Pilot
Service sheet 2/6 · Medium · 3–8 weeksProve one process on your real data in weeks, with a stop-or-scale decision at the end.
- A working system on one process, one channel and a limited set of users, running on your real data for a fixed period.
- Measured results against a baseline you agreed before the start (response time, extraction accuracy, leads captured, hours saved).
- A written stop-or-scale recommendation with the cost of scaling.
- Pilot charter (one page: outcome, measure, baseline, users, duration, stop rules)
- Working pilot on your data for 2–6 weeks
- Results memo with the numbers and the failure cases
- Scale plan or a stop note
- Nobody can predict how the AI will behave on your documents, accents or customers — which is nearly always.
- You need evidence before a board, a bank or a partner will fund the full build.
- Voice, vision or document extraction are involved: these depend on your recordings, cameras and scans.
| Effort band | Medium · 3–8 weeks |
|---|---|
| Your side | One owner; 5–20 users; 200+ real records or 1–2 weeks of live traffic |
| Success test | Agreed before start; measured by you, not the vendor |
| Pricing fit | Fixed scope with a milestone at go-live |
| Zimbabwe constraint | Run the pilot through a load-shedding week and a month-end; both change behaviour |
- A "pilot" with no end date and no measure — that is a build with a discount.
- The pilot runs on the vendor's demo data, not yours.
- The pilot quietly needs the full integration to work, so it never runs live.
Fixed scope · Milestone
See pricing models for how each behaves when volume changes.
A pilot answers one question: does this work on our data, with our customers, in our conditions? It does not answer “is AI good for us”. Keep it to one process. Keep it to one channel (usually WhatsApp). Give it a stop rule (“if extraction accuracy on 200 invoices is under 90% at week four, we stop”).
Sizing a pilot in Zimbabwe. Two conditions make pilots here different from the vendor’s last one abroad. First, month-end: invoice volume, EcoCash settlements and salary queries all spike, and a system tuned on a quiet fortnight fails on the 28th. Second, power and connectivity: run at least one week through a known outage window at your premises and watch what queues and what is lost. If the vendor cannot show you a queue, the pilot has taught you something already.
Judging the result. Use the baseline you wrote in the charter. A pilot that “went well” without numbers is a demo. Ask for the failure cases as a list — those are the scope of the build.