Most lenders already run some form of automation: rule-based decisioning, scheduled payment runs, templated arrears letters. The question is not whether to automate, but whether the system underneath can support an agent that reads a borrower's account, forms a view and acts, with every step visible to a regulator the following morning.
The answer depends almost entirely on the ledger and the audit trail, not on the AI model. Get those right and agents can do real servicing work. Get them wrong and the agent is a liability, because no one can reconstruct what it decided or why.
What an AI loan management system is made of
Four components have to work together before any agent touches a live account. Bolt AI onto a legacy core and you are missing at least three of them.
- Event-sourced loan ledger. Every balance change, fee, payment and rate adjustment is stored as an immutable event. The current state is derived from that sequence, so the system can reconstruct any account as it stood at any moment. This matters for FCA Subject Access Requests, FCRA dispute timelines and any complaint that turns on what the borrower was told when. See how the two-clock ledger design handles transaction time and business time separately.
- Decision engine with reason codes. Origination, limit changes and hardship arrangements all pass through a single decisioning layer that records not just the outcome but the inputs and the rule version that produced it. Under CONC 5 in the UK and Regulation B in the US, a borrower who is refused credit is entitled to a reason. The engine has to be able to produce one, not infer it after the fact.
- Agent runtime bound to policy. AI agents that handle payment deferrals, arrears outreach or complaint triage operate inside a defined procedure, not open-ended prompts. Each step is logged. Escalation paths are explicit. The Agent Operating Procedures model sets out how that works in practice.
- Jurisdiction pack. Consumer credit rules differ by market. In the UK, CONC and Consumer Duty govern how arrears must be handled and what a fair outcome looks like. In the US, the FDCPA and Regulation F govern collections contact, TILA and Regulation Z govern disclosure, and UDAAP sits across all of it. A jurisdiction pack binds the platform to the rules of a specific market so agents cannot take actions that are compliant in one country but unlawful in another.
- Audit trail. Not a log file: a structured, queryable record of every agent action, every system event and every human override, with timestamps that correspond to when the event occurred and when it was recorded.
How an agent works through a real servicing event
Take a borrower on a £5,000 personal loan who misses a payment. The arrears process in an AI loan management system runs like this.
- The ledger records the missed payment as an event. The account moves into arrears. No balance is adjusted speculatively. The event is timestamped in both transaction time and business time, so if the payment later arrives with a value date three days earlier, the ledger can backdate correctly without corrupting the audit trail.
- The collections agent reads the account state and selects an outreach path. It checks the borrower's contact preferences, any existing vulnerability flags and the number of prior arrears episodes. Under Consumer Duty, the FCA expects firms to act in borrowers' interests, not just follow a script. In the US, Regulation F limits the number of calls per week and prohibits contact at certain times. The agent applies whichever rules govern this account's jurisdiction before it sends anything.
- The agent makes contact and logs the response. If the borrower agrees to a one-month deferral, the agent records the arrangement, updates the schedule on the ledger and generates the required notice. In the UK, CONC 7 requires that any arrangement be confirmed in writing. The system produces that confirmation from the ledger event, not from a separate document store.
- If the borrower does not respond or the situation escalates, the agent raises an approval queue item. A human operator sees the account, the agent's action log and the reason for escalation. They can approve, override or refer. Every choice is recorded.
- The closed event is available to compliance the next morning. What the agent did, why it did it, which policy version was active and what the outcome was: all queryable, all immutable.
That sequence is what the FCA means when it asks for evidence of fair outcomes under Consumer Duty, and what the CFPB expects when it examines complaint handling and collections conduct. A lender can answer both with a single audit trail rather than reconstructing events from emails and spreadsheets.
Why bolting AI onto a legacy LMS does not work
Legacy loan management systems were built to store state, not events. When a payment arrives, the balance updates and the prior state is gone. An AI agent reading that system can see today's account, but it cannot reconstruct yesterday's. It cannot tell whether a fee was applied before or after a complaint was raised, which is exactly the question a regulator or an upheld FOS complaint will ask. The agent's output is only as trustworthy as the data it reads, and a state-based ledger gives it a partial picture at best. This is not a model problem; it is an architecture problem. See what the architecture requires for AI lending platforms to understand the full dependency stack.
Kontor is built event-sourced from the ledger up, so agents have a complete, ordered account history to work from. That is what makes it possible to let an agent handle an arrears arrangement, a rate change or a complaint triage step without a human checking every action before it runs.
An AI loan management system earns that name when every agent action is grounded in a complete ledger, constrained by the rules of its market and recoverable by audit, not when a language model is given access to a database and told to be helpful. If you are assessing platforms, start with the ledger and the audit trail, because the AI is only as good as what it can read and prove. Talk to the Kontor team if you want to see how that works against your own book.