AI diligence starts with the work, not the demo.

To assess an AI value thesis, inspect one important workflow, its exceptions and the cost of changing it. A demonstration establishes technical possibility. Diligence must establish who will use the system, what operating change is achievable and how that change could affect the investment case.

Start diligence there. Select an important workflow and follow it from request to recorded outcome. The purpose is to expose the conditions an operating plan would have to satisfy.

Follow one transaction all the way through

Consider a hypothetical distributor whose management team proposes AI-assisted quotation. Watch an operator prepare an ordinary quote, then an order involving a discontinued item, disputed pricing, or an unusual delivery requirement. Record which information is available, which system is authoritative, and where a person supplies knowledge that the systems do not contain.

Keep the observation separate from the explanation. A manager may describe a standard process while the operator works around it. Both perspectives matter. The gap is part of the implementation scope, even if it never appears in the software inventory.

Ask questions that produce evidence

Use the working session to answer a small set of questions:

• Can we inspect source records under the permissions and contractual conditions that would apply in production?
• Which decisions follow documented rules, and who can approve a change to those rules?
• What happens when a record is missing, two systems disagree, or an exception exceeds someone's authority?
• Who will own the redesigned workflow, train the team, and maintain it after launch?
• What baseline would let finance distinguish released capacity from an actual cost reduction?

Request supporting records for each answer. An access diagram, a resolved exception, and an existing operating report are more useful than a readiness score unsupported by evidence.

Separate opportunity from underwriting assumptions

For the hypothetical quotation workflow, faster preparation could release time for selling. It does not follow that revenue will increase. The thesis also depends on demand, sales capacity, adoption, and what the team does with that time.

Build the case in layers. Document observed workload and rework first. State the proposed intervention next. Then identify the assumptions connecting operational improvement to financial effect. Include implementation, review, support, and ongoing ownership costs. Give each assumption an owner and a way to test it.

Where evidence is incomplete, narrow the claim. A bounded opportunity that can survive scrutiny is a better planning input than an ambitious benefit estimate with hidden dependencies.

Make the diagnostic lead to an operating decision

Separate released capacity, removable cost and revenue opportunity in the investment brief. Record the observed baseline, implementation and operating cost, management owner and assumption connecting the workflow change to each benefit. Give uncertain assumptions a range or leave them unquantified.

Classify the proposed initiative before including its benefits in the plan:

• Ready for a bounded test: the team can access representative transactions, an operator owns the change and the proposed benefit has a reproducible baseline.
• Dependent on remediation: the case requires data cleanup, system access, a policy decision or management capacity that is not yet available. Include that work and its cost in the plan.
• Unsupported at present: the benefit depends on demand, staffing reductions or adoption for which the team has no evidence or approved action. Keep it separate from supported operating improvements.

A faster quotation process may be worth funding as a capacity improvement. A revenue forecast needs the additional evidence that available demand and selling capacity can use it. A cost-reduction forecast needs an actual spending decision. Those are separate investment claims.

Leave with a testable next step

The diligence output should name a workflow, an accountable operator, the evidence still required, and a decision gate. For quotation, that might mean testing historical cases in a read-only environment before authorizing any customer-facing action. Define what would justify proceeding and what would stop the work. A useful AI thesis gives management a concrete decision to make next.