Volume 39: AI Got the Keys to Your Refills but Not Your Denials

In one state, an AI agent renews people's prescriptions on its own, no human signature required. In that same state, the same software cannot deny anyone's care without a licensed person behind the decision. One technology, two opposite rules, because the law keeps governing a word when it should govern the work itself.

🧭 Founder's Corner: Why a wave of new healthcare AI laws keeps regulating the word instead of the work, and the line between routine paperwork and a real clinical decision.

🧠 AI Education: An honest first week inside a second AI tool: what the free version delivers, where it pushes back, and how to tell which tasks belong in which tool.

✅ 10-Minute Win: Build your own map of which AI tool to reach for at each stage of a task, so you stop forcing one model to do five jobs.

Let's get into it.

1) Introducing the OpenAI Partner Network

Summary: OpenAI launched the OpenAI Partner Network on June 14, 2026, its first formal global partner program backed by $150 million, targeting 300,000 certified consultants by year-end, with launch partners including Accenture, BCG, McKinsey, Bain, PwC, Eliza, and Artium. The company stated AI model capabilities are no longer the main barrier to enterprise adoption.

Why it matters: The biggest AI lab on earth just publicly conceded that the model isn't the bottleneck, the implementation is. For any leader being told "we just need the right tool," this resets the conversation, the next phase of AI value will be won by the organizations that get governance, change management, and workflow redesign right, not by the ones chasing the latest benchmark.

2) AI can help close workforce gaps while keeping humans in the loop

Summary: CommonSpirit Health's CMIO says AI's greatest value may be helping clinicians scale cancer screening, identifying overlooked findings and reducing administrative burden while preserving human oversight. CommonSpirit is one of the largest nonprofit and Catholic health systems in the United States, operating more than 2,200 care sites and hospitals across 24 states.

Why it matters: This is what scaled, governed clinical AI looks like at a system serving 24 states. The pattern, AI finds the signal, clinicians make the call, is the model healthcare leaders should be benchmarking their own rollouts against right now.

3) House Appropriations Committee votes to end funding for WISeR pilot

Summary: The House Appropriations Committee voted to cut funding for the Wasteful and Inappropriate Services Reduction (WISeR) Model, which would introduce AI-driven prior authorization into traditional Medicare. The model would hire private companies to use artificial intelligence to automate prior authorization, with vendors compensated based on a share of "averted expenditures."

Why it matters: The first federal AI program to put approval-or-deny decisions on Medicare patients is now politically radioactive. For any organization building AI into payer workflows, the message is that incentive design, especially anything that rewards denials, is going to be the first thing regulators and Congress scrutinize.

4) AI in spotlight at G7 as Trump, world leaders joined by tech chiefs

Summary: Sam Altman, Dario Amodei, Demis Hassabis, and around a dozen other tech leaders took part in a working lunch at the G7 summit in Evian-les-Bains on June 17, 2026, with frontier AI risks, infrastructure and sovereignty all expected on the agenda.

Why it matters: This is the first G7 where AI lab CEOs sit at the same table as heads of state, and after the Fable 5 export-control episode, it is no longer an academic conversation. Expect coordinated frameworks on frontier model deployment, child safety, and biosecurity to start moving from communique language into actual rulemaking.

5) Improving health intelligence in ChatGPT

Summary: OpenAI announced that more than 230 million people use ChatGPT every week for health-related topics and that GPT-5.5 Instant is now better at identifying situations that may require urgent medical attention, asking follow-up questions when more information is needed, and explaining complex medical topics in simpler language. GPT-5.5 Instant had fewer instances of not tailoring to local healthcare context, missing red flags or referral to care, or failing to seek additional context from the user when needed than both older models and physicians.

Why it matters: A general-purpose chatbot is now the first health touchpoint for hundreds of millions of people every week. For health system leaders, the question is no longer how to keep patients off ChatGPT, but how to design intake, education, and follow-up so the work it does well strengthens your care plan instead of competing with it.

In Utah, an autonomous AI agent renews people's prescriptions on its own. Regulators there cleared it to handle 192 different chronic-condition medications with no human signature required, quietly keeping patients on the drugs they depend on. That same year, the same legislature went the opposite way on prior authorization. When AI helps review one of those requests, any denial now has to come from a licensed person, not the software. One state. One technology. Two opposite answers.

Prior authorization is the approval an insurer requires before it will cover certain treatments. For many medications it is a routine step in the prescription lifecycle. The criteria behind it come from clinical guidelines that flag when an expensive or specialized drug deserves a closer look. It is also where a large share of the system's friction collects. So why would one legislature hand AI the keys to refill a prescription on its own, yet insist a human stay in charge the moment that same software might deny a patient's care?

Utah's split runs along an access-versus-risk line. Renewing a stable patient's medication is mostly a question of access, so the state let the agent move fast. Denying care is a question of risk, so it kept a human accountable. That instinct is sound, but it stops one level too high. Even inside a single process like prior authorization, AI is still not one decision, and until the laws taking shape across the country see that, they will keep regulating a word instead of the work underneath it.

Every State Is Writing Its Own Rules

In the last two years, AI in prior authorization has jumped from conference panels to legislation in statehouses across the country. And in traditional political fashion, no two states have landed in the same place. California's Physicians Make Decisions Act keeps AI from being the sole basis for denying or delaying care. A licensed professional still has to make the medical-necessity call, the judgment that a treatment is genuinely needed. Georgia took a different path, passing a law that lets insurers use AI in prior authorization to automate routine work and ease the paperwork load. Utah added a disclosure rule of its own, requiring insurers to say when AI is part of the review at all. Dozens more states sit somewhere in between, each writing its own version of how AI interacts with this part of the healthcare system.

Those laws look like a country that cannot make up its mind. But almost all of them are arguing about the same wrong question, how far AI should reach across prior authorization as one undivided process. That framing is the real problem. Prior authorization is not a single act to wave through or rein in. It is several jobs bundled together, most of them routine and only a few of them clinical. A law that cannot tell those apart ends up treating all of it the same way.

The Yes Was Always Coming

Underneath the disagreement, the states do share one rule. Whatever a law says about AI, it stops short of letting software deny care on its own. A person has to stand behind any refusal. That common ground is real, but it settles only a small piece of what prior authorization does. Most requests never reach a real clinical judgment. The bigger story is the part the debate keeps stepping over.

A December 2025 white paper from the National Association of Insurance Commissioners puts approval rates for prescription-drug prior authorizations at around 90 percent. Nine in ten requests end in a yes. After more than a decade on the technology and operations side of specialty pharmacy, where the most expensive and complex drugs almost always require this step, that number is no surprise. And it changes what prior authorization really is and what it means. If nine in ten requests are approved anyway, most were never a clinical gate. They are administrative throughput, paperwork moving between systems while everyone waits for a stamp that is already coming.

Running that nearly-automatic process still costs a fortune in time. A 2024 American Medical Association survey, the research the NAIC drew on, found that doctors and their staff spend about 13 hours a week on prior authorizations. Thirteen hours a week, feeding a process that says yes nine times out of ten. Most of that time is pulled straight from patients, who wait while the paperwork clears. So why not scrap prior authorization altogether? Because the criteria underneath it do real work. Clinicians write them to catch the cases where an expensive or risky treatment deserves a second look, and that judgment is worth protecting. The waste is everything wrapped around it, the faxing, the re-keying, the chase for one missing field.

Automate that, and the yes that was always coming arrives in minutes instead of days. The thirteen hours go back to patients. The human stays where the judgment is real, on the small share of cases that genuinely need one. That is the line worth drawing, and it is exactly the line these laws get wrong.

When a Blank Field Counts as a Denial

Under the hood, much of what gets counted as a denial is not one. A prior authorization can come back rejected because someone left the patient's weight off the form. Nothing clinical happened. No reviewer weighed the treatment against the criteria and said no. A required field was blank, so the system kicked it back. The NAIC names this directly, noting that incorrect or missing patient information can delay a request or produce an unexplained denial. The record still calls it a denial, but it is really a clerical gap, the kind of gap software should catch and fix before a person ever opens the file.

Now compare that to a real clinical denial, a clinician reading a complicated case and deciding the treatment does not meet medical necessity. That is a judgment call, and it carries weight a missing data field never could. Yet most of the laws on the table treat the two as the same act, because both run on software and get filed under the same word. Govern them together and there is no good setting. Loosen the rules to clear paperwork and you weaken oversight on the denials that need it. Tighten them to protect patients and the clerical work that should take seconds drags for days. Every one of these laws makes the same mistake. It writes a single rule for a single word, AI, when that word covers both a clerical task and a clinical decision.

A pharmacy that fills prescriptions across state lines now follows a different AI rule in each one. Every rule has its own disclosure language and its own definition of what even counts as AI. The burden skips the largest, best-resourced players and lands on the providers and pharmacies already stretched thin. Fifty versions of a question no one has framed correctly are overhead dressed up as oversight.

Healthcare Knows How to Build Inside the Lines

Legislation will keep reshaping how AI works in healthcare, state by state and year by year, and prior authorization is only what happens to be in front of us right now. None of that is a reason for pessimism. Healthcare is adept at working in a regulated environment. This is the industry that learned to innovate inside HIPAA and FDA review, under decades of rules far heavier than a disclosure form, and the innovation will continue.

What makes me optimistic is not a hope that the rules get lighter. It is what AI is genuinely good at. The friction in this whole story, the thirteen hours, the bounced forms, the waiting that lands on a patient, is exactly the kind of waste the technology clears best. Aim AI at the administrative half of prior auth and you strengthen patient protection, because the time handed back flows to the people the process exists to serve. The optimism here is pro-patient.

And if you want more access for people, change the system. You do not slow down the one tool that creates speed just to keep the pace where it has always been.

Whether AI belongs in healthcare was always the wrong frame. The real question lives one level down, inside the process, where administrative throughput and clinical judgment turn out to be two different jobs that happen to share a name. Good governance automates the throughput and keeps the human on the judgment. Get that line right, and AI does what it should, reduce friction across the healthcare journey and improve the experiences we all go through.

The Setup

A professional in healthcare operations has used ChatGPT every workday for two years. It drafts her emails, helps her think on her commute, and turns her messy meeting notes into something forwardable. She is fluent in it and not looking to switch.

A colleague suggests that the long strategy documents she wrestles with each quarter might go better in Claude, Anthropic's AI assistant. She is skeptical, but she has a quieter week before her next planning cycle, so she runs an experiment. One week, Claude only, the free version, her actual work. The point is not to crown a winner. It is to learn what a second tool can and cannot do, so she knows when to reach for which.

This is what the week taught her, and most of it is true of the free plan anyone can open today.

Monday: Getting Oriented

The first thing she has to do is find her footing, and the interface makes that easy. It is a chat box, like the one she already knows. The notable difference is small and worth understanding: a control that lets her pick how hard the model thinks about a given request.

This is the first real concept of the week. Claude offers more than one model, and the free plan runs on the fast, capable everyday one. For most of what she does, drafting, summarizing, talking through a problem, that model is the whole job. The heavier model that the paid plans unlock is built for the hardest reasoning and coding, and she does not need it on day one. Knowing the distinction means she is not left wondering whether she is using the tool wrong. She is using the right model for the work in front of her.

Tuesday: The One Feature That Changes the Most

Her real test is a forty-page strategy document. She needs every place it commits to a timeline.

In her ChatGPT habit, a document this long meant working in pieces. Paste a section, ask, paste the next, stitch the answers together herself. She drops the whole document into Claude in one go and asks once. The answer references a commitment near the front and a caveat near the end in the same response.

The concept underneath this is the one most worth learning from the whole week: the context window, which we covered back in Vol 8. It is the amount of text a model can hold in active view at one time. Claude's is large, large enough that a forty-page document fits inside it whole. That is not a magic trick and it is not unique to Claude, but it is generous on the free plan, and it is the single feature that most changes what kind of work feels natural here. If her job involved holding a lot of material in view at once, this is where she would feel the difference.

Wednesday: Where the Free Plan Pushes Back

By midafternoon she hits a wall. A notice tells her she has reached her usage limit and needs to wait for the window to reset.

This is the honest cost of the free plan, and it is worth stating plainly rather than burying. Free Claude is built for occasional use, not all-day reliance. The limit is not a simple count of messages either. It scales with how much text the model is working through, which means the forty-page document she loved on Tuesday spent her allowance faster than a dozen short questions would have. The capability and the constraint are the same coin.

Here is where the comparison turns genuinely two-sided. Her ChatGPT habit of firing off quick questions all morning never ran into this kind of ceiling as fast. For rapid, lightweight, all-day use, that experience was smoother. Neither tool is wrong. They are tuned for different rhythms, and the free plans make those tradeoffs more visible than the paid ones do.

Thursday: An Honest Accounting of What Is Missing

By Thursday she is keeping a running list of what Claude does not do that ChatGPT did, because an honest evaluation requires it.

Claude does not generate images. When she wanted a quick illustrative graphic, she went back to ChatGPT, which does. The ecosystem of third-party add-ons and custom tools around ChatGPT is larger and more mature. Some of the consumer-facing conveniences she had grown used to were simply not there. None of this is hidden, and none of it is a flaw exactly. It is a different product with a different center of gravity. Claude is built first around careful reasoning and text. ChatGPT spreads wider across features. A fair evaluation names both halves.

What the free plan does give her, and what she confirms by week's end, is the genuine core. The same writing and reasoning quality the paid plan offers, file uploads, web search, and the workspace features the rest of this series will dig into. The free tier is not a crippled demo. It is the real product with a ceiling on how much she can use it.

Friday: What She Actually Decided

She did not delete her ChatGPT account. That is not how a working professional adopts a tool, and pretending otherwise would be the tell of a sales pitch rather than an honest week.

What she decided was narrower and more useful. The two tools have different shapes. ChatGPT suited her fast, scattered, all-day questions and the moments she needed an image or a niche custom tool. Claude suited the heavy, document-driven thinking that defined her hardest weeks. The free plan was enough to learn that. Whether she ever pays the twenty dollars a month, the same price as the ChatGPT plan she already had, is a question she can now answer from experience rather than from a review.

She ended the week understanding what free Claude is. Not better, not worse. A different tool with a clear strength, an honest limit, and a free tier real enough to judge it on.

What to Take From This

• The free version of Claude is the real product, not a teaser. Full writing and reasoning quality, file uploads, web search, and a large context window are all available without paying. The ceiling is on how much you use it, not on what you get.
• The context window is the concept that most shapes the experience. It is how much text the model holds in view at once, and Claude's is large enough on the free plan to work through long documents whole. Vol 8 covered the idea; this is what it feels like in daily use.
• The free plan's usage limits scale with how much text you run through it, not just how many times you hit enter. Long documents are powerful and expensive at the same time. Plan your heavy lifting accordingly.
• Claude does not do everything. No image generation, a smaller third-party ecosystem, tighter free limits than some alternatives. An honest tool evaluation names what is missing, not just what shines.
• The useful question is never which AI is best. It is which of your tasks fits which tool. A week of your real work answers that better than any benchmark.

How This Connects

This opens a four-part Claude Deep Dive, a guided tour of what the platform offers and where it falls short, anchored in the free plan anyone can use. This volume was orientation: what Claude is, what the free tier includes, and the honest pros and cons of first contact. Vol 40 goes under the hood on Projects and memory, the features that let Claude build context across conversations. Vol 41 covers Artifacts and long-form work. Vol 42 closes with Claude Code and the agentic future, and makes the case that the agentic shift reaches well beyond developers. The fine-tuning series that just ended explained how these models are built. This series is about what it is like to actually work inside one.

Part 1 of 4 in the Claude Deep Dive series.

Stop Forcing One AI to Do Everything

You picked a tool a year ago and use it for everything. Research, drafting, cleanup, polish, all in the same chat window. Lately you can feel the edges. The research is shallow, or the writing is generic, or it cannot hold the whole document in its head. You suspect another tool would do that one part better, but switching mid-task feels like overhead. So you keep forcing one model to do five jobs.

Matching the tool to the job is the single biggest lever in AI fluency, more than any individual prompt. The people who get the most out of AI are not the ones with the best single tool. They are the ones who know which tool to reach for at each phase.

Why this matters: most work has phases, and the phases want different things. Research wants breadth and current sources. Drafting wants reasoning and long context. Polish wants speed. No single tool is best at all of them. This workflow builds your own map of which tool fits which phase.

The Workflow

1. Pick a real multi-phase task (1 minute). Choose something with several stages: a report, a proposal, a research memo. Not a one-shot question. Something you are actually working on.

2. Break the task into phases with an AI (3 minutes). Open Claude or ChatGPT and have it map the work before you do it. Paste this:

Copy/Paste Prompt: "I am working on [YOUR TASK]. Break it into its natural phases, for example research, drafting, refinement, and polish. For each phase, tell me in one line what that phase actually needs from an AI tool, such as current web sources, deep reasoning, long context, speed, or formatting. Do not recommend specific products yet."

Read the list and mark the phase where your usual tool feels weakest. That is the one to test next.

3. Run a head-to-head on that phase (3 minutes). Write one prompt for the weak phase you just marked. Run that same prompt in two tools, your usual one and one other (Claude, ChatGPT, Gemini, or Perplexity). Read both outputs side by side. You will feel the difference faster than anyone could explain it.

4. Write your switch rule (2 minutes). Decide where in this task you will switch tools and why. Keep it to one line, like "research in Perplexity for the live sources, then draft in Claude for the long context." The rule matters more than getting it perfect.

5. Save your phase-to-tool map (1 minute). Write a three-column note: phase, tool, why. Four or five rows. Next task, you start from the map instead of forcing one tool through all of it.

The Payoff

You walk away with a personal phase-to-tool map and a switch rule you can reuse on every multi-stage task. The bigger gain is the instinct underneath it: you stop asking "what can my tool do" and start asking "what does this phase need." That question, asked over and over, is how AI taste actually develops.

The AI Concept You Just Used

Tool-to-task matching. Every model has a shape: strengths, blind spots, a context limit, a personality. Forcing one model through every phase is like using a chef's knife to also open cans and tighten screws. It works, badly. The head-to-head test is the fastest way to learn each tool's shape, because you judged them on your own work instead of trusting a review. Run it a few more times and switching stops feeling like overhead and starts feeling like skill.

Transparency & Notes

• This workflow runs entirely on free tiers. Claude, ChatGPT, Gemini, and Perplexity all let you run the same prompt at no cost. Free tiers use lighter models with message caps, so you are comparing the free versions, which is what matters if you do not pay.
• Tool strengths shift with every model release, so treat any phase-to-tool map as a living note, not a permanent ruling. Re-run a head-to-head when a tool ships a major update.
• The head-to-head is more honest than any published benchmark, because it uses your real task and your judgment. Trust what you see over what you read.
• Switching tools means re-pasting context. For long documents, keep a master copy of your content somewhere outside the chat so you can move it between tools cleanly.