Every quote starts the same way. An email arrives with a PDF drawing, sometimes a STEP file, and the quantities written out in free text. Someone opens it, works the geometry in Excel, and re-keys the result into the ERP. It takes hours, and the estimate is only ever as good as the person was on the day.
So the brief looked obvious: point AI at the drawing and let it produce the quote.
We built it three times
We built that. Then we built it again, and a third time. The early versions leaned on the model to interpret the drawing: read the shapes, infer the dimensions, estimate the work. In a demo it was impressive. In production it was a coin toss. A model that is confident and occasionally wrong is more dangerous than a slow human who knows when to check twice.
There is a reason the demo flattered us. A demo runs on the example you picked. Production runs on the drawing that turns up on a Tuesday with a layer mislabelled and half the spec in a margin note. Interpretation breaks down exactly where nobody is watching.
The answer was already in the file
What turned the project around was noticing where the answer already lived. The CAD and STEP files were not images to be read. They were exact geometry: lengths, areas, counts, precise to the millimetre and already sitting in the data. The model had been guessing at numbers the file could simply tell us.
So we stopped asking it to guess. We read the geometry deterministically, straight from the CAD data, and kept the model only for the genuinely ambiguous parts: the free-text notes, the details a drawing leaves open. The estimate fell from 240-270 hours to roughly 40.
Where the data is already exact, deterministic code wins every time.
Know the half that shouldn’t use AI
The lesson travels well beyond quoting. AI earns its place where the input is messy: language, images, unstructured documents, the things rules cannot parse. Where the data is already structured and exact, deterministic code wins. It is faster, it is cheaper, and it can be audited, because you can point at the line that produced the number.
The hard skill in an AI project is not using the model. It is knowing the half of the problem that shouldn’t.