Why our AI pilots reach production in four weeks

Most enterprise AI projects stop after the demo. The tool works, leadership is convinced, and six months later nothing has changed in operations. The problem is rarely technical.

A deliberately narrow scope

We never open a pilot on an entire process. We take one task, one team, one metric. The constraint forces the decisions a broad project lets you postpone indefinitely: what data goes in, who validates the output, what happens to the cases that aren't handled.

On a recent retail engagement, the initial scope fit in one sentence: answer product availability questions on WhatsApp, during opening hours, for a single store. Everything else came later.

A pilot that can't fail cleanly isn't a pilot, it's a demo.

Metrics before the first prompt

We measure the starting point before changing anything: average first response time, number of requests handled per day, error rate on inventory information. Without this baseline, the end-of-pilot discussion comes down to impressions.

• A volume metric, to know whether usage is real.
• A quality metric, to know whether the output is usable without rework.
• A workload metric, to know how much time the team actually got back.

The four weeks, week by week

• Week 1 — Scoping and baseline: interviews with the team involved, recording current metrics, choosing the task to cover and the cases explicitly excluded.
• Week 2 — First usable version: building the agent on real data, connecting the required sources, defining escalation rules to a human from the start.
• Week 3 — Supervised use: putting it in the hands of two or three users, systematically reviewing outputs, fixing the business rules that usage reveals.
• Week 4 — Switchover and handover: opening it to the whole team, comparing against the baseline metrics, documentation and handing maintenance over to the in-house team.

What remains open

Four weeks is enough to put one use case into production, not to transform an organization. What follows is governance work: who owns the agent, who adjusts it, how you decide to open a second one. That is where most companies need help, and it is rarely what they ask for at the start.