Budget judgment, not intelligence

Companies are budgeting for intelligence the way they once budgeted for bandwidth: licenses, tokens, seats. It is the wrong unit. Intelligence keeps getting cheaper. The bottleneck is the organization's capacity to decide what to do with it.

Consider a hypothetical credit team. Suppose a model prepares a hundred memos in an afternoon. Someone must still decide which to approve, answer for the ones that go bad, and explain the policy to a regulator. If that person's calendar has not changed, the organization has not become faster. It has moved the queue.

Three lines belong in every AI plan:

• Decision capacity. Who reviews machine output, how many consequential decisions a week they can make well, and what happens when volume exceeds that.
• Verification cost. The time and tools needed to check outputs before they reach a customer, and who owns the failures.
• Apprenticeship. The share of work deliberately kept human so that the next generation of experts can form.

None of these appears on a vendor invoice, which is why they rarely appear in a business case. The firms that do well this decade will not be the ones with the most intelligence per dollar. They will be the ones that know how much judgment each dollar of intelligence consumes.

My view: before approving another AI budget line, ask to see its judgment budget. If no one can produce one, the plan is not finished.