Can AI help a finance broker match deals to lenders?

This is one of the better fits we see, and commercial finance is a sector we already build in. It works because the problem is genuinely structured underneath, even though it does not feel that way day to day.

The problem in plain terms

A broker with access to sixty lenders is realistically placing with eight. Not through laziness, but because holding sixty sets of criteria in your head, each with its own view on loan to value, sector, term, adverse credit, minimum size and property type, and each changing without notice, is not something a person can do. So the case goes to the lenders you know well, and the better deal at the lender you had forgotten never happens.

What a matching engine does

Criteria live in one place, structured, with a date against every change. A case is entered once. The system checks it against every lender and returns three groups: fits, fits with a caveat, and does not fit, each with the specific criterion that decided it.

The reasoning matters as much as the result. A broker will not trust a bare list of names, and quite rightly. They will trust “out on minimum term, everything else passes”, because that is checkable, and it is the sort of thing you can repeat to a client.

Where AI belongs, and where it does not

The matching itself should be deterministic. It is comparison against known thresholds, it has to give the same answer twice, and you need to be able to explain any decline. Building that on a language model would be a worse product wearing a better badge.

AI earns its place at the two ends:

• Getting criteria in. Lenders publish updates as PDFs, emails and product sheets. Reading those and proposing structured changes for a human to approve is exactly what models are good at, and it is the task that otherwise never gets done.
• Getting the case in. A first enquiry arrives as a phone note or a forwarded email. Extracting the parameters from that, so the broker is checking rather than typing, removes real friction at the point where cases go cold.

The part that decides whether it works

Maintenance. Criteria change constantly, and a matching engine three months out of date is worse than no engine at all, because people will believe it. Whoever builds this has to answer how updates get in, who approves them, and how a broker sees when each criterion was last confirmed. Design that in at the start or the tool gets quietly abandoned within a quarter.

Where the boundary sits

The system is a research tool for the broker. It narrows a field and shows its working. It does not decide, and it does not speak to the client. Keeping that line clear is the difference between a useful internal system and a regulatory problem you did not need.