From Pilot to Production in 4 Weeks: How We Run an AI Pilot

"We tried AI, but it didn't go anywhere." We hear this often. Usually the technology wasn't the problem — the pilot was designed as a demo instead of as the first version of a real system.

Here's how we structure pilots so they have a real chance of reaching production.

Week 0: Discovery & audit

Before building anything, we spend a few focused sessions with the people who do the work.

• What is the process today, step by step?

• Where is time lost, and where do errors happen?

• What data exists, and where does it live?

• What would "success" look like in numbers?

The output is one use case, a baseline measurement, and a clear success metric.

Week 1: Solution architecture

We design the smallest system that can solve the problem with real data — how it connects to existing tools, who uses it, what happens when the AI is unsure, and what security and access rules apply.

Weeks 2–3: Rapid pilot

We build and run it with real users and real data — not a sandbox demo. Feedback is collected daily and the system is adjusted as we go. This is where most of the learning happens.

Week 4: Decision and deployment plan

We compare the results with the baseline:

• Go — move to full production rollout with monitoring

• Adjust — change scope or approach based on what we learned

• Stop — sometimes the honest answer is that the use case isn't worth it yet. Learning that in four weeks is far cheaper than learning it in a year.

After go-live: Optimisation

Production systems need care: performance monitoring, handling new edge cases, retraining where needed, and expanding to the next use case once the first one is stable.

What makes the difference

• One use case, not five

• Real data and real users from week two

• A number to beat, agreed up front

• A clear decision at the end, not an open-ended "phase 2"

If you have a process that feels like it should be automated but don't know where to start, a short discovery session is the easiest first step.