Start with watching, not asking
Ask a team what takes the most time and you get opinions. Watch them for two half-days and you get numbers. A real assessment sits in the systems, times the manual steps, counts the errors, and captures the workarounds nobody put in a process document because they are embarrassed by them. Those workarounds are where the money is.
The scorecard
Every recurring workflow that touches money or records, with four numbers: hours per month, loaded cost of those hours, error rate (and what an error costs), and the estimated cost to automate. Sort by return. The list is usually a surprise to the owner, in both directions: the thing they hate is often small, and something nobody mentioned is large.
The top three, designed
For the three highest-return rows: what triggers the automation, what it reads, what it writes, what it checks against, and what a human still approves. Designed to the screen so a quote is possible and so the team can object before anything is built. This is where "AI" gets specific: which steps are language (a model can do), which are arithmetic (code must do), which are judgment (a person keeps).
Fixed quotes
A price for each of the three, written, so the decision is a number. The assessment fee should come off whichever you build. If a firm will not credit it, ask why.
The roadmap
First, second, third, and leave alone. The "leave alone" column is the honesty test: an assessment that recommends automating everything is a sales document.
What to refuse
- A "maturity score" with no hours or dollars attached.
- A deliverable that is a slide deck rather than a document you could hand to another engineer.
- An assessment only sold bundled with implementation.
- A diagnostic that takes longer than four weeks for a company under 200 people.
- A price you have to ask for.
Mine is on one page: $4,500, two weeks, the four deliverables above, credited to a build within 60 days.