Thousands of bid files, priced by hand

What they had

A telecom parts dealer receiving bid files faster than they could be priced. Thousands of them, and pricing that depended on judgment held by a single in house expert. Throughput was capped by how fast one person could read, and the logic holding it together existed nowhere but in their head.

What we did

A paid proof of concept first, over 27 development days, to establish whether the hard parts were possible: parsing files of inconsistent quality, matching parts against supplier data that rarely agrees with itself, reasoning about price across multiple agents, pulling live market intelligence, and holding on to what the system learned between runs. It produced 37 commits, 24 completed stories, and a measured build velocity.

That velocity sized the production build at 97 days across seven areas of work. One senior architect was named before any price appeared, and their pricing expert sat in on every decision, because the goal was to encode the judgment rather than replace it with something generic. After launch a fractional architect stayed available for changes inside the original scope.

They were being asked to fund a build that depended on something nobody had proved yet, which was whether a system could reliably reproduce a specialist's reasoning about price. Committing to a full engagement on that assumption would have meant carrying the technical risk and the budget risk at the same time, on a process the business was actively running on.

Starting with a paid proof of concept capped that exposure at one month. If the hard parts had failed, they would have found out for the cost of four weeks rather than the cost of a year. Because they did not fail, the same month produced something more useful than confidence: a real rate of progress, which is what made the 97 day figure an arithmetic result rather than an experienced guess.

The three payment structures existed for the same reason. The cost was not the obstacle. Where the cost landed was. Offering one scope with different splits between upfront and monthly let them fit the engagement to how their business is funded, instead of negotiating the work down until the number fit.

And their pricing expert stayed in the room throughout because the asset being built was his judgment. A model trained without him would have produced generic pricing, which is exactly what their business does not sell.