An SME that is about to move its banking relationship rarely announces it. The account does not close. It goes quiet. Payment volumes thin out, the operating balance drifts down, a payroll run appears somewhere else, and the treasury products that were meant to make the relationship sticky stop being touched. By the time somebody notices, the decision was made a quarter ago.
A design put to a consumer and SME banking group operating across several European markets treats that quiet period as the product. One of three lending and SME use cases in the deck, it puts an agent on portfolio health rather than on credit risk: monitor accounts for declining usage patterns, identify the early warning signs of attrition, and hand the relationship manager a proactive engagement script.
Nothing in this had been run when the deck was written. The autonomy level, the outcome figures and the accuracy targets are all design commitments in a board workshop document, which is how they should be read here.
Two questions that read the same account
A commercial bank already watches SME accounts. It watches them for credit. Exposure, covenant headroom, arrears, the things that decide whether a facility is still safe. That machinery exists, it is regulated, and in the same programme it is being extended so the whole book gets continuous monitoring rather than the top slice getting periodic review. I have written about that side separately, in Why Only the Top Fifth of a Credit Portfolio Gets Real Monitoring.
This use case reads the same account and asks a different question. Not whether the client can still pay, but whether the client is still there. The two questions diverge in an important way: a departing SME is very often a perfectly good credit. It pays on time all the way out of the door. Everything the credit machinery is tuned to detect stays green while the relationship dies, because nothing about leaving looks like distress.
That is why the retention signal has to be built as its own thing rather than bolted to the risk model. The risk model is not blind by accident. It is looking somewhere else on purpose.
Nobody files a complaint before they leave. They just use you less, and every part of that decline is already sitting in your transaction data, unread.
What the agent is actually doing
The design assigns three tasks, and the ordering matters more than the technology.
Watch usage, continuously. Not a quarterly portfolio review. The pattern being detected is a trend, and a trend sampled four times a year is mostly invisible. The value of putting an agent on this is not that it detects something a human could not. It is that it looks every day, at every account, including the ones too small to have earned a human's attention.
Identify the early warning. Decline is not one number. It is a shape across payment volume, balance behaviour, product usage and channel activity, and the shape means different things for a seasonal business than for a steady one. This is the part that needs a model rather than a threshold, and it is also the part the deck is honest about being unbuilt.
Suggest the engagement script. This is the one most banks would drop, and it is the one that makes the rest useful. The output is not a flag on a dashboard. It is a specific thing to say to a specific client about a specific observed change, delivered to the person who already has the relationship.
The step that decides whether any of it works
I have watched enough early-warning systems get built to know where this one fails, and it is not in the model.
An alert without an owner is a notification, and notifications get muted. The design here routes to the relationship manager, which is right, because the relationship manager is the only person who can act on it. But it also puts a new obligation on a role that is already full. Somebody has to decide that responding to a quiet-account flag is part of the job rather than an interruption to it, and that decision is made in capacity planning, not in the model.
The script is what makes that possible. A relationship manager who receives "account 4417 shows declining engagement" has been given work. One who receives "their payment volume has fallen by a third over two quarters, their payroll run stopped appearing in March, here is what to ask" has been given a call to make. The difference is small on the page and total in practice.
The measurement problem nobody solves in the deck
Retention interventions are genuinely hard to evaluate, and the design does not pretend otherwise, because it does not address it at all.
The problem is the counterfactual. When the relationship manager calls and the client stays, you cannot tell whether the call worked, whether they were never leaving, or whether they left six months later for the same reason. Credit models get clean labels eventually, because default is an event. Attrition intent is not an event, and a saved relationship looks exactly like a relationship that was never at risk.
The only honest way through it that I have seen work is to hold out a control group deliberately, accept that some of those accounts will leave, and measure the difference. That is an uncomfortable thing to propose to a board and it is the difference between a retention programme with a number on it and one with a story. If you are building this, decide about the holdout before you build, because afterwards there is no way to reconstruct it.
What I would ask before funding it
Who owns a quiet account today. If the answer is nobody, the model is not the first thing to build. The role is.
What does the relationship manager receive. A score, or a sentence they can say. Only one of those changes a quarter.
Is the signal separate from the credit signal. If retention is a field on the risk model, it will inherit the risk model's blind spot, which is that good payers leave.
How will you know it worked. Ask before you build, and be willing to hold accounts back to find out.
The broader pattern is one I keep meeting in agentic designs. The model is the part everyone specifies and the cheap part to get right. The expensive part is the human step at the end, and whether anybody's Monday actually changes when the system produces its output.
Drawn from one slide of a board-level AI transformation deliverable for a consumer and SME banking group operating across several European markets, covering three lending and SME use cases at a stated autonomy level. Nothing described here had been built when the document was written: the accuracy objectives and the autonomy level are design commitments, not outcomes. The credit-side monitoring work from the same programme is a separate published piece, linked above. The reading of the retention signal as structurally distinct from the credit signal, and everything said here about ownership and measurement, is ours.
“Nobody files a complaint before they leave. They just use you less, and every part of that decline is already sitting in your transaction data, unread.”
Get in touch
Put RealAI’s applied-AI team on your hardest data problem.
We help enterprises move from pilots to production: sovereign models, governed data, and agents you can audit. Start with a value-first assessment.