“AI audit” can mean anything from a tool list to a sales call in a trench coat. Here is what a real one delivers, which doubles as a checklist for evaluating anyone's.
Trace the workflow with the person who does the work: the inputs, steps, decision points, handoffs, and places where work stalls or gets rekeyed. Include the workarounds that an org chart misses.
An audit written without access to the person who runs the workflow comes from the brochure.
For each step: which decisions need human context, and which actions follow a repeatable pattern. This line decides everything downstream: what gets automated, what stays with your team, and where the approval gate sits.
An audit that proposes automating judgment is a red flag you can spot from this section alone.
Rank multiple candidate builds by the measured cost of the current workflow, feasibility, and risk, with reasons attached. A deliverable that presents only the vendor's preferred sale functions as a proposal.
Hours-saved arithmetic with every assumption stated, so you can recompute it yourself. Scenario math labeled as scenario math, with nothing presented as a result before anything has been built. Then a 90-day plan and the first sprint: what to build first, what it needs to start, what “working” will mean, and how it gets measured against the baseline the audit just established.
Your whole data estate, a long discovery phase, or a commitment to build. The scope is the smallest useful set of systems and examples needed to map the target workflows.
And the recommendation has to survive the possibility that you take the report and build it yourself. That path is fine. The map, the priorities, and the plan are yours after payment either way. The audit page has the full scope; current rates are on the services page.
Describe the workflow that drags and get back where I would start. The audit is the paid, full-depth version of the same discipline.