BvLogic
Diagram · AI

Does This Need AI? — Decision Diagram

The questions that decide whether a problem warrants a model, drawn as a sequence — with the two exits that end most projects before they start, and what to build instead.

Download the diagram Read the full guide

SVG. No sign-up, no email.

The most expensive AI projects are the ones that should have been a rule, a form or a query. Nobody sets out to build those; they happen because the question below was never asked in order.

Each step has an exit. Reaching the end is the uncommon outcome, and that is the point.

Four questions, in order 1 · Can the rules be written down?: Yes → write the rules (cheaper, testable, explains itself), Partly → rules + exception queue (usually the right answer), No → continue (). 2 · Is the knowledge written down?: No → documentation project (nothing to retrieve from), Scattered → fix the sources first (), Yes → continue (). 3 · What does being wrong cost?: High, no reviewer → stop (or add a human first), High, reviewed → continue (), Low → continue (). 4 · Is the gain worth the cost?: Measure the baseline (what rules already achieve) → Express the gain as work/day () → Against build + monitoring forever () → Then decide (). 1 · Can the rules be written down? Yes → write the rules cheaper, testable, explains itself Partly → rules + exception queue usually the right answer No → continue 2 · Is the knowledge written down? No → documentation project nothing to retrieve from Scattered → fix the sources first Yes → continue 3 · What does being wrong cost? High, no reviewer → stop or add a human first High, reviewed → continue Low → continue 4 · Is the gain worth the cost? Measure the baseline what rules already achieve Express the gain as work/day Against build + monitoring forever Then decide measure, do not estimate Exit — build this instead The exit most often skipped Proceed
Each question has an exit. Most problems leave at the first two — which is a good outcome, arrived at cheaply. Shaded exits are the ones that save the most money.

Question 1 — can the rules be written down?#

If someone can enumerate them, enumerate them. Rules are cheaper to build, testable, explain their own decisions, and fail visibly when no rule matches — which is a feature.

The partial answer is the one worth noticing, because it is usually correct: a handful of rules covering most cases, plus an exception queue for the remainder. That combination frequently delivers most of the benefit for a fraction of the effort, and it makes the residual problem visible and measurable.

Question 2 — is the knowledge written down?#

Retrieval works over documents. If the answer lives in the heads of three experienced people, there is nothing to retrieve, and no model changes that.

This is a documentation project wearing an AI project's clothes. Saying so is unpopular and it is cheaper than discovering it in month three.

Question 3 — what does being wrong cost?#

Not "is it accurate" but "what happens when it is not". A system that is wrong occasionally and reviewed by a human is fine. A system that is wrong occasionally, acts automatically, and affects someone's money, health or employment is a different proposition entirely.

If the cost is high and no reviewer exists, the answer is not a better model. It is a reviewer.

Question 4 — is the gain worth the cost?#

The step people skip, because by this point everyone wants to build it.

Measure the baseline — what the rules already achieve — then express the model's advantage in units of work rather than percentage points. "Eight points more accurate" persuades; "sixteen documents a day, about twelve minutes" invites the right comparison.

Set that against the build, plus a monitoring obligation that does not end. See the worked example for this arithmetic done in full.

The lanes have no arrows#

Deliberately, for the first three. They are not stages of a pipeline — they are independent gates, and a project can fail any one of them regardless of the others. A problem can have unwritten rules (passes 1), no documentation (fails 2), and the sequence stops there.

See AI in business, the readiness checklist and the use-case one-pager for these questions in fill-in form.

Back to AI