The acceptance criteria
are the product.

Most AI engagements get sold on a sentence like "help our advisors prepare for meetings."

The implementation team inherits every question that sentence leaves open. Six months later the demo works, the customer shrugs, and nobody can say whether the thing passed.

I've built that demo. In 2026 I spent 2 weekends on a rep-coaching feature before agreeing with anyone what a sales manager should do differently after reading it. The output looked useful. I couldn't show it changed a single coaching conversation.

Building a meeting-preparation agent for a Minneapolis wealth-management firm showed me the counter-move. I worked directly with the owner to scope the engagement, and before rollout the owner signed pass/fail thresholds for retrieval accuracy and unsupported claims.

The build had a target the customer helped define. That's what I mean when I say the acceptance criteria are the product: they describe the result the buyer is paying us to deliver and the evidence we owe them.

"Help our advisors prepare for meetings" gives a sales conversation a direction. It leaves the implementation team holding the decisions.

What information belongs in the brief? Which sources should support it? What happens when those sources disagree? Who decides whether an advisor can use the result?

I want those questions answered in the commercial conversation because each answer moves the price. A requirement to cite the source behind a statement creates work. So does handling a missing document, and so does routing a questionable result to a person. The buyer should see that cost before committing.

Construction figured this out a century ago. A building reaches "substantial completion" when it passes the inspections named in the contract, and the contract names them before ground breaks. An AI engagement deserves the same clause.

Technical discovery may still change the plan. When it does, the customer and the delivery team revise the promise together. An acceptance criterion is an agreement to maintain as we learn.

Before calling an evaluation complete, the buyer and the implementation team work through:

• the situations and source material that will judge the workflow
• what counts as a missing fact, an unsupported claim, or a result that needs review
• who assesses those results, and what threshold they'll accept
• the record that lets the customer inspect the decision and its exceptions

Those are design questions. The specific checks depend on the workflow, which is why selecting the test conditions belongs alongside selecting the acceptance threshold.

I also want the person doing the work in that agreement. The owner can approve the engagement; the advisor can say what they'd still need to verify before trusting the brief. Both belong in the definition of a useful result.

If you have a deal in flight, stop here and try it: write "We'll consider this ready to use when…" and finish the sentence. If you can't, that's the next meeting.

The meeting-prep agent produced cited, two-page briefs from CRM and document-store context. In one session, one advisor finished total preparation in under five minutes, including reading the brief and checking questionable details.

An average takes more sessions.

Repeatability takes more advisors.

A firm-wide time saving takes a documented before-and-after.

One session supports one claim: that advisor, that day, was done in under five minutes with the review included.

The boundary is commercial, and it cuts both ways. If the promise concerns preparation time, the measurement has to include the advisor's review and corrections. Timing only the generation answers a smaller question than the buyer asked.

For an AI company, the agreed workflow, the acceptance conditions, and the open uncertainties travel from sales to implementation as part of the deal. The delivery team can see what the customer expects, how that expectation gets checked, and who can accept the result.

Before a deal moves forward, the seller, the buyer, and the person responsible for delivery each finish the same sentence: "We'll consider this ready to use when…"

The differences in their answers are the work to do while scope and expectations can still change.