney.ai) [Listen on Apple Podcasts](https://podcasts.apple.com/us/podcast/explore-ai-out-loud/id6812370441?ref=mindovermoney.ai) ## AI Education for You ****Evaluating Your Organization's AI Policy** #### **The Situation** The partners approved Priya's condition. She had recommended the[ prior authorization appeals](https://www.mindovermoney.ai/prompt-library/ai-prompt-to-write-a-health-insurance-appeal-letter/) tool with a ninety-day run attached, every letter scored on payer approval against the numbers her team posted before the tool arrived. Now the pilot has a start date, and the practice administrator wants written confirmation that the tool is covered by the practice's AI policy before the first letter goes out. Priya has never seen an AI policy at the practice. She goes looking. #### **What They Try First (And Why It Falls Short)** In the employee handbook she finds two paragraphs, added sometime last year, telling staff not to put patient information into public AI tools and to run anything customer-facing past a supervisor. Priya reads it twice, decides the pilot clears both lines, and starts a reply saying so. Then she stops. Appeal letters are not customer-facing in the sense the handbook means, and the patient information they depend on is the thing that makes them work at all. Two sentences of guidance now have to carry a workflow nobody had in mind when they were written. #### **The Concept, Through the Scenario** What Priya is holding is a prohibition. A prohibition names what staff must not do, while a policy names who decides, who answers for it, and what happens when conditions change. Five questions separate the two, and they are the same five whether the document runs three pages or three sentences. 1. What has to be disclosed, and to whom? Part 2 established that[ the labeling duty runs on three conditions together](https://www.mindovermoney.ai/do-you-have-to-disclose-ai-use-at-work/) and that most workplace writing fails the first. Sitting underneath it is a question about the tools themselves. The Commission's guidelines treat grammar correction, spellchecking and AI translation as standard editing that carries no marking obligation, while a summary or a rewrite that changes meaning does carry one. That line runs straight through the practice's most common use of the tool, and no document in the building draws it. 2. Who holds editorial control, and how far does the tool run before a person takes it? Part 2 covered what the Commission means by editorial control. What a policy has to add is where the handoff sits. The guidelines describe an AI agent as something that should announce itself to the person instructing it at the points of authorization, reporting and validation. Those three moments are exactly where a practice has to name a human. A physician signs the appeal letter, which settles the last one. The first two belong to nobody yet.[ Vol 42's autonomy dial](https://www.mindovermoney.ai/how-much-autonomy-to-give-an-ai-agent/) set that distance for one tool at a time, and a policy sets it once, for every tool that arrives after. 3. What data leaves the building? The handbook covers this one, in the only way it covers anything. It draws a line around public tools and says nothing about the vendor tool the practice is about to pay for, where chart notes and denial letters travel somewhere by design and a contract decides what happens to them once they land. 4. What training exists, and who receives it? Part 1 established that[ the literacy duty attaches to the staff who actually use the tools](https://www.mindovermoney.ai/does-the-eu-ai-act-apply-to-us-companies/). The transparency guidelines close the loop and state that the literacy requirement applies to providers and deployers of the systems those transparency rules cover. Two duties this series treated separately turn out to be one obligation with two halves. A policy that assumes training happened is a different document from one that says what the training covered and who sat through it. 5. What happens when the tool changes? A vendor can swap the model inside a product, and the behavior changes with it. Part 2 described a marking grace period that runs out in December, which means output carrying no mark this month may carry one before the year ends, with no announcement from anyone. A policy written against how a tool behaves today has a shelf life unless it names who watches for the change. #### **What Changes** Back at her desk, Priya runs the five against the two paragraphs. Question three comes back partly answered. The other four come back blank. Writing the policy is not her job. She is the operations manager rather than counsel, and a practice policy on AI belongs to the partners and whoever advises them. So she writes the gap list, one line per question, with the pilot as the worked example on every line. That takes her under an hour, and it goes to the administrator alongside the confirmation the administrator asked for, which now reads very differently. Reading it back, she notices the shape of the thing. Four questions got her the truth about a product. Five more get her the truth about the rules standing around every product that follows it. One of the five she takes herself. The people using this tool are her staff, which makes the training question hers to answer. She books thirty minutes before the pilot opens, covering what the tool gets wrong and what to check before a letter reaches a physician for signature. Then she adds a line to the ninety-day condition. If the vendor swaps the model during the run, the run restarts. #### **What This Reveals** Most readers who go looking this week will find what Priya found, a short prohibition written before these tools arrived and never revisited since. That is the ordinary condition of the document rather than a failure of whoever wrote it, because the workflows moved faster than the handbook. An afternoon spent on the five questions produces something nobody else in the building is holding, a written account of what the organization has not decided yet. Stripped of Priya's practice, the page fits on one screen. 1. What has to be disclosed, to whom, and by which of us? 2. Who has the authority to change or refuse what the tool produces, and at which step? 3. What data leaves the building, where does it land, and what does the contract say happens to it there? 4. What training exists, and who has actually had it? 5. Who watches for the tool changing, and what happens to our approvals when it does? Running them takes no permission and no legal training. Priya had neither, and the gap list she came back with is now the most useful thing anyone at that practice has written about AI. #### **How This Connects** This series opened with the duty that already reaches your desk, moved to what disclosure actually requires and who owes it, and closes on the document your employer either has or does not.[ Vol 49's four questions](https://www.mindovermoney.ai/how-to-evaluate-ai-tool-safety-at-work/) evaluated one tool before Priya recommended it, and the five above evaluate the rules standing around every tool she will be handed next. Whether a tool is safe to use and what you are expected to say about using it were always different questions, and both of them now have a page. Vol 42's dial returns inside question two, set once for an organization instead of once for a product. The two questions this page can only point at, what happens to your data inside an enterprise tool and how to judge the vendor holding it, open the Enterprise series later this year. *Part 3 of 3 in the AI Governance and Disclosure series.* ## Your 10-Minute Win ****A step-by-step workflow you can use immediately** ### **Your Free Trial Is Not a Test** You hit the paywall on a tool you actually like. That is a good place to be standing. It means you used the thing long enough to reach its edges, and edges tell you more than any review will. What usually happens next is a guess. Notion now keeps full Notion AI on its paid Business and Enterprise plans and leaves free users a small allowance of complimentary responses. That is the kind of change that turns a quiet Tuesday into a pricing decision. Back in[ Vol 42](https://www.mindovermoney.ai/how-much-autonomy-to-give-an-ai-agent/) we audited the tools you already pay for. This is the other half of that question, the one that arrives before the money does. So build a test instead. You hand the paywall to an LLM, let it design a trial against your real work, let it argue the case against upgrading, and let it draft the memo you read when the test ends. #### **The Workflow** **1\. Name the gate (2 minutes)** Open a blank note. Write the one thing you cannot do on the free tier. Then write the last three times you hit it and what you did instead. Those three moments are your evidence base. **2\. Let AI design the test (3 minutes)** Open Claude, ChatGPT, or Gemini. Any free tier works. Fill in the reset field honestly, because it decides how long your test can run. **Copy/Paste Prompt:** *I am on the free tier of \[TOOL NAME\] and I keep hitting this limit: \[DESCRIBE THE GATE IN 1-2 SENTENCES\]. When I hit it, the limit \[RESETS DAILY / RESETS MONTHLY / IS A ONE-TIME ALLOWANCE THAT NEVER REFILLS\]. Here are the last three times I hit it and what I did instead: \[PASTE THE THREE MOMENTS FROM YOUR NOTE\]. Design the longest test that fits inside the headroom I actually have, up to 7 days, and tell me why you picked that length. A 3-day test that finishes beats a 7-day test that runs out of allowance on day two. Build the daily tasks out of the pattern in those three moments, not out of a demo or a feature list. For each day, name what I should watch for, what result counts as evidence for upgrading, and what result counts as evidence against. End with the single question the test is meant to answer and the date I should decide.* **3\. Make it argue against you (2 minutes)** The same chat works best here, because the model already has your context. **Copy/Paste Prompt:** *Now argue the case against upgrading. Name three ways I could get the same outcome without paying, including a free alternative tool, a change to my own workflow, and simply doing without. For each one, be specific about what I would give up. Then name the one condition that would have to be true for the paid tier to be right. Do not soften this. \[IF YOU ARE IN A NEW CHAT, PASTE THE TEST PLAN AND YOUR THREE MOMENTS HERE FIRST.\]* **4\. Write the memo before you need it (2 minutes)** **Copy/Paste Prompt:** *Today is \[TODAY'S DATE\] and my decision date is \[THE DECISION DATE FROM MY TEST PLAN\]. Write a one-page decision memo template for me to fill in on that date. Include the question, a line for each of the \[NUMBER\] days in my test, the strongest argument against upgrading, and a final choice of upgrade or do not upgrade with a one-sentence reason. Keep it under 200 words so I will actually fill it in.* **5\. Schedule the test and the decision (1 minute)** Put the daily task list where you will see it each morning, a calendar series or the top of your task app. Then drop the memo template into an invite on your decision date, titled with the tool name. Ten minutes of setup turns ordinary work into evidence. When that invite fires, you fill in the memo and you decide. The model does not decide. You do. #### **The Payoff** You walk away with three things you did not have this morning. A test built around your work rather than a feature list, the argument against the purchase written while you were still calm, and a dated appointment with your own decision. The portable move is pre-commitment, and it works on any subscription or any tool someone asks you to buy. #### **The AI Concept You Just Used** Set your decision criteria before the decision arrives, while you are still thinking clearly and nothing is expiring. That is pre-commitment. AI makes it practical, because designing a real test used to take longer than the decision was worth. Step 3 asked the model to build the case against what you were leaning toward, the fastest way to find out whether you were leaning on evidence or a countdown timer. #### **Transparency & Notes** - On the free tiers of Claude, ChatGPT, or Gemini, every step here costs nothing. - According to Notion's own help center, Notion AI is only available on Business and Enterprise Plans, and Free and Plus users receive a limited allowance of complimentary AI responses to try it. Vendor packaging moves, so check the current terms before you run your own test. - Describe the shape of your work rather than its contents in the three moments. No client names, no internal metrics, no confidential documents. - This test measures whether a tool earns its price at your current volume, not whether that volume will change. Revisit the call if your work does.