A borrower misses a payment. Your servicing queue grows. A human agent opens the account, checks the arrears balance, reads the hardship flag, decides whether to send a letter or make a call, logs the outcome, and moves to the next case. At volume, that chain takes long enough that a material number of accounts age past the point where early intervention works. The rule in CONC 7.3 requiring firms to treat customers in default with forbearance does not care how large your queue is.
AI servicing agents handle that chain automatically, account by account, without the queue. The question worth asking is not whether they are fast, but whether they are auditable, constrained and safe to run under Consumer Duty.
What a loan servicing agent is made of
A servicing agent is not a chatbot wrapped around a database. It is a process that reads account state, reasons over a defined set of permitted actions, executes one, and writes the result back to the ledger as an immutable event. The components that matter to a lender are these.
- A live ledger read, the agent pulls current balance, arrears days, scheduled instalments and any flags such as hardship or dispute before it does anything. Stale data produces wrong decisions; the ledger must be the system of record, not a cache.
- A bounded action set, the agent can only take actions explicitly listed in its operating procedure: send a templated notice, apply a fee waiver, defer an instalment, escalate to a human. It cannot improvise outside that list.
- A jurisdiction rule layer, before any action executes, the agent checks whether it is permitted in the borrower's jurisdiction. A payment holiday valid under UK consumer credit rules may not be available under a different regulatory pack. Jurisdiction packs encode those constraints so the agent does not have to guess.
- An event-sourced audit trail, every decision, the inputs read, the rule applied, the action taken, is written as a timestamped, immutable event. A regulator or an upheld complaint can replay exactly what happened and why.
- A human escalation path, any case that falls outside the agent's permitted action set, or that triggers a vulnerability indicator, is routed to a human immediately. The agent does not attempt to handle what it is not authorised to handle.
How an agent works through a past-due account
Take a concrete case. A direct debit returns unpaid on day one of arrears. Here is what a governed servicing agent does, step by step.
- Read account state. The agent pulls the ledger record: balance, days past due, payment history, any open disputes or hardship flags. It checks the business-date clock to confirm the arrears calculation is correct before proceeding.
- Select permitted action. The agent matches the account state against its operating procedure. Day-one arrears with no prior hardship flag: send a standard missed-payment notice. The notice template is pre-approved; the agent fills the variable fields from the ledger and does not draft free text.
- Check jurisdiction constraints. The operating procedure confirms the notice is compliant with CONC 7.9 timing requirements and that no cooling-off or dispute window blocks it.
- Execute and write the event. The notice is dispatched. The agent writes a ledger event: action taken, rule applied, timestamp, agent version. Nothing is held in the agent's own memory; the ledger carries the record.
- Set the next trigger. If payment is not received within the window defined in the procedure, the agent schedules the next step automatically: a follow-up contact attempt, a forbearance assessment, or escalation to a collections specialist depending on what the procedure permits at that stage.
What this gives a lender is consistent first-response time on every account regardless of queue depth, a complete decision log that satisfies past-due credit audit requirements, and a clear boundary at which a human takes over. For a compliance team preparing for a Consumer Duty review, that boundary and that log matter as much as the speed.
Why a large language model on its own is not enough
A general-purpose language model can draft a sympathetic arrears letter. It cannot guarantee that the letter is sent at the right point in the arrears timeline, that it does not contradict a live dispute flag, or that the action it chose was the one your credit policy authorised for this account type on this day. Those guarantees come from the operating procedure and the ledger, not from the model. Without them, every decision is unrepeatable and unexplainable, which is a regulatory problem before it is a technology one. The FCA's expectation under Consumer Duty is that firms can demonstrate outcomes, not just intentions.
Kontor's agent operating procedures sit between the model and the ledger. They define what the agent may read, what it may do, and what it must escalate. The model handles language and reasoning within that envelope; the procedure enforces the boundary. A lender can update the procedure when policy changes without retraining anything, and the audit trail records which version of the procedure governed each decision. That separation is what makes AI servicing agents safe to run at scale, rather than safe to demo.
AI agents make loan servicing faster at scale only when they are constrained by explicit operating procedures, grounded in a live ledger and bounded by jurisdiction rules: speed without those constraints is liability, not efficiency.