How long should an AI pilot take? Realistic timelines from discovery to go-live
Almost every AI vendor pitch includes some version of “we can have this live in a couple of weeks.” So it’s worth asking plainly: how long should an AI pilot actually take? Gartner’s 2024 AI Mandates for the Enterprise survey found organisations spend an average of 33.6 weeks going from idea to production on a non-generative AI project, and 29.3 weeks on a generative AI one - that’s seven to eight months, not weeks. Of that, roughly seven to eight weeks goes on idea vetting alone, before any build work starts. A single, well-scoped pilot on one workflow should move faster than that enterprise-wide average, but “faster” still means months, not days.
How long should an AI pilot take for one workflow?
The Gartner figure covers the full spread of enterprise AI projects, including large rollouts across multiple teams. For a single pilot on one workflow - which is what most small and mid-sized businesses should be running first - a realistic timeline looks like this:
- Discovery (1-3 weeks). Picking the one workflow, naming an owner, and capturing the baseline numbers (turnaround time, error rate, cost per job) before anything is switched on.
- Setup and configuration (2-4 weeks). Connecting the tool to the real data it needs to work with, not a demo dataset.
- The pilot run (8-12 weeks). The actual trial, measured against the baseline from day one, not reconstructed afterwards.
- Review and decision (1-2 weeks). Comparing the result against a threshold agreed before the pilot started, and deciding to scale, adjust, or stop.
Add it up and a single pilot, done properly, runs around three to four months end to end. That’s longer than a weekend build, and considerably shorter than the seven-to-eight-month enterprise average, because it’s one workflow with one owner rather than a firm-wide programme trying to move several things at once.
Why the “weeks” claims don’t hold up
The gap between a two-week promise and a three-to-four-month reality isn’t a sales trick so much as a scope mismatch. A demo can be built in a couple of weeks. A pilot that produces a number a business owner or a board can trust takes longer, because the baseline, the real data connections, and the measurement period all have to happen before anyone can say whether it actually worked. MIT’s Project NANDA found 95% of generative AI pilots show no measurable return on the P&L - and the pattern behind most of that 95% isn’t a bad tool, it’s a pilot that skipped the slow parts to hit a fast deadline. Separate industry tracking found the share of businesses abandoning most of their AI programs jumped from 17% to 42% in a single year, which tracks with pilots that were rushed into a result nobody can defend.
On the other end, a pilot can also run too long. If there’s no review date and no pre-agreed threshold, a trial can drift on for six months or more without anyone deciding whether it’s actually working - which is just as unhelpful as rushing it. The fix in both directions is the same: agree the timeline and the success criteria before the pilot starts, not while you’re already running it.
What actually slows a pilot down
The stages above assume the groundwork is already in reasonable shape. In practice, two things routinely blow out the timeline before the pilot even begins:
- Messy or scattered data. If customer or job information lives in three different systems, the setup stage stretches because the tool has nothing reliable to connect to yet. Discovery turns into a data clean-up project instead of a two-week scoping exercise.
- No named owner and no governance answer ready. A pilot without a single accountable person tends to stall at every decision point. And if nobody can say where the data goes, which vendor sees it, and how that sits against the Privacy Act, that conversation happens mid-pilot instead of before it - adding weeks you didn’t budget for.
Both of these show up before day one of the actual pilot, which is exactly why they’re worth checking first.
Where to start
If you don’t know whether your data and systems are in good enough shape to hit the three-to-four-month timeline above, guessing is slower than checking. Our complimentary AI-Readiness Data Check is twelve questions, about ten minutes, and gives you a plain picture of where your foundations stand before you commit to a start date.