An Answer in the Right Shape: Why Every Question Should Not Return a Memo
Some questions want a verdict. Some want a forecast, a diagnosis, a comparison, or a plan. Forcing every answer into one format is the most common quiet failure in AI decision support — here is how we fixed it.
Ask a competent human advisor two different questions and watch what comes back.
"Should we raise prices?" gets a recommendation with a reason and a condition attached. "Why did churn spike in March?" gets a diagnosis — a ranked set of candidate causes with the evidence for each. "What does this look like in eighteen months?" gets a forecast with named assumptions. Nobody has to be told to do this. Matching the form of the answer to the form of the question is most of what expertise looks like from the outside.
Software is much worse at it. Almost every AI product has exactly one output shape, and every question gets poured into it. Chat interfaces return prose whether or not prose is the right container. Decision tools return a memo even when the question was a five-second sanity check.
The cost is not aesthetic. A forecast rendered as prose loses its assumptions. A diagnosis rendered as a recommendation skips straight to the fix and never names the cause. A comparison written as a paragraph makes you hold six variables in your head that a table would have held for you.
So The Consult picks the shape. There are 11 of them.
The shapes
Answer — the direct read. A real question with a real answer and no ceremony. Most turns land here and should.
Prose — the written thing itself. When you ask for the investor update, the difficult reply, or the announcement, the deliverable is the draft, not advice about drafting.
Decision — a verdict with a confidence level, the reasoning that produced it, and the conditions under which it flips. This is the shape people picture when they picture "AI decision support", and it is a minority of turns.
Analysis — the structured read of a situation: what is actually going on, which parts are load-bearing, what follows.
Comparison — options against criteria, with the trade-off named rather than averaged away. Two vendors, three pricing models, four candidates.
Plan — sequenced work with owners and dependencies, for when the decision is made and the question is how.
Steps — the short procedural version. Do this, then this. No narrative.
Ideas — divergent generation, when you need surface area rather than a verdict.
Research — a grounded brief with live sources and citations, for when the answer depends on facts outside your walls.
Forecast — a projection with its assumptions named and separable, so you can disagree with an assumption instead of the number.
Diagnosis — ranked candidate causes with the evidence for each, for when something broke and the fix depends on knowing why.
Why the router matters more than the list
A list of formats is a feature. Choosing correctly, without being asked, is the product.
The routing happens before the expert writes. The question is read for what it is actually asking — not keyword-matched, but classified by intent, prior context, and how much is riding on it. A question that opens with "why" and describes something that already happened is a diagnosis. A question that names two options is a comparison. A question that asks what happens if is a forecast.
Two design choices make this work in practice rather than in a demo.
The floor is low. The router's default is the cheapest honest shape, not the most impressive one. If a question can be answered directly, it is answered directly. Shape escalation has to be earned by the question, because the alternative — everything becomes a memo — is exactly the failure we set out to fix.
You can overrule it. If the router reads your question as a comparison and you wanted a plan, you say so and you get a plan. The router is a good default, not a verdict about what you are allowed to want. Explicit asks always outrank inference.
What this changes about cost
Shape and effort are the same lever. A direct answer does not run the research pass, does not build the counter-case, and does not spend the tokens a forecast spends running scenarios. That is why a typical Consult turn lands around 5–40 credits while a full board sitting runs 55–150.
This is the part that quietly changes behaviour. When every question costs the maximum, you ration your questions — and the ones you stop asking are the small ones. But most decision damage is not one catastrophic call; it is fifty small calls made on autopilot because getting a second opinion on each one felt like too much friction. Making the cheap shape genuinely cheap is what lets the small questions back in.
The failure this prevents
Here is the concrete version. A founder asks: "Our enterprise churn doubled last quarter. Should we build SSO?"
The single-shape answer is a recommendation. It will be thoughtful, it will weigh engineering cost against deal size, and it will probably say yes, build SSO, here is a phased approach.
The shaped answer notices that this question contains two questions, and that the second one is unanswerable until the first is. So it returns a diagnosis first: here are the four plausible causes of doubled enterprise churn, here is the evidence for each in what you have told me, and here is the one cheap check that would separate them. SSO is candidate three of four. If the real cause is candidate one — onboarding never completed for the cohort you signed in November — then building SSO is a quarter of engineering spent on the wrong problem, and you would have felt very well-advised the whole time.
That is the actual damage a wrong shape does. It is not that the prose was worse. It is that the answer confidently addressed a question you had not established yet.
Shapes and the board
When a Consult escalates to the full Boardroom, the shape question resolves differently: a board sitting exists to produce a decision memo, and that is what it produces. Rounds of structured disagreement across 24 experts, positions held under pressure, dissent preserved rather than averaged out, one memo at the end with a verdict, a confidence level, and the minority view intact.
That is the right shape for the biggest calls, and the wrong shape for almost everything else — which is the whole argument for having both rooms. More on the escalation itself in Meet Your Named Expert.
How to get the most out of it
Say what you want when you know. "Give me the comparison" or "just the short answer" beats hoping the router agrees with you. It always defers to an explicit ask.
Let the small questions be small. The router will keep them small. That is the point.
Push back on the shape, not just the content. "This is a forecast, I wanted the diagnosis" is a completely legitimate correction and it takes one turn to fix.
Watch the assumptions in a forecast. They are broken out separately on purpose. The number is downstream of them; if you disagree with one, say so and the number moves.