Decide whether a business capability should use AI at all

In plain terms

Decide, for this capability, whether AI is the right answer. A decision not to use AI is recorded, because an organization that never documents restraint cannot distinguish judgment from inattention.

What is being judged

Whether the capability’s problem is one AI solves, and whether the organization would accept the failure mode that comes with it.

Three questions, and the third is the one usually skipped.

Is the problem shaped for AI? Prediction, classification, generation and extraction at volume. A process that is failing because it is badly designed will fail faster with a model in it.

Would a simpler mechanism work? A rule, a lookup, a form change. This is not a cost argument; a deterministic mechanism has a failure mode a person can reason about, and that is worth paying for in many contexts.

Is the failure mode acceptable to the people affected? A model wrong 3 percent of the time may be better than the humans it replaces and still unacceptable, because the errors fall differently and the affected person has no one to argue with.

Proportionality applies to the record, not the decision. Every capability decision is made. What varies is how much is written down, governed by the materiality mechanism. A low-materiality capability may warrant two sentences.

What this decision does not cover

It does not fund or prioritize; that is L5-01. It does not classify, authorize or accept risk. It decides whether the thing should exist as an AI capability at all.

When it fires

On event. When a capability is proposed for AI, from any source including a vendor proposal or an internal initiative. On material change to an existing capability’s approach.

On cycle. None. Event-driven.

What you need before deciding

What the capability is for and who it affects. What currently performs it and how well. The failure mode of the current mechanism and of the proposed one. Whether the capability is material, which sets the documentation weight.

How this goes wrong

The unrecorded no: capabilities decided against leave no trace, so the same proposal returns every eighteen months and is evaluated from scratch at full cost. From outside, an organization that examined and declined AI for a process is indistinguishable from one that never considered it. The decision that was never made: an initiative starts, gets funded and reaches production without anyone asking whether AI was the right approach, because the question had no owner and no trigger. Deciding on capability rather than consequence: a model can do this, therefore it should.

Related decisions

Upstream L6-01 strategy, L4-CLS-02 materiality for the documentation weight.

Downstream L4-CLS-03 formal governance determination, L5-01 funding.

Reverse L6-10 discontinuation.

Instrument references

None. Every instrument in the crosswalk governs the AI an organization has. None creates a record of the AI it considered and declined. Crosswalk gap 10, and the reason restraint is invisible in most organizations.