How AI Actually Works: A Field Guide for People Who Make Decisions

Fifteen cards, four weeks, one free PDF. For people approving the strategies, signing the contracts, and answering for the outcomes.

TL;DR. A four-week series. Fifteen Tufte-coded cards. Today's piece is the manifesto and three preview cards. June 16: What Happens When You Hit Send. June 23: Where the Answers Come From. June 30: Why You Can Trust Some Systems and Not Others. All fifteen will live as a free PDF after the series ends, called The AI Field Guide.

---

A field guide is a book you take into the wild for situations where the thing you are trying to identify isn’t going to pose, isn’t going to explain itself, and isn’t going to pause for questions. You need the marks ahead of time. The bird is gone in three seconds.

Right now, in most boardrooms and most executive teams, AI is the bird.

The flap is fast. Internal pilots that worked in the lab and never shipped. Contract reviews where the legal team is reading the same paragraph for the third time because nobody is sure what the system actually does at runtime. By the time the room agrees on what just happened, the bird has moved.

I have been making infographics this spring to put the marks on the page. Not “how to use AI”: there are a thousand of those, and most of them age out inside a quarter. Not “what AI will do to your industry”: there are a thousand of those too, and most of them are wrong. The plumbing. What actually happens between a person hitting send and a system returning an answer: where the request goes, what the model can see, what it remembers (almost nothing), what it costs in water and watts, who is allowed to see what under whose identity, what happens when the model is wrapped in an agent and given a goal and a budget and a few hours.

The cards are Edward Tufte-coded. Dense, restrained, no decoration. They are field guide cards, not posters. They speak for themselves, which is why I am not going to spend the next four weeks paraphrasing them. I am going to drop them into the page and write the thesis around them.

Here is the thesis, in one paragraph.

Most executive AI literacy is anecdotal. It comes from product demos, vendor decks, dinner-party stories, and one or two Substack posts somebody forwarded. That is how we ended up with boards approving AI strategies they could not draw on a whiteboard. The fix is a small number of well-drawn pages that show how the system actually behaves and where enterprise value sits. We have arrived at the part of the cycle where “trust me, it works” is no longer enough. The people writing the checks need the same operating fluency a doctor has about anatomy or a banker has about leverage. Not surgical, not actuarial, but enough that they cannot be sold a fairy tale.

Three previews, one from each cluster.

Card 1: The Context Window Is Not Memory

If you ask a roomful of executives what an AI “remembers,” you will get an answer that assumes the model recalls previous conversations the way a colleague does. It doesn’t. There is a model: a frozen set of weights from a training run that ended months or years ago. There is a context window: a temporary working set for this request, the size of a small book, thrown away when the chat closes. There is whatever the application around the model decided to keep — your name, your preferences, a thread summary, sometimes nothing. The model is not “learning from you.” The application is taking notes on your behalf, or not, depending on a configuration somebody set up.

This is the card I would frame and hang in every legal and policy meeting at every company that has bought an AI system in the last two years. Because the answer to did the AI retain that customer’s PII is not yes or no. The answer is which of these four stores held it, for how long, under whose identity, and with what audit log. That is a different question than people are asking, and the answer is on this card.

Card 2: A Prompt Footprint, in Human Terms

The “AI uses a lot of water” headlines have been...pardon the pun...recirculating for a year. The numbers in those headlines are usually old, or wrong, or both. The good news, from OpenAI’s and Google’s own disclosures across late 2025 and early 2026: a single short text query is small. About a quarter of a milliliter of water. About a third of a watt-hour of electricity. About three hundredths of a gram of CO₂. Roughly equivalent to nine seconds of television.

The harder news: ten thousand of those queries equal a sheet of paper in water, a smartphone charge in energy, and a short flight’s worth of carbon when run on dirty grids. Billions of prompts scale this from negligible to meaningful, and heavier inference (reasoning models, video, audio) climbs another order of magnitude per request.

That is not a moral argument. It is an operating cost argument. Anyone writing an AI policy without these numbers is winging it.

Card 3: Permissions and Data Boundaries

This is the card most likely to make an executive reader say we don’t have this and we should. Same prompt (”Summarize the customer escalation for ACME”), and five users get five different answers, because identity, role, residency, and policy filter the context before the model sees it. The sales rep gets a redacted, public-facing summary. The legal counsel sees the DPA and retention notes. The external partner sees a refusal: outside your permitted scope.

That refusal is not a model decision. It is a system decision the model is not even consulted on. The model only sees what the layer in front of it decided to let through. If your AI strategy talks about “connecting everything,” you have skipped this card. The right framing is connect the right things, under the right identity, with the right controls. That distinction is the difference between an AI deployment that scales and one that ends up in a regulatory disclosure.

Why this series, and why now

I’ve been building on two ideas across Beyond Reason this year.

The first is that skills are the new software: the durable unit of organizational AI capability is the packaged combination of instructions, tools, scope, and persona that a non-developer actually invokes. Markdown beats a hosted custom GPT for the same reason a recipe beats a restaurant: portability is leverage. That thesis got its first formal validation when Anthropic shipped Agent Skills as a published spec in October 2025, and again when Google mirrored the concept inside Workspace at Cloud NEXT in April 2026.

The second is that knowledge work is code: the most valuable thing your knowledge workers produce this decade is the artifact that captures how a decision gets made, version-controlled, instrumented, replayable.

The cards in this series are the fabric underneath both arguments. If you are going to treat skills as software, you have to understand the system that runs them. If you are going to treat knowledge work as code, you have to understand the runtime.

Part 1 covers the request: what happens when somebody hits send, where the context comes from, what it costs, what the model can actually see. Part 2 covers the answers: where the model’s knowledge came from, what it does not know, how retrieval gets it past the cutoff, and where the answer is physically running. Part 3 is the trust arc: agents, identity, routing, evaluation, and the layer that ties them together. The last card of Part 3 is the closer of the whole series. I am not going to spoil it here. It is the card that distinguishes a consumer AI deployment from a system you can hand to a regulator without flinching.

The PDF lands the week after Part 3 ships. Fifteen cards, printable on letter or A4, suitable for the boardroom or the bench.

Receipts

• Anthropic — Agent Skills, overview and spec. The October 2025 announcement that formalized the packaging unit this series builds on.

• OpenAI — disclosed per-query energy and water estimates in 2025–2026 press materials and internal blog posts; the numbers on Card 10 reconcile those disclosures with Google’s median-query estimates and methodology notes.

• Edward Tufte — The Visual Display of Quantitative Information (Graphics Press, 2001). The visual grammar this series borrows from, badly and on purpose.

• Beyond Reason, prior pieces — Skills Are the New Software (March 10, 2026), Knowledge Work Is Code (April 15, 2026).

The bird is still moving. But you will know what it is. See you next Tuesday.