Most AI roadmaps read like a technology wishlist: a platform, a few use cases, a vague timeline. A finance leader reads the same document and asks three questions it never answers — what does this cost, what does it return, and how will we know in a quarter whether to continue? A roadmap that can't answer those isn't a plan. It's an aspiration with a logo.
Start from a number, not a use case
Pick one metric the business already tracks and already cares about: hours spent on a recurring task, cost per processed invoice, time to first response, leads lost to slow follow-up. The roadmap is the smallest sequence of work that moves that number, with a before value measured on real data, not estimated in a workshop.
The 90-day structure
- Days 1–15, baseline and scope — measure the current process end to end, name the owner whose workflow changes, and agree the target in writing. Deliverable: a one-page brief with the baseline number and the success threshold.
- Days 16–45, prove it on real data — build the narrowest version that can be tested on your actual, messy inputs. Deliverable: an honest accuracy figure and a list of the cases it gets wrong.
- Days 46–75, run it beside humans — shadow mode first, then assisted mode. Log every correction. Deliverable: a cost-per-task figure from real usage, not a projection.
- Days 76–90, decide — compare the measured result with the agreed threshold. Fund it, change it, or stop it. Deliverable: a decision, in writing, by the person who owns the number.
A roadmap a CFO signs has an exit built in. Every phase ends with a gate where stopping is a legitimate, pre-agreed outcome. That is what makes the spend approvable.
The four lines finance will look for
- Cost to reach the first gate — a small, fixed amount, so the first decision is cheap.
- Run cost at expected volume — measured per task during the run phase, including model usage, hosting and review time.
- Return in the owner's own metric — hours returned, errors avoided, revenue protected. One metric, defined before the build.
- Kill criteria — the specific result that ends the project, and who decides.
What to leave out
Leave out the platform vision, the multi-year transformation arc and the list of eleven other use cases. They are not wrong, but they make the first approval harder. Win one funded, measured result; the next roadmap writes itself from evidence instead of enthusiasm.