Where AI earns its place: fit, then cost

Use a model where a wrong answer is visible and recoverable, then divide what you spend by the answers people actually use.

The best AI decision is often not to use it. A model earns its place when the task tolerates variation and the cost of each useful answer beats the alternative.

Check the fit

THREE TESTS FOR FITA candidate should pass all three

ReviewableCan someone tell a good answer from a bad one, quickly?

RecoverableWhat does a wrong answer cost, and can it be undone?

TolerantIs an approximate answer genuinely useful here?

Fails one? Reshape the task firstDraft-and-review often passes where decide-and-act does not.

A poor fit is often a framing problem. Work that can’t be automated safely can often be assisted safely: the model proposes, a person decides, and the review is faster than doing it from scratch. Where an exact answer is required and the rule is knowable, write the rule; code is cheaper, auditable and correct every time. Decide what good means, measured against today’s process, before you build a prototype.

Count the cost per useful answer

THE NUMBER THAT MATTERSDivide by the answers people could actually use

Cost per callInput and output volume, plus retrieval and tool calls

Attempts per answerRetries, agent steps, and rejected drafts

Cost per useful answerWhat the feature actually costs the business

A cheap call that fails twice is not cheap.

Example: When we weighed buying a document-extraction product against building our own, three things set the cost per usable document: the product’s per-document price at our volume, the documents that needed a second pass or failed outright, and document types a general product didn’t handle well. We built it.

The biggest savings come from removing calls: put arithmetic and lookups in code, route easy requests to a cheaper path, cache what repeats, and cap runaway loops. Set a latency budget too, since more context and more steps cost time as well as money. Measure per feature, not per department, because an aggregate bill hides the workflow consuming the budget.