The unit economics of collections: why human-led recovery stops working at high volume.

There is a number every collections leader knows but rarely says out loud. The breakeven point. The debt value below which recovering the money costs more than the money is worth.

For most digital lenders operating at scale, that number is somewhere between £100 and £300, depending on market, channel mix, and team efficiency. Below that line, human-led collections is economically irrational. The accounts sit. They age. They write off. And the loss rate becomes a structural feature of the business model, rather than a problem to be solved.

Understanding why this happens — and why the unit economics change fundamentally with AI — is the key to understanding why collections is broken at scale, and how it gets fixed.

Building the model

Start with what it actually costs to collect a debt with a human agent.

A single contact attempt — agent time (approximately 8 minutes including preparation and logging), dialler costs, telephony, and infrastructure overhead — runs to £6 to £14 in most markets. That is a fully-loaded cost per attempt, not just the call charge.

Reaching a delinquent borrower typically requires 8 to 12 contact attempts. Some will answer on the first try. Many will not answer for days. By the time you have first contact, you have spent £80 to £150 in operational cost.

Not every first contact produces a payment. Recovery rates from first contact vary by portfolio quality, vintage, and market — but 20% to 35% for early-stage delinquency is a reasonable range. On a £200 debt with a 25% recovery rate, the expected recovery value is £50. The expected cost to reach that outcome is £100.

The model is upside down.

Why digital lending books are particularly hard

Traditional consumer lending — personal loans, mortgages, auto finance — tends to produce high-balance delinquency. The economics of human collections were never great, but they worked well enough on large balances.

Digital lending inverted the distribution. BNPL transactions average £80 to £250. Salary advance products advance £200 to £600. Micro-loans in emerging markets are smaller still. The entire business model is built on high-volume, low-balance lending — which means the delinquent book is high-volume and low-balance by design.

For these books, the human collections model does not just underperform. It fails structurally. The cost of attempting recovery on a £120 debt is greater than the maximum possible recovery value. There is no version of human-led collections that makes the economics work at this balance point.

The problem is not that digital lenders are bad at collections. The problem is that they inherited a model built for a different type of debt.

What changes with AI

The fundamental shift is in the cost structure. Not optimised — inverted.

When an AI agent initiates a collections conversation, the marginal cost per contact attempt approaches zero. The infrastructure exists. The conversation is initiated. No agent time is consumed. No telephony cost per call in the traditional sense. The tenth thousand conversation costs the same as the first.

This changes the breakeven calculation entirely. On a £120 debt, if the cost of the contact attempt is effectively zero, then any recovery is positive. A 15% recovery rate on a portfolio of 10,000 small-balance accounts that would previously have been written off represents significant recovered value — at near-zero incremental cost.

The portfolio that was not worth chasing becomes worth chasing. Every account becomes worth chasing.

The compounding effect

The change in unit economics does something else: it shifts where investment is worth making.

In a human-led model, investment goes into contact volume. More agents, more calls, more attempts. The constraint is capacity.

In an AI model, contact volume is not the constraint. The constraint — and the opportunity — is conversation quality. Every pound invested in improving how the AI negotiates, how it handles objections, how it proposes repayment arrangements, directly improves recovery rate across the entire portfolio. Not just the high-balance tier. Every account.

At 80,000 accounts, the difference between a 15% and a 25% recovery rate is not marginal. It is the difference between a collections function that bleeds and one that performs.

The new economics

The lenders who understand this are not trying to make human collections more efficient. They have recognised that the model itself is broken for the type of lending they do — and they are building on a different economic foundation entirely.

For them, the question is no longer whether they can afford to chase a debt. Every debt is worth chasing. The question is how to make every conversation as effective as possible.

That is a very different problem. And it is a solvable one.