If you read our Trustpilot reviews, you’ll see brokers thanking specific underwriters by name. That doesn’t happen at a high street bank, where the underwriters are hidden behind a system.

There’s a reason for this: at Gen H, there is a human behind every decision, and brokers have direct contact with the underwriter on their case. We don’t hide behind systems, we don’t hide behind automation, and we won’t hide behind AI.

But that isn’t to say AI isn’t changing things for us.

AI is the newest tool in our kit, and we use it for the same reasons as everyone else. We’re looking for efficiency, more powerful data analysis, and insights that can supercharge how we do business. But more than that, we’re looking for ways to give our team more time to do the work a human is best placed to do. This resonates with our brokers and customers.

This has proven especially powerful in underwriting.

Plenty of the work around a mortgage application is already automated, as it should be: we instruct the valuation, run KYC checks, and compare the credit file against policy automatically. Human judgement doesn’t add much value there.

Underwriting at Gen H is what’s left: the real risk-based call. It’s seeing the missed payment three years ago for what it actually was: a one-off, not a structural credit problem. It’s the customer who’s temporarily supporting their parents, whose affordability needs a closer read than the headline numbers. It’s the contractor whose income is irregular on paper but, read properly, is as stable as the PAYE on our books. That’s the part we keep manual.

But this very high touch underwriting process has trade-offs.

Namely, it depends on the application arriving fully packaged. A payslip could be missing; the underwriter raises a query, the broker comes back hours or days later, and the file gets picked up cold. Packaging is a known administrative burden across the industry, and a clear place for technology like AI to earn its keep.

So we built an AI agent to help us with packaging checks. It reviews each case as it lands, compares the documents against what the case needs, and suggests questions or clarifications the underwriter might want to raise. The underwriter accepts those suggestions, edits them, or deletes them, and sends them back to the broker. The agent has been in production with our underwriters for months and helps save time on every case.

The intention is to make the agent broker-facing, too – it will, eventually, proactively raise queries directly with brokers via email before the underwriter picks it up. We meant to launch it a few months, but we haven’t yet – and that’s deliberate.

The agent is trained on our lending policy. But we’ve learned that our underwriters are routinely overriding its suggestions because they’re applying judgement the agent doesn’t have. They accept a bank statement that’s a little out of date because they’ve seen enough of the customer’s pattern to be comfortable. Or they have enough confidence in the income that it’s not strictly necessary to get the latest payslip.

The agent doesn’t understand this yet. If we shipped it as it currently exists, we would actively make the broker experience worse (though it would sound cool on LinkedIn).

So this is what we’ve learned: AI has the potential to fundamentally change how our team works. But it’s worth being precise about what we’re protecting. Our lending policy provides a baseline of consistency, but we give underwriters deliberate latitude to exercise judgement and common sense.

That non-determinism is hard to automate. It’s also, like the manual approach itself, a feature, not a bug. AI, itself, is non-deterministic – you can’t predict what it would do. But it can be trained to behave more, often than not, in a way that you consider to be net-good. That’s our target before shipping.

We know what good looks like – you can enable human judgement and be way more efficient with AI. For example, we built a dashboard in a few hours with Claude Code for our head of underwriting that shows every DIP that has referred or failed. About twice a day, he picks up the phone to a broker, out of the blue, to tell them that although their DIP has failed, he thinks he can make an exception on a piece of criteria. He has the time for this because work that used to fill his calendar now happens in the background before it lands on his desk.

The incredible thing is that dashboard didn’t take a quarter of planning; it took an afternoon. We actually built it during an internal AI hackathon – a twice-yearly event where our engineers take a day to hack away at whatever business problem they’d like.

That’s the other thing AI has changed for us: how fast anyone on our team can build the small tools we need. A month ago I built an email that goes out to brokers automatically when their client’s product switch window opens. It took two days and was reviewed by real engineers like any other piece of software here. The difference is I don’t know how to code.

With AI, you can just do things. Fast. The potential to improve how you do business is limitless. The real challenge – for us, and everyone else – is building tools that amplify our humanity rather than efface it.

Luke Calton is head of product at Gen H