What the improv stage taught me about shipping production AI

The first thing they teach you on an improv stage is that you can lose a scene by saying the right thing too soon. It is Tuesday night at the National Comedy Theatre on India Street in San Diego, and a player walks out and offers, Officer, I swear I didn’t take it. The other player, who is supposed to be the officer, hears this and decides — for reasons that will haunt them later — that the truthful, helpful, correct answer is, “I know you didn’t, sir, you can go.” The scene ends in eleven seconds and the audience laughs the kind of laugh that is not really a laugh. The player who refused the offer has done what every junior engineer learning AI inevitably does on their first production deploy. They optimized for being right at the cost of being in the conversation.

The improv principle that fixes this is Yes, and… You accept the offer your scene partner makes, and you add to it. The officer says, “I know you didn’t take it, but I have to write you up anyway, because the precinct sergeant is looking for a stat.” Now there is a scene. Now the audience is in. Yes, and… is not blind agreement. It is the structural commitment that the scene continues — that whatever is offered, the system that receives it has somewhere to go next.

I have spent enough time on the consulting side of AI to be confident that Yes, and… is also the only sustainable design principle for any AI system that has to hold a conversation, an interface, or a workflow with a real user. The systems that ship and stay shipped are the ones whose authors understood, at the level of design instinct, that the user’s input is an offer, and the response has to continue the scene.

What ChatGPT learned and most enterprise systems didn’t

Consumer AI got this right by accident. ChatGPT, in its earliest public form, would sometimes refuse a request — but it would almost always say something in the same shape as the request. You could ask it for a haiku about your cat and you would get a haiku-shaped object. The “scene” continued. The system was, in improv terms, taking the offer.

A meaningful number of enterprise AI systems do not take the offer. They time out. They throw a 500 with an opaque trace ID. They return I’m sorry, I cannot help with that request in a context where the user has clearly indicated that they are trying to do something the system can, in fact, help with. They are the officer who said I know you didn’t, sir, you can go. Each of these is a scene-ender.

The cost of a scene-ender, in production AI, is precise and quantifiable. We measure it in user re-engagement rate. We measure it in support tickets that begin with a screenshot of a refusal. We measure it in the difference between a 12-second median session and a 3-second median session — the difference between the user who got somewhere and the user who walked out.

Graceful degradation is a script direction

When a director writes a stage direction that says the door does not open, the actor’s job is not to stop the play. The actor’s job is to find the next move given the door not opening. The stage direction is a constraint on the scene, not the end of it.

Graceful degradation is the same idea, in a different vocabulary. When the model you most wanted to call is rate-limited, the answer is not to throw. The answer is to fall back to the model that will respond, summarize what the cheaper model can summarize, and tell the user — explicitly — that you are doing so. Yes, the primary path is unavailable, and here is the secondary path I am taking, and here is what you should know about how this changes the answer. That is a continuation of the scene.

I have built systems where the fallback chain is four deep, and the system tells the user, in plain English, when it has dropped to fallback two. The users do not mind. They mind opacity. They do not mind degradation that is announced. The improv crowd that would have walked out on a scene-ender will sit through a scene that goes somewhere weird, as long as it goes somewhere.

What the Officer / I swear I didn’t scene teaches about evaluation

A useful evaluation suite for an AI system is a set of offers — inputs of varying weirdness — and a set of acceptable continuations. Not “correct answers.” Not a single ground truth. Acceptable continuations. You write down what a human director would consider an in-scene response, and you score the system against the space of in-scene responses, not against a single golden output.

This is closer to how editors evaluate dialogue than to how unit tests evaluate code. A line of dialogue can be in-character and not be the line another writer would have written. The model can give an in-scene answer that is not the one in your head — and as long as it advances the scene, the user gets where they need to go.

The teams I work with that ship and stay shipped have all, in some form, internalized this. They stop asking did the model say the right thing and start asking did the model take the offer. The first question is unanswerable in production. The second question is answerable in fifteen minutes with a CSV.

The trap of no, but…

The opposite of Yes, and… is No, but… — sometimes spelled Well, actually, sometimes spelled I’m sorry, I cannot, sometimes spelled Did you mean X? All three are scene-enders. All three are, at the level of system design, the same failure: a system that cannot continue the scene the user opened. The fix is the same in every case. Find a way to take the offer. If you cannot take it as offered, take it modified, and tell the user what you modified.

The improv stage taught me that you can lose a scene by saying the right thing too soon. The production-AI stage taught me that you can lose a user by saying the right thing in the wrong shape. Neither lesson is original. Both are operationally load-bearing. The teams that learn them ship. The teams that don’t, don’t.

The systems we build at this practice take the offer. They tell the user what offer they took, and what offer they could not take, and where the boundary is. The scene continues.

That is, in the end, all anyone in the room is asking for.