Lenders use AI agents today for the work around the credit decision. An agent reads the documents, pulls out the facts, checks them against the lender's rules, asks the applicant for what is missing and writes up the file. A person on the credit team still makes the decision. That split is the safest place to start.
What an agent does
From what I know, older automation follows fixed steps on fixed layouts. It breaks when a bank statement looks different or a document arrives in the wrong order. An AI agent reads a document it has not seen before and works out what to do next. If a figure is missing, it knows to ask for it.
In lending, the jobs look like this:
- Read and sort what the applicant or broker sends in.
- Pull out the facts, such as turnover, the amount asked for and the owners.
- Check the file for gaps against your list of requirements.
- Compare facts with each other and with registers, such as Companies House.
- Ask the applicant or broker for what is missing.
- Write a draft of the file for the credit team to review.
None of these jobs decides anything. They get the file ready, and that is where much of the time goes.
How credit and insurance underwriting differ
The two jobs sound alike but ask different questions.
An insurance underwriter prices the risk of something that may happen later. A claims handler checks if a claim falls inside the policy. A credit underwriter asks if the borrower can repay the loan, and what the lender can take if they cannot.
Both start from the same thing, which is a complete file. At DAS, a legal expenses insurer, I built an AI workflow that checked whether claims were covered. What I saw there was how much time went into working out what was missing and going back to customers to chase it. I think the same is true in lending. A file that arrives half complete needs someone to find the gaps and write the email.
What stays with people
The decision stays with people. That covers whether to lend, on what terms and at what price. It also covers exceptions, such as a file that fails one check but is strong in other ways. And it covers any case where a person is judged, because the rules on that are stricter. I wrote about those in is AI allowed in business lending decisions.
A failed check should stay visible until someone has looked at it. A system that hides a failure to make the file look neat is worse than no system.
What to ask a vendor
- Where does each fact come from, and can I click through to the source?
- What happens when a model call fails? The answer should be that it stops and says so.
- Who sees a failed check, and can it be hidden?
- Does it score or decide? If it scores a person, what are the rules for that?
- Can I write my own checks in plain words?
- Where is the data kept, and for how long?
A worked example
This is an example with made-up details. It is not a real lender or applicant.
A broker emails an application at 16:40. It has a form, three bank statements and no ID for the second director. Your policy asks for six months of statements.
The agent reads the form and the statements. It sees the gap in the months and the missing ID, and it writes to the broker the same evening, listing both. It also checks the company on the register. When the broker replies, it reads the new papers.
The next morning, the credit officer opens a file with the facts, each linked to its document. It shows two items that were open and are now closed, and one check that failed. The officer reads the failed check and decides what to do. Without an agent, someone reads the email in the morning and writes to the broker. The reply comes a day or two later.
Where to start
Pick one product and write down what a complete file needs. I wrote about how to do that in write your intake checklist before you buy software. An agent can only check against a list that exists, so the list comes first.
Then count the follow-up emails on your last ten files. That count is your starting point. Pick the part of the file that goes wrong most often, such as missing bank statements, and run the agent next to your team for a few weeks. Compare what it found with what your team found. Your team keeps the decision the whole time.
What can and cannot be automated
Reading, sorting, finding facts, spotting gaps and chasing papers can be automated. So can register lookups and the first draft of a write-up. A model can misread a blurry scan, or invent a value that is not there. So every fact needs a source. A missing value should stay marked as missing. And ordinary code should do the sums and the format checks.
The credit decision cannot be automated, and neither can your risk appetite or the exceptions. The same goes for the relationship with a borrower you have known for years.
Klarefi is built for the first group. It does the intake work, so the file is complete when your credit team opens it. It does not score or decide.
See how Klarefi works and the document checklist.