Collections teams oversee agents that contact customers in arrears, propose payment plans and draft notices with real consequences. Article 14 of the EU AI Act applies to high-risk systems, and collections agents can cross into that category when they assess creditworthiness. Either way, overseers need the competence, authority and practical means to stop an agent treating a struggling customer badly.
What does Article 14 ask, in plain words?
Article 14 requires that high-risk AI systems be overseen effectively by people while in use. Those people must be able to understand what the system can and cannot do, monitor it for anomalies, stay alert to automation bias, interpret its output correctly, decide not to follow it or reverse it, and intervene or stop it. Oversight must be proportionate to the risk, the system's autonomy and the context. Article 26 then requires deployers to give oversight to people with the competence, training and authority to do it.
"Oversight" here is not a signature at the end of a process. It is a person with the knowledge and power to change the outcome. The distinction is set out in what meaningful human review is.
Is a collections agent high-risk?
It depends on what the agent does. The Act treats creditworthiness assessment and credit scoring of individuals as high-risk. A collections agent that only schedules contact and drafts reminders is not assessing creditworthiness. One that decides whether a customer qualifies for a new payment arrangement, a restructured loan or further credit may well be, because that decision rests on an assessment of the customer's ability to pay.
Most lenders do not split hairs. Collections touches customers at their most financially vulnerable, consumer protection rules in every EU market expect fair treatment of borrowers in difficulty, and the GDPR applies to the data throughout. Applying Article 14 discipline to every collections agent is simpler and safer than arguing about which ones need it. Insurers face a similar question about which claims agents fall inside, covered in Article 14 oversight for AI agents in insurance claims.
Where does a collections agent need a person?
A typical governed collections agent works at several depths:
- Depth 1: reads account history, payment records and contact notes.
- Depth 2: drafts reminder messages, payment plan proposals and arrears notices for a person to send.
- Depth 3: sends routine reminders within agreed templates, frequency and hours, and records promises to pay.
- Depth 4 (held): applies a fee, agrees a plan outside standard terms, refers to an external agency, or starts formal recovery. These are hard to undo, and a named person decides.
Two moments demand the most from overseers. The first is vulnerability. A customer who mentions illness, bereavement or job loss changes what fair treatment means. The agent should hold or refuse further automated contact when such signals appear, and the overseer must know how to respond: pause contact, adjust the plan, involve a specialist. The second is escalation to formal action, where the overseer is the last check before a step that affects the customer's credit record or legal position.
What competence, authority and intervention look like in collections
Competence means the overseer can read the agent's proposal and the facts behind it. They should know which facts are sealed, from the loan ledger; signed, from a verified party such as an employer confirming income; declared, by the customer on a call; or external, such as a forwarded email. A plan built on a declared income figure deserves a different conversation from one built on verified data. They should spot when the agent's tone or frequency would breach the lender's own standards.
Authority means the overseer can stop contact, waive a fee within limits, or reject an escalation without asking a manager, and is not measured only on cash collected. An overseer whose bonus depends on recovery rates has a reason not to intervene; Article 14 is about removing that tension, not ignoring it.
Intervention means knowing how to pause the agent for one customer, for a segment, or entirely, and what happens to scheduled contact when they do. An agent's certificate lapses after 90 days, and a run that clears no depth drops it to depth 1 the same day, which stops automated sending. Teams must be able to take over without missing contacts that customers are relying on, such as confirmations of an agreed plan.
How do you meet Article 14 for collections agents?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Overseers understand the agent's limits | Training on depths, templates, bounds and known failure cases in collections | Training and assessment records |
| Awareness of automation bias | Seeded cases, including vulnerability signals, placed in the review queue | Detection rates on seeded cases by overseer |
| Correct interpretation | Overseers trained to check provenance of income and hardship facts | Decline and edit reasons citing the facts |
| Power to override | Overseers can pause contact and reject escalations without approval | Pauses and rejections in the record |
| Ability to stop | Named staff can suspend the agent and run manual contact | Suspension events and cover timings |
| Competent people assigned (Article 26) | Approval of depth 4 actions tied to a current credential | Credential register matched to approvals |
The pattern of proving training this way is common to other operational sectors, as logistics companies proving staff are trained to oversee AI agents shows.
What will a supervisor ask?
A supervisor or consumer protection authority is likely to start from a complaint. Show the contact history for this customer. Who approved the fee and the referral? What did they see, and did it include the hardship the customer mentioned? How many vulnerability signals did the agent flag last month, and what happened to each? How are overseers paid and measured? Can you stop all automated contact within minutes if something goes wrong?
What does Article 14 not cover?
It does not set collections conduct rules; national consumer credit and debt collection rules do that, and they apply whether or not an agent is used. It does not apply to agents that are not high-risk. It does not prescribe training content. It does not stop a lender from collecting debts, only from doing so through a system no one can check. And it does not displace the GDPR. For a related sector with similar customer impact, see Article 14 oversight for AI agents in payments.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where the Academy fits
Pryme Intelligence Academy teaches human review, teaching the agent and handling what goes wrong on the real product, with module tests set in the learner's own industry and role, so a collections team lead practises on arrears cases. Our guide to AI literacy and certification for people who work with agents shows how the Operator credential fits alongside the other levels.
Questions people ask
Can an agent agree a payment plan on its own?
Within standard terms set by the lender, an agent at depth 3 can record and confirm a routine plan. Plans outside those terms, or involving a customer showing vulnerability, should be held for a person.
How should overseers handle a customer who asks to speak to a human?
The agent should hand over without friction, and the overseer should treat the request as a signal to review the case. A customer who asks for a person is often one the agent is not serving well.
What training do overseers need on vulnerability?
They need to recognise the signals, know what changes in treatment each one triggers, and practise on realistic cases. Training should include cases where the agent missed the signal, so overseers learn not to rely on it.