What the first eight weeks of an AI product look like

A founder with a proven method, a clickable mockup and money to build usually asks one question: how fast can this become real? On an AI sales-coaching product we built, the plan said eight weeks for the core. It was working in eight weeks, as planned. This is what those weeks held.

Weeks 1–3: the foundation, not the features

• Learn the method before writing code. The client’s scorecard, their coaching style, and their real sales calls.
• Rebuild the mockup as a real product. The mockup looked right but ran on nothing: no data, no scoring, no cloud.
• Clean the data. The client’s call transcripts were read, normalised and de-duplicated, so every score would rest on clean input.
• Connect the CRM read-only. The product could read the client’s CRM and never write to it.

Weeks 4–5: live on the client’s own accounts

From 1 May the product ran on the client’s own cloud and AI accounts, not ours. Building there from the start means nothing has to be migrated later, and the client can see the real thing, not a demo environment.

Slips reported the week they happen

In week 5, a database migration and the move to the cloud cost about a week. The client heard about it in that week’s report, not at the deadline. By week 6 it was back on schedule.

The same weeks showed a slowdown in the first cloud version: scoring a call took 120–180 seconds. It was found and fixed within the week. Scoring now takes 1.6 seconds per call, about 75 times faster.

Weeks 6–8: the core loop

• AI call scoring against the client’s own scorecard, every score with quoted evidence.
• Pipeline warnings: 16 rules across 3 agents watching the CRM.
• A weekly briefing for managers.

At eight weeks the core was complete. The client’s own sign-off reviews were still open, which is normal: engineering finishing is not the same as the client having used it in anger.