Ask for evidence from the current process
Start with the people doing the work. Have them walk through a recent example, including what happens when something is missing or wrong. The procedure in the shared drive may describe the clean version. The inbox usually contains the exceptions.
Record the volume, handling time, rework and people involved. Separate active staff time from time spent waiting. They are different problems. A faster draft may help the first without fixing the second.
If nobody can produce representative examples, the first deliverable may be a plan for collecting them. That is more useful than assigning a confident savings estimate to a process nobody has observed.
Make the economics inspectable
The report should show how each estimate was calculated. What volume did it assume? Which staff cost? How much of the work can change? How much review remains?
Subtract the cost of running and supervising the system. Model usage, integration maintenance and correction time are part of the job. Saved capacity is also different from a reduction in payroll. A team that gets time back may use it to serve more customers rather than reduce spending.
An estimate is a reason to test a project, not a result to put in a case study.
Expect a recommendation that can say no
An audit should compare AI with simpler options. An existing software setting, a standard integration or a clearer process may solve the problem with less maintenance.
Anthropic's guidance recommends beginning with the simplest approach and adding agent complexity only when it improves the task enough to justify the cost. That is a useful check on a proposal full of agents. Anthropic, Building effective agents, checked September 6, 2026.
Ask the consultant to show at least one process they would leave alone. If every workflow requires an ambitious build, examine the incentives behind the recommendation.
Leave with a first scope
The strongest opportunity should be specific enough to quote. Name the input, the output, the systems involved, the person who approves access and the conditions for acceptance.
Include the exceptions. What happens when a record cannot be matched? Who sees a failed run? Which actions require review? These details determine whether the workflow can operate outside a demonstration.
Keep the report portable
You are paying for a decision and the evidence behind it. The report should be useful to your team even if you choose a different builder. Clarify that before buying the audit.
Our AI Opportunity Audit is a standalone engagement. It ends with the opportunity list, the reasoning behind the priorities and a proposed first build. You decide whether to take the next step.