AI for collections in banking: early arrears, affordability and the evidence trail
A bank's collections function is judged twice. Once on cash, through roll rates, cure rates and the balance recovered before charge off. And once on conduct, through whether each customer in difficulty was treated fairly, whether forbearance was offered where it was due, whether vulnerability was recognised, and whether that can be shown from the record eighteen months later rather than from an agent's memory.
Those tests only conflict when the process is weak. Most of what an arrears book loses, it loses to friction: contact that arrives late, affordability packs abandoned at the second call, arrangements set above what anyone could sustain, files that reach litigation missing the document that would have justified getting there. That friction is back office work between the collections system, the core platform, the document store and the customer's own paperwork.
The arrears book is several different books
A card balance in cycle one, a mortgage three payments down, an overdraft that never cleared and an SME term loan with a personal guarantee behind it share a worklist and nothing else. Each carries different forbearance options, different notice requirements and its own definition of a good outcome, so segmentation has to be structural before it is behavioural: product, security, arrears stage, live arrangement, vulnerability flag, open dispute, moratorium.
Pre-delinquency and early arrears contact
The cheapest arrears to cure are caught before a missed payment hardens into a stage. Two populations matter. The administrative miss: a direct debit that failed on a date change, a card that expired, a salary that landed late. And genuine deterioration, visible in patterns that precede default: buffer eroding month on month, utilisation climbing while repayment stays at the minimum, payments to other creditors failing first.
An agent suits both jobs, ranking those signals into a review queue with the evidence attached and drafting the outreach. Early arrears contact that reads as a demand pushes the customer away from the conversation you wanted, and the difference between opening a dialogue and triggering a complaint is wording, sequencing and channel. It should cite figures it actually read, then hold the draft for approval rather than send unsupervised, because this is a regulated communication to someone who may already be in difficulty.
Affordability, income and expenditure
Every decision downstream rests on an affordability picture: income including benefits, essential expenditure, priority debts such as rent, mortgage and council tax, then non-priority credit, ending at a disposable figure to be shared fairly across creditors rather than claimed by whoever called.
Bank statements and income evidence
The slow part is not judgement, it is transcription. Statements arrive as PDFs and photographs, wage slips as scans, award letters as screenshots, and somebody retypes them. An agent reads those documents, categorises the figures against your own expenditure headings, identifies recurring income and its true frequency, separates transfers between the customer's own accounts from real income, and produces a draft statement with every line traceable to the page it came from. Where the bank holds the current account, it can also reconcile declared figures against observed transactions and surface the gaps for the adviser.
The arithmetic belongs in code
Subtotals and threshold comparisons belong in deterministic code, not in a model. A plan built on a wrong subtotal fails by the second instalment and reads later as an unaffordable arrangement the bank imposed. And the customer's own account of their circumstances stays the record: an agent may question an inconsistency, never replace a declared figure.
Forbearance is a decision, the file is the deliverable
Reduced payment, suspension of interest and charges, term extension, capitalisation of arrears, part interest only on a secured loan, token payments, referral to free debt advice, and in hardship cases partial or full write off. Each carries consequences for the customer's credit file, cost of borrowing and, on a mortgage, their home, so each belongs to a person with authority.
The work around them does not. An agent can assemble the file before the decision: the affordability statement with its evidence, arrears history, prior arrangements and how each ended, the vulnerability record, the options open for that product, the projected outcome of each, and why each rejected option was rejected. It can draft the documentation, including the disclosures owed about interest still accruing, then check the file against your own policy checklist. That leaves the approver deciding rather than gathering.
Vulnerability is a gate, not a score input
A vulnerability indicator must act as a gate that runs before any prioritisation or treatment logic, never as a feature inside a score. A feature is a weight, and a large enough balance eventually outweighs it. As a gate it removes the account from automated contact and from every stage advance path.
The useful direction for an agent runs the other way: finding the vulnerability nobody recorded. Indicators sit in call notes, inbound letters, advice agency correspondence, references to a health condition or a carer role, signs of coercion or financial abuse, and language suggesting the customer did not follow what was explained. An agent reading the whole history can propose a flag with the evidence quoted. A trained human confirms it and chooses the treatment. Hardship follows the same shape: terminal illness, bereavement, loss of capacity and severe mental health crisis need a route that bypasses the worklist, and the agent's job is to recognise that fast and hand over.
Promises to pay and arrangements that break
A promise to pay is a small contract, and most systems track only whether it was kept. The useful reading is graded: paid on time, paid short, paid late, partial and recurring which usually means the plan was set above affordability, or nothing at all. An agent watching for the earliest indication, the failed direct debit representation rather than the missed instalment date, buys the window in which a plan can be revised instead of terminated: it gathers what changed, drafts the contact and proposes a revision. A plan that broke because it was never affordable calls for a lower plan, not a stage advance.
Triaging inbound arrears correspondence
Inbound is where banks quietly lose control of the timeline. One mailbox receives income evidence, complaints, advice agency letters proposing an offer, moratorium notifications, probate correspondence, insolvency notices and disputes about the balance. Several start a clock and some stop collections activity outright. An agent can classify each item, attach it to the right account, identify what it obliges the bank to do and by when, and open the correct queue with a draft reply. That prevents the expensive failure: recovery continuing while an unopened letter says it should have stopped.
Litigation, recovery handover and portfolio sale
Litigation is a trigger, not a stage, and it belongs behind a preflight list rather than a timer: balance verified against posted transactions, default notice issued accurately and served, no open dispute, no live arrangement, no vulnerability flag, forbearance genuinely offered and evidenced, address confirmed, limitation period checked, no insolvency or moratorium in force, correct legal entity, which on commercial lending includes whether a guarantee is relied on and whether it is enforceable.
An agent can evidence every item and refuse to mark the list complete when it cannot. It should not pull the trigger. The same holds for placement with an external agency, where recall conditions need monitoring afterwards because nobody internal is watching a placed account.
Sale of debt portfolios
Sale exposes whatever the book was hiding, and the exclusions are the whole exercise: live arrangement, open dispute, vulnerability flag, moratorium, probate status, statute barred balances, unresolved billing or interest errors. An agent screening the proposed tape against every exclusion and reconciling balance and documents account by account does work that is otherwise sampled. The bank warrants that data, stays accountable for the buyer's conduct, and may face buyback obligations on anything mislabelled.
The evidence trail is the product
For any account you should be able to show what contact was attempted and on which channel, what the customer said, what affordability evidence existed at the moment of decision, which options were considered, why the chosen one was chosen, which policy version applied and who approved it. Built properly the system produces that trail as a by-product, because each step records its inputs, output and confidence. Automation without an audit trail is automation you are not allowed to keep.
Where AI fails here
Poor account data defeats it. Duplicate customers, stale addresses, unposted payments and balances that disagree between systems produce confident, wrong output at volume.
Transaction categorisation is wrong at the edges in ways that matter: irregular self-employed income, transfers between a customer's own accounts, a benefit paid into a partner's account. Each can move a disposable income figure enough to change a decision.
Circumstances nobody wrote down are invisible. If a bereavement was mentioned on a call and never noted, no model recovers it.
Fluent and wrong is the dangerous failure. An arrangement letter quoting a balance that changed yesterday reads perfectly and is a complaint in waiting. Bind every figure in a customer-facing draft to a value read at drafting time.
Vulnerability detection errs both ways. A missed indicator harms someone, a false one is intrusive, and inside a score it is discriminatory.
A model trained on historical outcomes learns historical practice, including whatever in it was unfair. If past agents pressed hardest on the customers least able to push back, that is the pattern in the data. And non-response is not evidence of ability to pay.
Where to start
Begin with affordability document extraction into an adviser review queue. It has a checkable right answer, it touches no customer, the adviser still owns every figure, and it removes the slowest manual step. Inbound triage is a strong second. Forbearance file assembly and litigation preflight follow, both leaving the decision where it was. Shadow run each against current practice and read the disagreements before granting authority.
Why a recovery-maximising objective fails here
Set the objective to maximum recovery and an optimiser finds the customers where pressure works, many of them paying the bank instead of rent, pushing past a broken arrangement rather than reducing it, treating a vulnerability flag as a cost, pressing an account a limitation check would have closed. In arrears the right outcome is often a smaller, sustainable, evidenced arrangement rather than the full balance. So the objective is arrangements that hold and decisions that can be defended, with recovery as the consequence rather than the target, behind gates no score can override. When the system is not confident, it should stop and route to a named owner rather than guess.
If your arrears book holds arrangements that break faster than you can reset them, describe the process to Adopt AI Partners at hello@adoptaipartners.com and we will tell you which parts an agent should touch and which it should not.