Contents
Written by
There has not been a more exciting time to work in financial risk than right now.
Agentic AI, the kind that can read, reason, and act on a goal across a sequence of steps, is starting to do real work inside real risk functions. Teams across credit, procurement, and compliance are putting it to good use. Frankly, the pace of adoption I am seeing in the industry is faster than anything I have seen in my 30 years plus career.
But there is a conversation we need to have, and it is not the one most people are having.
The discussion right now is almost entirely about models. Which one. How big. How fast. Whether it can reason. Whether it can act autonomously. These are interesting questions and valid ones to ask, but for financial risk specifically, they are not the place I would start. The right question is what the agent is reading.
Strip the buzzwords away and the picture is simple. An agent reads a goal you give it, plans the steps to get there, queries the right tools and data, and either takes an action or recommends one.
Run a portfolio review. Check 200 suppliers against three watchlists. Pull every director change on your top exposures from the last 90 days. Build a one-page risk view on a potential acquisition. The agent does the legwork. The analyst makes the call.
When an agent gets a credit decision wrong, the instinct is to blame the model. In our experience, failures are more often rooted in the data layer than the model itself.
Give a brilliant model stale, shallow, or unstructured data and you get a brilliantly wrong answer, fast, with full confidence. In risk that is dangerous. A wrong call on a counterparty, a missed flag on a director, a supplier onboarded with quiet distress signals already showing in the accounts. These are not abstract problems. They show up on the P&L statement.
A large language model, on its own, can talk fluently about credit risk, but cannot calculate it.
An LLM does not know a company’s H-Score®. Nor can it produce a Probability of Distress® on demand. It cannot tell you that a director’s last three companies all failed within 18 months of incorporation. And it will not pick up a quietly filed change of accounting reference date, a group structure shift, or a fraud pattern building across multiple risk zones.
We have spent decades building exactly those capabilities. H-Score®, TextScore®, PoD® and Vigilance™ are not language model outputs. They are purpose-built risk models, running on the deepest UK and Ireland company dataset available anywhere, refreshed continuously.
When an agent gets access to that scoring layer, the change is not subtle. The agent stops summarising and starts assessing. It stops describing a company and starts forming a view on it. That is the line between a basic MCP and an agent built to surface real risk.
MCP (Model Context Protocol) is an open standard, introduced by Anthropic in late 2024, that defines how AI applications connect to external tools and data. Expose your service once through an MCP server, and any compatible AI client can use it, no custom integration required. It’s the USB-C for AI: one standard connection replacing a tangle of one-off adapters, and it’s what lets Company Watch plug directly into any AI assistant.
The most common question I get from clients is whether AI is going to take their analysts’ jobs.
It is not in the way people fear. The best analysts become dramatically more productive, and their judgment matters more, not less.
What it is going to do is change what one good analyst can get through in a week. Take a strong credit analyst, give them an agent built on Company Watch data, and let them work the way they want to. Our research suggests they can get through materially more data than they would without the agent, often by an order of magnitude. Triage thousands of counterparties overnight, walk into a prioritised stack of the dozen that genuinely need a human eye, and sign off the rest with confidence because the underlying scoring is sound.
That kind of leverage is huge, and we’ve seen first-hand what a difference it can make in the financial risk world.
Most of the AI tools risk teams are starting to lean on, copilots, in-house assistants, agentic workflows built on top of GPT, Claude, Gemini, and the rest, are only ever going to be as good as the data they can see. Right now, on UK and Ireland company intelligence, most of them are flying blind. Ask one of them about a small contractor in Birmingham or a recently restructured group in Dublin and the answer you get back is, charitably, a guess.
Your analyst types a question into their copilot, “is this supplier safe to onboard,” and the answer comes back in seconds, with the same evidence base they would get logging into our platform manually, embedded directly in the tool they are already using. No tab switching or copy and paste. And no more “the AI does not really know about UK companies.”
Plug it in once and every agent, every copilot, every LLM in your stack starts speaking Company Watch.
We are launching it soon. If you are exploring agentic workflows in financial risk, I’d welcome a conversation or a practical comparison of what you’re testing versus what we’ve built. My view is simple: the firms that connect their AI to trustworthy, decision-grade data early will move fastest over the next two years.
If you take one thing from this piece, take this.
Before you let an AI agent touch a real risk decision in your business, ask four questions.
The firms that get this right will be the ones that treat agentic AI as a capability layered on top of trustworthy intelligence, not as a substitute for it. The models will keep getting better. The data discipline is the part that has to be earned.
We’ve spent decades developing our data. Now we are making it usable everywhere your team works. If it is helpful, we can show you exactly how the scoring layer plugs into your existing AI tooling.
