The win

Our customers are taking AI deep inside their core lending process without rebuilding their stack. The pattern across engagements is consistent: pick the most painful manual underwriting workflow, automate it, demonstrate business value, and turn that first win into a foundation to spread AI across the org.

What that looks like in practice:

• Intake workflows that used to eat days of headcount collapse to underwriters spot-checking exceptions.
• Validations and checks that used to depend on individual judgment now run methodically on every deal, with an audit trail.
• Historical lending decisions get audited into structured data alongside the new flow, which becomes the foundation from which we support customers owning their data going forward.

Critically for the future, every customer starts their AI journey with something more valuable than just the workflow we automated: a data layer to train their own models on, with decision traces, evals, and validation rules that fall out as byproducts of using our AI system. The customer owns all of it, and we support them as their AI maturity increases.

Why this is happening

The lenders we are working with have already made their platform bet. They have a CRM or core banking system, a self-managed cloud, and years of tech stack investment. They are not in the market for another tool for their business users to get lost in.

They want to own the data, the system of record (and the decision traces).

AI is forcing teams to consider where their data flows and how decisions are made. Handing that to a platform vendor means handing over the most strategic asset in the business, and to a stack they no longer control.

A year ago, the conversation was about end-to-end underwriting platforms, each platform with its own agent framework, workflow engine, and lending abstractions.

Three things have become clear since:

1. Learning someone else’s framework is a permanent cognitive load.
2. Sending data to a vendor’s APIs is a perimeter that the security team has to defend.
3. Waiting on someone else’s roadmap means living with the median of all their customers, often designed for a different geography or scale.

In financial services, mistakes haunt you for years. The robust way to put a credit process into production is not to encode it inside someone else’s hosted agent loop.

We have also watched lending tech follow high-growth sectors into new markets and quietly ignore early customers. All these forces are why pure SaaS is ending for anything that touches the underwriting workflow. Lenders want the AI running where the data already lives, under their own cloud commercial agreements, with their own model choices.

As frontier model pricing drifts out of viability for high-volume internal workflows, smaller specialised models on their own infrastructure are how the unit economics work.

What we embed in customer stacks

• A library of pre-built, lending-specific AI primitives. Document extraction, validation, classification, fraud detection, chat, and audit trails. We make them lending-shaped to deliver outcomes in weeks, low-level enough that we can work with the team to compose them for the specifics of each workflow.
• A decision layer that the customer owns. Every extraction, classification, and validation runs with decision traces, evals, and validation rules attached. Over time, those become a proprietary dataset: a structured record of how the business actually decides, ready as audit material, training data, or to train the customer’s own specialised low-cost models.
• Runs inside their own cloud environment. No data egress. No new vendors on the security review.
• Designed for embedding. AI-assisted integration means most of the work is configuration, not glue code. We extend the team where it matters and are on the hook for delivering in weeks, not months.
• A clean API contract between our layer and their core system. They can interrogate exactly what goes in, what comes out, and replace any of it if they want to.

The two halves of the AI market

The market is dividing cleanly.

General-purpose AI is going to the labs. You can prototype a solution in an afternoon. The labs will win the generic capability fight every time. Ad hoc tasks make sense to do with Claude and Codex.

The other half is your core operational process: the cost of doing your business, the people you employ on volume tasks, the bottlenecks and the drivers of margin. None of that is something a generalist model can own for you. “Just use Claude” is its own form of key-person risk, where the key person is Claude, and the model provider gets to decide how many of your customers they want this quarter. When their price subsidy ends, they pressure your margins.

Decision traces are the durable advantage and insurance policy. These are not chats with models; they are the full operational record of what the model did to make every response.

We capture these from day one. They become a structured, owned data set that lets you analyse the process at scale, train your own specialised models on it, and survive whatever the frontier labs decide to change next.

If you aren’t storing your AI decision traces, they aren’t your decisions.

If the following rings a bell

• You have built on your core, with self-managed cloud and proper security, and want to compound that investment.
• You have manual document workflows costing real headcount and creating real risk.
• You want AI on your data, you don’t want to ship your data anywhere, and you want to own your AI model destiny.
• You have looked at the full-stack platforms and walked away from the technical debt and implementation-and-security nightmare.

We are having this conversation every week with business, commercial, and mortgage lenders across Europe, the US, and elsewhere. If you have made the same bet, we would love to speak.