A falling insolvency figure is not the same as falling risk

There is a particular kind of comfort that arrives with a negative statistic going down, and the recent company insolvency figures for England and Wales gave the market one. The insolvency rate over the twelve months to June 2026 was 50.5 per 10,000 companies on the effective register, down from 52.4 a year earlier. That is one in 198 companies. Read the headline on its own and you would conclude that the worst is behind us and that credit conditions are easing.

I would be careful. In my thirty years plus working with distressed businesses, the most dangerous moment is rarely the one where the topline number is climbing and everyone is paying attention. It is the one where the topline is falling and people have stopped looking underneath it.

What the headline leaves out

Look one layer down and the same set of figures tells a different story. I would work from the measure that statisticians themselves trust most i.e. the twelve-month rolling rate, because it strips out the noise of any single month and is far more comparable over time than a raw monthly count. On that measure, you’ll notice that the overall rate is easing. 

However, just as importantly, the administration rate is doing the opposite. It climbed from 3.3 to 4.0 per 10,000 companies over the year, while creditors’ voluntary liquidations, the bulk of the total, fell from 40.9 to 38.4. Administrations are not the failures of small companies quietly winding up. They are larger, more interconnected businesses, the ones whose collapse pulls suppliers and lenders down with them.

That is the part that should hold a credit team’s attention. The overall rate is easing while the weight of failure is concentrating. A voluntary liquidation of a dormant company and the administration of a property group with dozens of entities attached to it both count as a single insolvency in the rate. They are not the same risk, and treating them as interchangeable is how a portfolio ends up exposed to exactly the kind of failure that does real damage.

So the reassurance in the headline is not wrong so much as incomplete. Falling insolvencies are not the same as falling risk. They can be the surface of a market where risk is migrating from many small places to a few large ones, which is a harder problem to pinpoint. 

Why this matters more than it used to

For most of my career, the gap between a reassuring headline and a shifting reality was absorbed by human judgement. An experienced analyst reading a monthly release would feel the discomfort in that administration figure and go and look. That instinct was the safety net.

We are now automating away the very layer where that instinct lived.

Credit and risk teams across the market are moving decisioning into AI systems, and in 2026 that increasingly means agentic AI: models that do not just score an application but take steps, pull data, reach a conclusion and act on it with far less human handling in between. 

The promise is real and I am largely in favour of it. Handled well, it removes drudgery and lets good analysts spend their time where judgement actually pays. But handled badly, it does something more subtle and more dangerous. It industrialises the reading of the headline while discarding the instinct to look underneath.

I always say, an agent is only ever as good as what it has been given to reason over. I have written before about data being the missing layer in agentic AI for financial risk, and this is precisely what I meant. Give a brilliant model stale or shallow data and you get a brilliantly wrong answer, fast and with full confidence. Point one at a falling insolvency rate and it will, quite correctly, report that conditions are improving. It has no unease, no gut instinct. It will not pause on an administration rate climbing while the headline falls unless the data feeding it carries the granularity to see it and the signals to know it matters. So when you strip the technology back, the question becomes simple. What is this decision actually being made on?

The regulators have already noticed

It is true that none of this has been lost on supervisors. As agentic systems have moved from experiment into live banking operations this year, the consistent message from regulators has been that the absence of agent-specific rules does not mean that there aren’t expectations attached to users. The European Union’s AI Act has begun placing explainability requirements on financial systems, and supervisors on both sides of the Atlantic have made clear that existing model risk principles, from ongoing monitoring to the ability to challenge an output, apply to these tools whether or not the rulebook names them yet.

I read that as a warning worth taking seriously. A model you cannot interrogate is a liability the moment it is wrong, and an automated decision built on thin data is fast but risky. 

It’s the foundation, not the model

The instinct in a moment like this is to reach for a more sophisticated model. I would argue the opposite. The model is rarely the weak point. The data underneath it is.

A falling insolvency rate that hides a rising administration rate is a data problem before it is a modelling one. The failure was visible. It was sitting in the composition of the figures, in the connections between the entities behind a single collapsing group, in the early distress signals those businesses will have been giving off well before they filed. The tools to see it existed. What was missing was a system built to weigh the signal over the headline.

This is the work we have always tried to do at Company Watch, and it is why our H-Score® looks at the financial health of a business as a continuous, forward-looking measure rather than a snapshot, and why our distress signals and early warning monitoring exist to flag the movement beneath a stable-looking surface before it reaches a filing. What matters now is getting data of that quality in front of the agent at the moment it decides, which is why we have made ours reachable directly by AI systems through our MCP. An agent making credit decisions is a powerful thing when it is reasoning over data like that. 

Where I think this leaves us

As we know, the sector is going to automate more of its credit decisioning. That is settled, and on balance I think it is a good thing. The open question is what we automate on top of. If we hand these systems a diet of reassuring toplines, we will build machinery that is confident precisely when it should be worried, and we will find out where the risk really went when a large administration lands on a portfolio nobody flagged.

The falling number is only the surface. The shape of what is failing is the real story. Anyone building the next generation of credit decisioning owes it to their own book to make sure the data underneath can see that shape, and that their agents can actually reach it when it counts. That last part is the problem we built our MCP to solve.