How to plan an AI assessment and pilot
A focused first engagement can combine a 1 to 2 week assessment with about four weeks of pilot development and evaluation. Here is the shape we use, phase by phase, and what shapes the schedule.
A focused first engagement can combine a 1 to 2 week assessment with about four weeks of pilot development and evaluation. Data access, review availability and integration complexity shape the final schedule.
In our experience, unclear scope is a common reason early AI projects stall: the ambition is a transformation, the budget is a pilot, the timeline is a quarter, and nobody wrote down what "working" would mean.
This is the shape we use for a first engagement. It is not the only possible shape, but it is one that reliably produces a decision. The phases below are the same ones on our approach page, which also has a worked example and a sample evaluation report.
| Phase | Duration | Ends with |
|---|---|---|
| Fit call | 30 minutes, free | A decision on whether to assess |
| Assessment | 1 to 2 weeks, fixed fee | Roadmap, evaluation set, pilot proposal |
| Pilot, including evaluation | About 4 weeks, quoted | Evaluation report and a recommendation |
| Rollout | Scoped separately | Named owners, monitoring, support agreement |
The total is usually 6 to 8 weeks. It depends on scheduling, access readiness and integration complexity, and it does not include rollout.
Week 0: the fit call
Thirty minutes, free. We want to hear one workflow described by the person who lives with it. We want to know what a good result would be worth. And we want to say, early, if the honest answer is a simpler tool or a process change rather than AI.
Weeks 1 to 2: the assessment
The assessment is a fixed-fee engagement of 1 to 2 weeks. It produces:
- A map of the workflow with its friction points, drawn with the people who do the work.
- A prioritised shortlist of where AI could help, placed by value and feasibility.
- A build-versus-buy recommendation. A custom build is one possible answer, not the default.
- An evaluation set of real cases with agreed correct answers.
- The approval rules: which actions wait for a person during the pilot.
- A pilot scope with success measures, indicative effort and itemised running costs.
The assessment ends with a document you can act on with us or without us.
Weeks 3 to 6: the pilot, evaluation included
The pilot takes about 4 weeks and is bounded on purpose. One user group. One task. The integrations that task actually needs and no others. A review interface for anything uncertain.
The rhythm is a weekly review call and a written update every week. Progress is visible, and decisions are made by named people on both sides rather than waiting for a steering committee.
In the final week the pilot is scored against the evaluation set and, where possible, against one live week of real work. The report shows accuracy, error types, review rate and cost per run together.
After the evaluation: rollout, scoped separately
If the evidence supports it, rollout is scoped as its own piece of work: named owners, monitoring of exceptions and costs, an operating guide, which automatic actions are now agreed, and support hours agreed in writing. If the evidence does not support it, the report says so and why.
Either way, ownership, licences and handover were agreed before development started: bespoke code written for you is yours under the engagement agreement, third-party components keep their own licences, and infrastructure and accounts are arranged in scope. You keep the workflow map, the evaluation set and the findings.
Three ways this goes wrong
The scope grows in week three. Someone sees the pilot working on invoices and asks whether it could also do contracts. The answer is yes, in the next iteration. Adding it now means the evaluation set no longer matches the build.
Nobody can review during the pilot. A pilot needs someone on the client side who can look at results weekly and supply real examples. If that person does not exist, start with the assessment only and revisit.
The pilot is judged on a demo. A demo shows the cases that work. The evaluation set is there so the pilot is judged on the cases that were chosen before anyone knew what would work.
What it costs
We do not publish a price list, because scope drives cost. We do publish the structure: a free fit call; a fixed-fee assessment agreed before you commit; the pilot quoted after the assessment against a written scope; model usage, hosting and third-party services itemised at cost. No number is a surprise, and the running cost is visible before rollout.
If you have one workflow in mind, that is enough to start.
Have a similar workflow?
Discuss what a focused first project could involve. Tell us which article you read.