Ninety days is enough if it is spent mostly collecting evidence rather than building: a narrow version built on real data in the first one to two weeks, shadow mode for two to four weeks, supervised live use for three to five weeks, then a go/no-go review against thresholds agreed before the pilot started. It is usually not enough for low-volume, seasonal or long-feedback-loop workflows, where the fix is a narrower scope rather than a longer pilot. Build before the clock starts, reach real volume early, and decide once against evidence rather than a deadline.
Is a 90-day pilot enough to prove an AI agent works?
Ninety days is enough time if the pilot spends it collecting evidence, and not enough if it spends it building. The number itself matters less than what fills it: a pilot that reaches real users or real data in week two and spends the rest of the period measuring can prove plenty in ninety days. A pilot that is still wiring up integrations in week eight cannot, no matter how the ninety days are counted.
So the honest answer is: ninety days is a reasonable budget for the evidence-gathering part of a pilot, but only if the building part is mostly done before the clock starts.
What should ninety days of an AI pilot actually contain?
| Phase | Roughly how long | What it should produce |
|---|---|---|
| Narrow build on real data | 1-2 weeks | A working version on your actual documents or systems, not a demo set |
| Shadow mode | 2-4 weeks | The agent runs alongside the current process without acting, so its output can be compared to what a person actually did |
| Supervised live use | 3-5 weeks | The agent acts, with a person approving or reviewing before anything that matters happens |
| Evidence review and go/no-go | 1 week | A decision made against thresholds agreed before the pilot started, not during it |
If most of the ninety days goes to the first phase instead of the last three, the pilot has not actually tested whether the agent works - it has tested whether it can be built, which is a different and much easier question.
What evidence should a 90-day pilot produce?
- A comparison against the current process, not a vacuum: accuracy, time and errors measured against what actually happens today, on the same cases.
- Performance on real volume, not a curated sample: the messy, awkward and edge cases that show up in ninety days of real use, not fifty hand-picked examples.
- A stable trend, not a lucky week: several weeks of consistent numbers matter more than one strong week early on.
- What happened at the edges: how often it escalated, how often an escalation was actually necessary, and what the misses looked like - see how to escalate an AI agent to a human for how that signal should be built in from the start.
- A cost figure at the volume you actually saw: not a projection from a demo, but what it would cost to run the pilot's actual volume for a full month.
When is ninety days not enough?
| Situation | Is 90 days usually enough? | Why |
|---|---|---|
| High-volume, short-cycle workflow (support tickets, routine data entry) | Yes | Enough real cases pass through in weeks to see a stable pattern |
| Low-volume or seasonal workflow (quarterly reporting, annual renewals) | No | Too few real cases occur in ninety days to trust the result |
| A workflow with a long feedback loop (did the recommendation turn out right months later) | No | The outcome that actually matters is not known yet when the pilot ends |
| A narrow, well-scoped task with a clear right answer | Yes | Accuracy and time are measurable quickly and reliably |
| A workflow still being redesigned alongside the pilot | No | The pilot is measuring a moving target, not the real workflow |
Where ninety days genuinely is not enough, the fix is usually not a longer pilot but a narrower one: shrink the scope to a sub-task with a shorter feedback loop and high enough volume to produce real evidence inside the window.
How do you design a pilot that makes ninety days enough?
- 01
Agree the go/no-go thresholds before day one
Accuracy, time saved, cost per case and acceptable error rate, written down and agreed with whoever makes the final call.
- 02
Build the narrow version before the pilot clock starts
Treat the first working version as pre-pilot setup, not part of the ninety days you are using to judge it.
- 03
Run in shadow mode first, briefly
Confirm the agent's output is close enough to be worth supervised live use, without risking anything on cases it has not yet proven itself on.
- 04
Get it to real volume as early as possible
A pilot that only reaches ten cases a week needs far more than ninety days to say anything reliable. If volume is naturally low, shrink the scope instead.
- 05
Review weekly, decide once
Track the numbers every week so surprises show up early, but make the actual go/no-go decision once, against the thresholds agreed at the start.
For the broader reasons pilots stall before reaching this point at all, see why AI pilots fail to reach production, and run the result through the AI agent go-live checklist before it goes further than the pilot.
Is ninety days enough to prove an AI agent works?
Yes, if the pilot spends most of that time collecting evidence at real volume rather than building, and if the workflow produces enough real cases in ninety days to show a stable pattern. It is not enough for low-volume, seasonal or long-feedback-loop workflows, where the fix is a narrower scope rather than a longer pilot. Agree the thresholds before the pilot starts, build the first version before the clock does, and make the go/no-go call once, against evidence rather than a deadline.
If you are designing a pilot now, our AI agent development team can help scope one that produces a real answer inside the window you have.
