Why we don't write slide decks

The first time a client asked me where the slides were, I was standing in a parking lot in El Segundo, outside a content-tagging vendor's office. We had just spent six hours rebuilding their evaluation harness on a borrowed laptop. The harness worked. Two of their three model variants were demonstrably worse than the third; the third was demonstrably worse than the off-the-shelf API they had quietly stopped using six months earlier; and for the first time that week, the team in the room had stopped checking their phones. They asked, politely, where the deliverable was.

I said the deliverable was the harness.

Then came a pause I have learned to recognize. It is the pause in a rehearsal room when an actor has been running a scene from memory and you ask them to do it again — this time, on their feet. It is the moment a person realizes that the thing they have been calling "the work" — the script, the deck, the strategy memo — was not the work. The work is what happens when the script meets a body, a room, and an audience that never got the rehearsal notes.

This essay is about why the consulting practice I run does not produce slide decks. It is also, unavoidably, about why I think most enterprise AI consulting does — and why that costs more than it looks.

A deck is a rehearsal you never ran

I came to this work from a theatre degree at the University of Nevada, Reno and roughly eight years of phone-screening engineers. That is an unusual on-ramp to AI, and it left me with a specific bias: I distrust any artifact that has not been tested against a body in a room. A deck is one of those artifacts. Structurally, it is a rehearsal nobody ran. It promises a performance. It does not survive first contact with a real user, a real dataset, or a real auditor.

The expensive part of an AI engagement is everything after that first contact. The dataset turns out to have a leakage problem (training data bleeding into the test set) the deck never anticipated. The latency budget collapses on the first multi-tenant load test. Legal review surfaces a vendor-licensing exposure the deck waved off in a footer. None of these failures are unusual. All of them are predictable. None are visible in the deck — and the consultant who wrote it is, as a rule, no longer in the room when they surface.

The director's term for the rescue is deus ex machina: the second-act moment when a god descends on a crane to resolve the plot the playwright could not. In an AI consulting engagement, the god is usually a footnote reading "subject to data availability" or "pending model selection." That footnote is the trapdoor the unfunded work falls through.

What we ship instead

The no-deck rule is a structural commitment to artifacts that have to survive a real audience. There are two of them, and they trade off precisely.

The first artifact is the working system. It runs in your environment. It hits your data. It produces outputs your operating team can read, audit, and override. It comes with an evaluation harness — small, sometimes embarrassingly small, often a single Python file with a CSV next to it — that lets you and me argue about whether the system is getting better or worse on numbers we both trust. When the harness disagrees with our intuitions, we update the intuitions, not the harness.

The second artifact is the one-page memo. Memos are how senior practitioners argue with senior buyers about decisions, and the discipline of fitting an argument onto one page forces you to know what you are actually claiming. Bezos famously banned PowerPoint inside Amazon for the same reason: a memo cannot hide behind a build animation. A memo, like a script, has to land on its feet.

A memo also has a property the deck does not: it is read at the reader's pace, not the speaker's. A board member can re-read paragraph three without rewinding a YouTube link. A general counsel can put a sticky note next to the line about data licensing. A CFO can run the math herself. None of that happens with a slide.

The objection I hear most

The most common objection is that the deck is for the people who are not in the room — the steering committee, the parent company, the board. Fair. This is a real problem, and the no-deck rule does not pretend otherwise. The answer is that the executive summary of the memo is the artifact for those audiences, not the deck. A two-paragraph summary is not a deck. It is a piece of writing short enough to be read and long enough to claim something. If it cannot survive that test, it is not ready for the audience that is not in the room.

The version that does require slides — the all-hands, the conference talk, the sales-floor enablement — is downstream of the working system, not upstream. Make the system work first; let the slides come from the operator who actually shipped it.

What this costs me

I am aware that the no-deck rule costs me business. Some buyers genuinely need the artifact of the deck — for political cover, for procurement, for a steering committee that demands one. Those buyers are typically not in the small-batch cohort this practice can serve, and that is fine. The practice is structurally six engagements per year. The math works because the engagements that fit the engagement model fit it well, and the ones that need a fifty-slide stage prop go elsewhere.

The directors I trained under had a phrase for this: don't act in front of the audience; act with them. The deck acts in front of the audience. The memo and the working system act with them. That is the difference between a performance and an exchange — and the exchange is the only one of the two still standing in a Tuesday-morning Slack thread three months later.

So no — I don't write decks. I ship the rehearsal.