Quoting is the most expensive thing your engineers do badly

In most engineering and fabrication businesses, quoting is done by the people you can least afford to have doing it. It lands on a senior engineer or the owner, it is interrupt-driven, and it competes with work that is already late.

The result is a process that is slow, inconsistent between estimators, and almost entirely undocumented. It is also, for exactly those reasons, where AI most often pays for itself first.

What the tools are genuinely good at

Reading the enquiry. Customer enquiries arrive as email threads with drawings attached, half a specification, and a deadline buried in paragraph four. Extracting structured requirements from unstructured mess is precisely what language models do well.

Finding the comparable job. Most estimators price by memory: "we did something like this for Renishaw two years ago." That recall is the most valuable thing in the process and the least reliable. A system that searches previous quotes for genuinely similar jobs turns one person's memory into something the business owns.

Drafting the document. Once the numbers are decided, assembling the quote — terms, lead times, exclusions, the right boilerplate for that customer — is mechanical and fully automatable.

Flagging the unusual. A tolerance tighter than anything you have made, a material you do not stock, a quantity outside your normal range. Catching those before they are priced is worth more than the time saved.

What they are not good at

Deciding the price. Pricing reflects how badly you want the work, what the customer paid last time, what capacity looks like in six weeks and whether you trust their forecast. None of that is in your data.

Judging manufacturability. A model can read a drawing. It cannot tell you that the feature will chatter on your machine with your tooling.

Knowing what your previous quote left out. If your historical data contains quotes that lost money, a system trained on it will confidently reproduce the losses.

The structure that works

Draft, then review. The system produces a complete draft quote with its reasoning visible — which comparable jobs it used, which assumptions it made — and a human approves, adjusts or rejects it. Never send automatically.

Two reasons. First, a wrong quote that goes out unchecked costs more than the automation saves, and it can be contractually binding. Second, the review step is where your estimators teach the system, and where you discover what it gets wrong.

The data problem you will hit

The hard part is not the AI. It is that your quote history sits in a mixture of spreadsheets, PDFs, an accounting system and a folder structure that made sense to someone who left in 2019.

Before anyone builds anything, answer: can you retrieve every quote from the last three years, with what was quoted, what it cost to make and whether you won it? If you cannot connect those three facts, the first project is a data project, not an AI one. That is not a reason to stop — it is a reason to sequence it properly.

What good looks like after six months

Quote turnaround measured in hours rather than days. Estimators spending their time on the judgement calls rather than the assembly. Consistency between estimators that you can actually see. And a record of why each job was priced the way it was, which is the thing that lets you work out afterwards where you are making money.

That last point is usually the real prize. Most firms know their overall margin. Far fewer know which kinds of job earn it.