A 90-day AI roadmap, week by week
How to go from intention to a first use case in production within three months, with the decisions required at each stage and the milestones that mark progress.
The distance between deciding to do something with AI and having something working usually gets filled with meetings. This is a three-month plan that ends in a real case in production, not in a document.
Weeks 1 and 2: choose and measure
What you do: list candidate processes, assess them for impact and feasibility, pick one. Then measure how it works today: volume, time per case, error rate, cost.
The decision: which case, and which number it has to move.
The deliverable: half a page with the chosen case, the measured baseline and the success criterion expressed as a number.
This is the step most often skipped, and it determines whether anything that follows can be evaluated.
Weeks 3 and 4: validate feasibility
What you do: confirm the data exists, can be accessed, and is good enough. Test the candidate solution on a sample of twenty to fifty real cases.
The decision: whether the case is technically viable or the scope needs adjusting.
The deliverable: sample results, with the share of cases handled correctly and a breakdown of the errors.
If the numbers are poor here, this is the cheapest possible moment to change case.
Weeks 5 to 8: build the pilot
What you do: implement the working version on real cases, with the team that will use it later. Run in parallel: the system processes, but decisions stay human.
The decision: what stays in scope and what gets postponed. Additional requests always appear; the discipline of deferring them is what gets you to the finish line.
The deliverable: a system running on the real flow, measured daily.
Weeks 9 and 10: measure and tune
What you do: compare against the baseline, review the failures, adjust. Identify exceptions and define how they'll be handled.
The decision: scale, adjust or drop, against the criterion written in week 2.
The deliverable: a results report with the numerical comparison and a recommendation.
Weeks 11 and 12: production
What you do: integrate where needed, define permissions, configure monitoring, train the team, document the process and the exception path.
The decision: who owns the system in operations and who reviews the indicators.
The deliverable: the process running every day, with someone accountable.
What to resist during the 90 days
Expanding scope. Every addition looks small and adds two weeks. The right answer to almost every new request is "in version two".
Waiting for perfect data. If the chosen case requires impeccable data, the case was chosen badly.
Forming a committee. At this stage, three people with clear responsibility move faster than a committee of eight.
Buying an end-to-end platform. Before you know what you need, any large purchase is a bet.
What remains after day 90
More important than the system itself: the organization now has a measured baseline, a method for evaluating cases, a decision criterion, and real evidence about what works in its own context.
The second case will take half as long. The third, less. That curve is the real output of the first quarter.
Frequently asked questions
Is it realistic to have a case in production in 90 days?
Yes, for a narrow case with accessible data and no complex integrations. Projects that require unifying several systems or preparing historical data need longer, and that should be said up front.
What team is needed for these first 90 days?
A process owner in the affected area, someone from IT for access, and a technical profile to implement. Part-time but consistent: the project doesn't survive three-week interruptions.
What if the pilot misses its target?
That's a valid and cheap outcome. Document what was learned, adjust the scope or change the case. Ninety days invested prevents a two-year programme built on a wrong premise.