Volume 38: Why Your Best AI Idea Is Stuck in Someone's Inbox

Almost every company is using AI somewhere now. Far fewer have it running at full scale, and the reason is rarely the technology. The thing slowing your best AI work down is usually the structure it has to travel through before anyone can say yes.

🧭 Founder's Corner: The hidden reason your AI work keeps stalling is structural, not technical, and you can move faster without waiting for the org chart to change.

🧠 AI Education: A clear framework for knowing when to simply prompt a model, when to give it access to your documents, and when training the model is actually worth the cost.

✅ 10-Minute Win: Turn a pile of rough notes into a finished, designed slide deck in ten minutes instead of ninety.

Let's jump in.

1) WWDC 2026: Everything announced on Siri AI, iOS 27, Apple Intelligence, and more

Summary: Apple introduced Siri AI at WWDC 2026, with Google Gemini under the hood, making the new Siri more capable, conversational, and compatible with visual intelligence, housed in a standalone app in addition to working across existing apps.

Why it matters: The world's most-used voice assistant just outsourced its brain to a competitor. For non-technical professionals, this is the moment Gemini quietly arrives on more than a billion devices, whether you signed up for it or not.

2) Philips Future Health Index 2026: AI is already saving clinicians time and delivering measurable impact in healthcare

Summary: Philips' global Future Health Index 2026, based on 2,000+ clinicians across 10 countries, found 46% of clinicians reported time savings of at least 132 hours annually on average and 39% have already seen AI identify or prevent potential medical errors at least three times in the past three months.

Why it matters: This is one of the largest real-world data points on clinical AI to date, and the numbers move the conversation from "could it help" to "by how much." For healthcare leaders, the new question is whether your training, governance, and infrastructure can capture that value, or whether it leaks out of your organization.

3) Anthropic disables access to Fable 5 and Mythos 5 to comply with government directive

Summary: Three days after launch, Anthropic disabled its two most powerful models after receiving a US export control directive citing national security authorities, requiring it to suspend all access to the models "by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees." The company shut both models down for every customer to ensure compliance, while all of its other models will not be affected.

Why it matters: This is the first time the US government has reached in and shut off a deployed commercial AI model worldwide. The takeaway for any business building on a frontier model is that the most powerful tools now come with an off switch the vendor does not fully control, which makes model diversification and clean fallback paths a real business requirement, not a hypothetical.

4) EU Says Decision Not to Launch Siri AI in Europe Is Apple's Alone

Summary: One day after Apple blamed the Digital Markets Act for keeping Siri AI out of EU iPhones, the European Commission publicly rejected that framing, saying Apple simply requested a blanket exemption from its interoperability obligations under the Digital Markets Act, something the Commission says is not an available option.

Why it matters: This is the first time a major AI launch has been openly held up by an interoperability rule, and it will not be the last. For any company shipping AI features into regulated markets, the playbook is shifting from "ship and adjust" to "design for interoperability and access on day one."

5) HFMA 2026: AI front and center

Summary: At the Healthcare Financial Management Association annual meeting, hospital CFOs broadly expressed enthusiasm for AI's potential. Dennis Dahlen, the chief financial officer at the Mayo Clinic, cited AI as one of the reasons he thought the future of healthcare delivery is as bright as it's ever been. HCA's CFO Mike Marks expressed optimism about AI improving the prior authorization process.

Why it matters: Healthcare CFOs control the budget that decides whether AI scales inside a hospital or stays in pilot. When the people writing the checks are openly bullish, expect the next round of AI investment to shift from clinical innovation to revenue cycle, scheduling, and prior auth, the workflows where ROI is easiest to defend.

Your best idea this quarter is probably sitting in someone's inbox, waiting for a yes. Or sitting on a "roadmap" waiting for funding and focus through an outdated delivery process. This is typically chalked up to the way things have always worked, but there is a hidden mechanism doing the quiet work. The org chart decides who you can reach, whose permission you need, and how far a good idea has to travel before it can move. You do not see it. You feel it every time the work has to climb your reporting line and cross into another team's before anything happens.

None of this is new to anyone who has worked inside a company. Approval chains have always been slow, and for years that was acceptable. The market rewarded the big, careful, multi-year bet, and a long planning cycle made sense because the technology underneath it barely moved. AI inverted that. The capability now turns over in months, so the model you wanted last quarter is already replaced, the use case has shifted, and the moment the work was built for is gone before the last sign-off lands. The org chart did not get slower. The world it sits inside got faster.

The org chart was built to keep ownership clean. Every box knows its lane, its budget, and its boss. Decisions move up to someone with authority and back down to someone with a task, and within a single department that works. But AI work moves sideways. It crosses product, engineering, operations, and risk at once, and the org chart has no fast path across. It was built for a slower, more orderly world, the same one I have written about before, when the advice most of us still lean on took shape.

The Speed Problem Lives in the Structure

If this feels familiar, you are not the only one seeing it. The research keeps landing on the same uncomfortable point, that the thing slowing AI down is rarely the technology.

Walk into almost any company right now and you will find AI everywhere and nowhere at once, pilots in every function, a chatbot bolted onto the website, a dozen experiments running, and almost none of it woven into how the work actually gets done. McKinsey's latest numbers say the same thing, with 88 percent of organizations now using AI in at least one part of the business and only 7 percent having fully scaled it. Broadening that use, the firm notes, may require redesigning the workflows around AI so the work can run at scale.

So where does the rest of that work live, the part between starting and finishing? Boston Consulting Group breaks it down. In its 10-20-70 approach to deploying AI at scale, 10 percent of the effort is the algorithm, 20 percent is the technology and the data, and a full 70 percent is the people and process around it.

The model, the part everyone obsesses over, is the smallest slice. The 70 percent is the part nobody points to when AI stalls: who decides, who hands off to whom, whose sign-off the work needs, and how a decision moves from idea to shipped. That is not a technology problem. It is an org chart problem.

If the bottleneck is structural, then buying another tool or running another training cannot touch it. A new license does not redraw a sign-off chain, and a workshop does not shorten the road a decision travels before it gets a yes. You can buy the best model on the market and watch it stall anyway, because the software was never what slowed you down. So if your last two AI initiatives stalled in roughly the same spot, that is not a coincidence, and it is not a sign your team is behind.

What We Built Instead of Waiting

Last November, my own team stopped waiting for the org chart to catch up.

We are not restructuring, nor are we changing the org chart. We are building out a new operating model on top of the structure we already have, so the work can move now. My team began merging its efforts with our engineering group and an adjacent technology team, organizing around the experiences we want to ship for customers rather than around the boxes we each report into. None of this is set in stone, and we are still iterating on the fly, reshaping the model as we learn what the work needs. But we are already operating differently than we used to, and we are blurring the traditional lines of the org chart as we go.

We have seen it most in how fast we now ship alongside our engineering partners. They can iterate on customer-facing features at a speed that would not have been possible a year ago, and our job is to keep pace. Accepting the status quo is no longer an option, because AI has changed what coding delivery looks like. So we did not wait for an org change. We worked out how to operate as one team, under a new way of thinking that puts speed and collaboration ahead of dotted lines.

All of that is the part you can measure. The shift I keep coming back to is harder to put a number on, and I describe it the same way every time. "It's amazing what can get accomplished when everyone is marching in the same direction." That alignment, more than any new box on the org chart, is where the speed comes from.

And this is only the start. The experiences our customers want do not live inside one team. They run across platforms owned by a different part of the company, a group we used to treat as someone else's problem and now treat as a partner. So we are building the connections that let both teams work in parallel instead of in silos. The question itself is shifting, from "what does my roadmap look like next year?" to "what do we need to build together to solve our customers' biggest problems?"

What You Can Build Without a Reorg

You probably cannot redraw your own org chart, and you do not need to. It can stay exactly where it is. What you can change is how fast your corner of the work moves, by building a small operating model on top of the structure you already have.

Two questions are worth your time this week. What would your ideal operating model look like, the one that lets your best AI work move at the speed it deserves? And what is the single thing standing in its way, a tool you are missing, or a mechanism you have not built?

You do not have to wait for the org chart to catch up. The fastest teams already stopped waiting, and started building.

The Assumption

A common belief shows up in vendor meetings and internal AI pilots: when an AI model is not behaving the way a team needs, the answer is more training. Train the model on company data, train it harder, train it longer. The thinking goes that training is the universal solvent: if the model is wrong, more training will fix it. The assumption is reasonable because training is the AI concept most professionals already have a clear mental model for.

Meet the Specialty Pharmacy Operations Director

A specialty pharmacy operations director has three different AI problems on her desk this week. She wants one solution for all three. The vendor selling her an "AI platform" has been pitching fine-tuning as the universal fix. She has begun to suspect the vendor's framing is wrong, but she does not yet know which problems need which approach.

Problem one: she needs to draft a tailored summary of this quarter's adherence metrics for the executive team. The format is one her CFO requested specifically. She has done this kind of summary before, but each quarter the audience and the angle shift.

Problem two: her clinical team needs a tool that answers pharmacist questions about prior authorization rules. The rules live in PDFs from twelve different payers. The PDFs update almost weekly. Every answer must point back to the source document for compliance.

Problem three: she has a HIPAA-restricted workflow, meaning the data cannot leave her organization's secured infrastructure. Her team extracts structured information from clinical notes. Fifty thousand notes per month. The output format must be exact. The processing must run locally, without sending patient data to any cloud service.

She has been told fine-tuning is the answer to all three. Two of those answers are wrong. The third is correct, but for reasons the vendor has not bothered to explain.

Where the Assumption Breaks Down

"More training" is not a strategy. It is a vague gesture toward a tool kit. The real question is which lever solves which kind of problem.

The first problem is a one-time drafting task. There is nothing about it that needs to be baked into a model. The second problem requires the model to access information it does not currently have. That information changes faster than any training process could keep up with. The third problem is the only one where fine-tuning is even on the table, and even there, the reason has nothing to do with the model needing to be "smarter."

Three problems. Three different answers. None of them are solved by reaching for the heaviest lever first.

The Three Levers

There are three primary ways to change how an AI model behaves at work. Each one solves a different kind of problem.

Prompting is the cheapest and fastest lever. The model is given a clear instruction and any context it needs to produce the right kind of answer. Prompting works for one-off tasks where everything the model needs to know can fit inside the conversation. Most professional AI use is prompting, and should stay that way.

Retrieval is the lever for knowledge access. The technical pattern is called RAG, short for Retrieval Augmented Generation, which we covered in Vol 19 through 22. The model does not learn the information. The system retrieves the right document at the moment of the response and feeds that document into the prompt. The model then answers using the retrieved source. RAG is the right answer whenever the information changes, whenever traceability matters, or whenever the body of information is too large to fit in a single prompt.

Fine-tuning is the lever for behavior at scale. As Vol 36 established, fine-tuning shapes how a model responds, not what it knows. It is worth the investment when a team needs consistent, repeatable behavior across high volumes of similar tasks, when latency or compliance constraints rule out larger cloud-hosted models, or when a specific format cannot be reliably maintained through prompting alone. For most professionals, this lever is rarely the right one.

Through the Decision

Back to the operations director.

Her first problem is a quarterly executive summary. The right move is prompting. She writes a careful system prompt that includes the CFO's preferred format and a few example summaries from past quarters, then drops in this quarter's raw metrics. The whole task takes one chat session. There is nothing to deploy. Nothing to maintain. The next time the format needs to shift, she edits the prompt.

Her second problem is the prior authorization tool. The right move is retrieval. Her team builds, or buys, a RAG system that indexes all twelve payer PDFs and refreshes the index whenever a PDF updates. When a pharmacist asks a question, the system fetches the relevant passages and the model answers using those passages with citations. No fine-tuning is required. No training data is collected. When the rules change, the index updates and the answers update with it.

Her third problem is the only one where fine-tuning earns its place. Fifty thousand notes per month, structured extraction, HIPAA-restricted local processing. She fine-tunes a smaller open-source model on a labeled set of past notes so it produces the exact output format her team needs. Her team deploys that model on their own infrastructure. The model does not need to be smarter than a frontier model. It needs to be predictable, cheap to run, and locally hosted. Fine-tuning delivers all three.

The vendor's "fine-tuning solves everything" framing was not just wrong. It would have been catastrophically expensive. Two of her three problems would have been solved by no infrastructure investment at all.

The Decision Framework

Here is the test for any AI problem at work.

If the answer can be produced by giving the model the right instruction and the right context in a single conversation, the lever is prompting.

If the model needs access to information it does not have, or the information changes, or sources must be cited, the lever is retrieval.

If a high-volume task needs consistent output that prompting cannot reliably deliver, or if compliance rules out cloud-hosted models, the lever is fine-tuning.

Most professional AI problems are prompting problems. Most of the rest are retrieval problems. Fine-tuning is the narrowest of the three levers and the one with the highest deployment cost.

What to Watch For

• Vendor demos that propose fine-tuning before they have asked what kind of problem you are solving. The right vendor diagnoses first, prescribes second.
• Internal pilots that frame "we need our own model" as the goal. Owning the model is rarely the actual goal. Solving the work problem is.
• Time and cost estimates that grow large. A prompting fix takes hours. A retrieval system takes weeks. A fine-tuning project takes months. If the estimate does not match the actual problem, the lever may be wrong.
• Sycophantic AI responses inside the project itself, where the assistant agrees too readily with the framing that fine-tuning is the answer. RLHF, covered last week in Vol 37, taught the model to please. Push back on it.

How This Connects

This closes the three-part fine-tuning series. Vol 36 dismantled the assumption that fine-tuning adds knowledge to a model and established that fine-tuning shapes behavior. Vol 37 went inside RLHF, the most consequential fine-tuning method ever developed, and showed how human feedback became the engine that turned LLMs into products. This volume gives the practical decision framework that applies all of that knowledge at work. Fine-tuning is one lever. Retrieval is another. Prompting is the third. Most problems do not need the heaviest lever. Next week, Vol 39 kicks off a new series: a four-part Claude Deep Dive that follows one professional through their first week of switching from ChatGPT to Claude.

Part 3 of 3 in the Fine-Tuning series.

90 Minutes of Layout, Gone

You already have the content. The points are in your head or sitting in a notes file. Then you open PowerPoint, and the next 90 minutes disappear into text boxes, alignment, and picking a theme that does not look like 2009. By the time you are done, you give up and present a wall of bullet points anyway. The thinking was the easy part. The deck was the wall.

AI-native slide tools remove that wall. Gamma takes your raw content and generates a structured, designed deck in one pass. It is not PowerPoint with an AI button added on. It is built the other way around: you describe, it designs, you refine.

Why this matters: the work shifts from building slides to directing them. You stop nudging text boxes and start editing a draft that already exists. That shift is the whole point of this volume.

The Workflow

1. Get your content into one block of text (2 minutes). Paste your bullets, notes, or a rough outline into a single document. If it is messy, have Claude or ChatGPT clean it up first:

Copy/Paste Prompt: "Turn these rough notes into a clean presentation outline: a title, four to six section headers, and two or three short bullet points under each. Keep my wording where it is good. [PASTE YOUR NOTES]"

2. Generate the deck in Gamma (3 minutes). Go to gamma.app and sign in free. Click "Create new," choose "Presentation," then "Paste in text." Drop your outline in and let Gamma generate. It builds the structure, layout, and visuals together in about a minute.

3. Refine with Gamma's AI, card by card (2 minutes). Click any slide and use the AI edit option to fix it in place. Try instructions like "make this slide more concise," "turn this into a two-column layout," or "replace the image with something more professional." You are editing a draft, not building from zero.

4. Apply a theme and check the flow (2 minutes). Pick a theme from the side panel. Click through every slide once. Fix the one slide that reads wrong. Trust your judgment here; the AI sets the floor, you set the bar.

5. Export or share (1 minute). Share a link, or export to PDF or PowerPoint. Your deck is done in the time it used to take you to choose a template.

The Payoff

You walk away with a finished, shareable deck in 10 minutes instead of 90, and a repeatable move: describe your content, let AI generate the first draft, then direct the edits. The same pattern works for one-pagers, internal docs, and simple web pages, which Gamma and tools like it also build. You stop starting from a blank slide.

The AI Concept You Just Used

AI-native versus AI-bolted-on. PowerPoint added AI features to a tool built for manual slide-making, so you still drive. Gamma was built AI-first, so it drives and you steer. Learning to spot this difference helps you pick the right tool for any task: manual tools when you want to build by hand, AI-native tools when you want a fast first draft to react to. The asset you produced was generated, not assembled.

Transparency & Notes

• Gamma free tier includes 400 AI credits, which are lifetime rather than monthly. Each deck costs a few credits, so the free tier covers roughly your first 10 to 40 decks before you would need to earn more or upgrade.
• Free decks carry a "Made with Gamma" watermark on web, PDF, and PowerPoint exports. Fine for internal drafts; remove it before presenting to clients or executives, which requires a paid plan.
• PowerPoint export works on free, but complex layouts can flatten on .pptx. Check the exported file before relying on it.
• Do not paste confidential metrics, PHI, or NDA-protected content into a third-party tool.