Every AI consultancy says they ship fast. Most of them mean "we start fast." The hard part is keeping that velocity when the problem gets real.
We have learned that a 90-day cadence is the right size for almost any AI engagement. Smaller than that and you do not get past proof of concept. Larger than that and delivery risk compounds faster than you can retire it.
Why 90 days works
Three months is long enough to hit production and short enough to force a forcing function. Every week of the engagement feels urgent, because you can always feel the end date. That urgency translates into better scoping decisions, fewer nice-to-haves, and a more ruthless prioritization of what actually ships.
What 90 days is not
It is not the time to build everything you want to build. A 90-day cycle produces the smallest usable thing that moves a KPI. If you cannot describe that thing in one sentence, you are not ready to start.
What fits in 90 days
- One production model or copilot deployed to one team.
- One workflow automated end-to-end.
- One data pipeline running in production with real users downstream.
What does not fit: a data lake migration, a complete modernization of a legacy system, a multi-model platform. Those are multi-90-day programs — and each 90-day cycle should still produce a usable, shipped outcome.
The rhythm
Inside the 90 days, we run three 30-day phases: discover, design, deliver. Every phase ends with a demo and a decision. Every demo is real — no mock data, no slideware. The decision at the end of each phase is whether to continue, pivot, or stop.
That last option — stop — is the most important one. The 90-day cadence is the gift that lets us say "this is not working" without the sunk-cost trap that kills longer engagements.