Are collections agents accountable for an AI agent’s mistakes?

They are accountable for what they approve and send, and for spotting what the AI agent flags. A collections team member who approves a payment plan or a contact message owns it, including whether it treats a struggling customer fairly. They do not own an error the AI agent made inside bounds they never set. The record separates the two.

Why is collections a high-stakes place for approvals?

Collections touches people in difficulty. A poorly timed message, an unaffordable plan or a missed sign of vulnerability causes real harm, and regulators watch it closely. In the UK, for example, the FCA's Consumer Duty expects firms to deliver good outcomes for retail customers, which reaches directly into how arrears are handled.

An AI agent in collections drafts contact messages, proposes payment plans from income and spending data, records promises to pay and schedules follow-ups. The collector reviews what is held. That review is where judgement about the customer's situation enters, and where accountability attaches.

What may the AI agent do, and what stays with the collector?

• Allowed within bounds: sending routine reminders inside the firm's contact limits, logging a promise to pay, updating a due date the customer chose from approved options.
• Held for a collector: any payment plan, any message after a missed promise, any account showing signs of vulnerability, and any contact outside standard hours or frequency.
• Refused: contact that would breach the firm's own frequency limits, and threatening language outside approved templates.
• The collector keeps: assessing affordability when the figures are declared rather than verified, deciding on forbearance, and recognising distress in what the customer actually said.

What does the record show about who decided?

Each plan the agent proposes is filed with its sources. An income figure from bank statement data the customer shared through a verified channel is closer to signed. A figure the customer typed into a web form is declared. A note from a previous call saying "lost job last month" is declared too, and an agent trained to flag vulnerability should hold the account on the strength of it.

The record then names the collector who approved, the time, and what was on screen. If the vulnerability flag was there and the collector approved a standard plan, the decision is theirs. If the agent never raised it, the gap is in the agent's escalation rules, and in whoever certified them.

An example: the plan that looked affordable

Take a consumer lender. A customer is two payments behind on a loan with 1,200 outstanding. The agent proposes a plan of 200 a month, based on declared income from a form filled in a year ago. A call note from last week, also on file, records the customer saying they had been made redundant. The agent flags it and holds the plan.

The collector, handling a busy queue, approves the plan as drafted. The customer misses it, and the account rolls into further arrears. The record shows the flag, the old income figure and the approval. The collector is accountable for overriding a vulnerability flag without speaking to the customer. Their manager is accountable for queue targets that made that likely.

Which controls keep approvals fair?

1. Vulnerability flags cannot be cleared in bulk. Each needs a person to open it.
2. Stale income data triggers a hold. Declared figures older than the firm's limit should never drive a plan alone.
3. Collectors practise before approving alone. A two-week practicum for collections agents builds this habit, logged in a practicum log template for collections agents.
4. Skills keep pace with the AI agent. Its certificate lapses after 90 days and contact rules shift; see how to prepare collections agents for 90-day recertification.

Customer-facing teams face a similar split; compare customer support agents and AI agent mistakes. The wider topic is in AI literacy and certification for people who work with agents.