Every servicing team has the same problem: the work is rule-bound, but the rules are buried in policy documents, FCA CONC guidance, Consumer Duty outcome frameworks and, in the US, CFPB supervisory expectations under UDAAP. A payment arrives three days late. Does the account attract a fee? Is the borrower in a vulnerable circumstance that would make that fee inappropriate? Does the collector's next contact need to follow Regulation F's seven-day call limit? The answer is never one rule; it is a chain of them, and the chain must be walked consistently at volume.
Agentic loan servicing means software agents that read those rules as structured policy, decide what action to take, take it against the ledger and record every step with enough precision that a compliance officer, an FCA supervisor or a CFPB examiner can reconstruct exactly what happened and why. That is the shift. Not automation in the old sense of scripts, but agents that reason inside a defined policy boundary.
What an agentic servicing layer is made of
Agentic servicing is not a chatbot sitting in front of a legacy system. It is a set of components that share a common ledger and a common policy store, so that every agent acts on the same facts and within the same rules. The architecture has four parts that must work together.
- Policy as code. Servicing rules, FCA CONC forbearance requirements, Consumer Duty fair outcome obligations and US FDCPA contact restrictions are expressed in a form the agent can read and apply, not in a PDF it cannot. When a rule changes, the code changes once and every agent picks it up.
- An event-sourced ledger. Every balance, accrual, fee and adjustment is written as an immutable event. Agents read from the same loan ledger that produces statements and feeds reporting. There is no separate servicing database that can drift from the book of record.
- Agent runtime with bounded authority. Each agent has a defined scope: it can waive a fee below a threshold, issue a hardship arrangement up to a term limit, or send a contact, but it cannot exceed those bounds without an approval queue. The scope is explicit, not implicit.
- A full audit trail. Every action the agent takes is logged: the policy version it applied, the ledger state it read, the decision it reached and the output it produced. That log is the evidence a lender needs under Consumer Duty's outcome monitoring obligations and, in the US, under CFPB examination requests.
How an agent handles a late payment in practice
Take a borrower whose £350 instalment arrives two days after the due date. The steps below are not hypothetical. They are the sequence Kontor's servicing agents walk, governed by the policy the lender has configured.
- Read the ledger state. The agent retrieves the account's current balance, arrears history and any active vulnerability flag. It works from the event log, so the state it reads is the same state a human agent would see on the same screen at the same moment.
- Apply the policy tree. The late-payment policy is structured: if the account has a vulnerability flag, suppress the fee and route to a human review queue. If not, check whether the borrower has a prior arrangement in force. If so, the fee is waived. If not, a fee may apply, subject to the product's CONC-compliant terms.
- Write the outcome to the ledger. The agent posts the fee event, or posts a fee-waiver event with the reason code attached. Both are immutable. Neither can be quietly corrected later without a new event that records the correction and its authorisation.
- Trigger the next contact. If the account remains in arrears after the fee decision, the agent schedules outreach inside the contact rules: timing, channel and content governed by the same policy store. In the US, that means respecting Regulation F's call cap and the borrower's stated channel preference.
- Escalate if the boundary is reached. If the arrears exceed the agent's authority, say a third consecutive missed payment, the case moves to an approval queue for a senior agent or a human. The agent does not guess. It stops at the boundary and hands over with a full account of what it has already done.
The result is a servicing record that is consistent, explainable and ready for audit without a manual reconstruction exercise. For lenders subject to Consumer Duty's obligation to demonstrate good outcomes, and for US lenders under CFPB's UDAAP scrutiny, that consistency is not optional: it is the evidence base.
Why bounded authority matters more than model capability
The instinct when building agent systems is to focus on what the model can do. The harder and more important question is what the agent is permitted to do. An agent that can negotiate any arrangement, waive any fee or send any contact is not a servicing agent under CONC or UDAAP: it is an uncontrolled point of regulatory exposure. The Agent Operating Procedures pattern treats policy as the constraint, not the capability. The agent's intelligence is applied inside a defined space, and everything outside that space requires a human decision or a policy change that is itself versioned and audited.
This also answers the operational objection that agents will make inconsistent decisions. Inconsistency in servicing usually comes from human variability against poorly documented policy, not from agents operating inside well-specified rules. A well-specified agent will be more consistent than a team working from a PDF handbook, and its consistency is provable because the log exists.
Agentic loan servicing is not a feature to add to an existing platform. It is what you get when policy, ledger and agent runtime are built as a single system from the start, so that every action is governed, every outcome is recorded and every audit request can be answered from the log rather than from memory.
If you are working through what this architecture would mean for your book, speak to the Kontor team.