How we ship an AI prototype in six weeks

Every AI project we take on starts with the same promise: a working prototype in your hands within six weeks. Not a deck, not a Figma flow. Software you can open, break, and argue with.

People ask why not two. The honest answer is that two weeks buys you a demo, and a demo is a liability. It works once, on the happy path, in front of an audience, and then it sets an expectation nobody can keep. Six weeks is the shortest span in which we can build something and find out where it falls over.

Where the weeks go

• Week 1, Discover. We map the actual workflow: who does what, with which data, and where the pain is sharpest. We pick the one use case to build first and write down what "good" means as a number.
• Weeks 2 to 3, Prototype. We build the thin slice end to end. Real data, real integration, rough UI. The goal is a thing that runs, not a thing that impresses.
• Week 4, Harden. We put it under load and watch it fail. Bad inputs, missing context, the edge cases your team hits every Tuesday. We write evaluations so failure is visible, not anecdotal.
• Weeks 5 to 6, Hand over. We fix what the evaluations surfaced, tighten the interface, and hand you something you could put in front of a colleague.

The rule that makes it work

If we can't measure whether it's working, we don't build it yet.

Before a line of agent code, we agree on an evaluation set of a few dozen real cases with known-good answers. It's unglamorous and it's the difference between "the AI feels smart" and "the AI is right 94% of the time and here's the 6% we caught."

What you get at the end

A prototype that runs on your data, a short doc on where it's strong and where it isn't, and a straight recommendation: kill it, keep iterating, or harden it for production. Sometimes the right call after six weeks is stop, and knowing that for the cost of six weeks is a good outcome, not a failed one.

Six weeks. Something real. Then a decision made on evidence instead of vibes.