AI in Debt Collection: How to Govern Collections Automation Responsibly

Governed AI for Collections, Not Generic Automation

Why putting AI into production means putting policy, accountability and customer treatment around it

One question stayed with me after our August webinar on “Moving Fast Without Losing Control”:
What does it actually take to put AI into a live collections operation with confidence?

The demonstrations are becoming easier to deliver and the use cases more tangible, with clear potential across risk prediction, segmentation, contact strategy, case summarisation, agent support, conversational engagement and compliance review. But collections is not a generic business process. We are often dealing with people under financial pressure, including vulnerable customers, and making decisions that affect how, when and why they are contacted. A recommendation that looks efficient in a demonstration can produce a very different result when it meets incomplete data, a complex customer situation or a policy boundary.

This is what leads me to believe that the next phase of AI in collections is not a race to be won, because if it is used responsibly everyone stands to benefit; the most effective businesses will be those that can demonstrate, transparently and with confidence, where AI is being used, what it is allowed to influence, how people remain accountable and whether it is delivering the right outcomes for customers.

Why Generic AI Automation Falls Short in Debt Collection 

Simply put generic automation is the wrong starting point because it begins with the technology and asks what can be automated, rather than starting with a clearly defined collections problem and asking what the right outcomes and subsequent treatments should be for the customer.

Collections is not a generic process, as every decision involves a particular combination of customer circumstances, account history, vulnerability indicators, policy rules, and available treatment options. This means an AI capability cannot simply be placed over an operation and expected to make the right decisions without being connected to this wider context.

To work effectively, AI needs to support the complete collections journey, ensuring that any prediction informs a permitted decision, that the decision leads to an appropriate action or customer interaction, and that the resulting outcome is captured and used to improve what happens next. If this approach is not adopted even a technically strong model could apply the wrong strategy, present incomplete information as fact or keep a customer within a digital journey when human support is needed.

This is why governance must cover the whole use case rather than the algorithm in isolation, bringing together the data, model, policy, workflow, people, action, and outcome so the organisation can explain not only what the AI recommended, but why the resulting treatment was appropriate.

AI Automation in Collections: What to Automate and Where Human Oversight Is Needed

During our recent online seminar, my colleague Konstantinos Kentrotis introduced me to the term “bounded automation”, and it provides a useful way to distinguish this approach from assistive AI and decision-support AI. Each plays a different role within collections and does not carry the same level of risk, so they should not be governed in the same way. To clarify:

• Assistive AI can help an agent understand a long case history or enable a supervisor to review a greater number of completed interactions.
• Decision-support AI can recommend a priority, treatment, timing, or channel while leaving the resulting decision with a collections colleague.
• Bounded automation can complete a defined digital journey, provide approved information, present permitted options, and escalate the case when the situation falls outside the established rules.

The word “bounded” is important because the objective is not to create open-ended autonomy, but to provide a safe and appropriate customer journey within a clearly defined decision space, with explicit limits on what the technology is permitted to do and when human support must take over.

A controlled hand-off to a skilled colleague should not, therefore, be seen as a failure of automation; in many cases, it is exactly the outcome the system should have been designed to produce.

Where Can AI in Debt Collection Create the Greatest Governance Risks? 

In my view, collections teams are most likely to lose control when AI becomes embedded within the operation faster than the data, people, and processes around it can adapt. Four risks stand out.

AI cannot manufacture reliable context from unreliable records, but good data does not necessarily mean perfect data; it means information that is sufficiently accurate, current, complete, consistent, and relevant to support the decision being made. Contact details, channel preferences, payment history, previous treatments, customer responses, vulnerability indicators, and arrangement outcomes all need to be connected so the organisation can understand not only what happened, but what action contributed to that outcome.

Harnessing this data requires more than bringing it together in one place. Firms need common definitions, clear ownership, and ongoing quality controls, while also linking decisions and treatments back to customer outcomes so the AI can learn from what actually happened. Without that connection, a model may identify risk accurately but still recommend an inappropriate treatment because it lacks the operational context required to turn the prediction into the right action.

A confident score, summary or recommendation can discourage challenge, particularly when it is presented as authoritative or when an agent is working under time pressure. Over time, familiarity can lead to complacency and there is a natural human tendency to follow the easiest available route, especially if previous recommendations have generally appeared to be correct; what begins as trust in the technology can therefore become an assumption that the technology must know best.

This is where organisations need to be particularly careful because accountability for the treatment does not pass to the AI. Agents need enough information to understand and challenge a recommendation, while managers should look at acceptance and override patterns to determine whether people are applying genuine judgement or simply following the system. “The model decided” is not an operating principle and should never become an explanation for an inappropriate customer outcome.

There is a danger that automation is associated only with digital self-service journeys, when in reality it can influence almost every part of a collections operation, including account prioritisation, dialler strategies, channel selection, contact timing, agent prompts, affordability assessments, offer presentation and workflow routing. A human may still be involved in the interaction, but their choices could already have been heavily shaped or restricted by automated decisions made earlier in the process.

The risk is therefore not simply that a customer becomes trapped in a digital journey, but that automation gradually removes choice, context, and meaningful judgement from the wider treatment process. Vulnerability, disputes, complaints, and complex affordability situations should have clear routes to appropriately skilled colleagues, while automation should be designed to recognise when the available information is uncertain or the customer’s circumstances fall outside the intended journey.

AI cannot be treated as something that is tested, approved, and then left to operate unchanged, because customer behaviour evolves, policies and strategies change, models drift and front-line teams inevitably encounter circumstances that were not visible during development. A meaningful feedback loop therefore needs to capture more than whether a customer paid; it should also include overrides, escalations, abandoned journeys, broken arrangements, complaints, policy exceptions, customer outcomes, and the reasons why a colleague decided not to follow a recommendation.

These signals should then be reviewed together to identify whether the problem sits with the data, model, policy, workflow, or training, with any resulting change made through a controlled process rather than as an informal adjustment. Without this closed loop, the organisation cannot improve the use case or demonstrate that it remains safe and effective as circumstances change.

AI Governance in Debt Collection: 7 Controls for Responsible AI

Governance can sound abstract, particularly when it is discussed separately from the day-to-day operation, but in collections it should be visible within every part of the customer journey. For me, there are seven elements that should sit around any material AI use case, with the level of control reflecting the nature and potential impact of the decisions being made.

Every use case should begin with a clearly defined business problem, an understanding of the decisions the AI will influence and a named business owner who remains accountable for the resulting customer treatment. The boundary between recommendation and execution also needs to be clear, including what the technology can do independently, where approval is required and which circumstances would cause the use case to be changed, paused, or stopped.

Firms need to understand which data is being used, where it comes from, how current and reliable it is and whether it is appropriate for the decision being made. They should also be able to connect the recommendation and resulting treatment to the customer outcome, because without this relationship it becomes difficult to assess whether the AI is improving the journey or simply producing technically accurate predictions that do not lead to better decisions.

Collections policy should be translated into practical and enforceable boundaries, including permitted and prohibited actions, contact rules, channel permissions, offer limits, mandatory disclosures, and escalation triggers. These guardrails need to operate across the whole collections process, rather than applying only to customer-facing digital journeys, so that automated prioritisation, agent recommendations, and workflow decisions remain within the same policy framework.

AI should be tested for accuracy, robustness, fairness, and its ability to deal with the less predictable cases that inevitably arise within collections. Explainability should then reflect the needs of the audience: an agent requires a clear and usable reason for the recommendation, compliance teams need evidence of the data, rules and controls applied, model-risk teams require technical validation, and customers or regulators may need a straightforward explanation of why a particular treatment occurred and how it could be challenged.

Human oversight needs to involve more than placing a person somewhere within the process, as that person must have the information, authority and time required to question the recommendation, override it, or take control of the customer journey. Appropriate challenge should be encouraged rather than treated as a failure to follow the system, particularly where vulnerability, complaints, disputes, affordability, or other complex circumstances require experience, empathy, and professional judgement.

The organisation should be able to reconstruct the customer journey, including the information available at the time, the model and policy versions used, the recommendation made, the action taken, any human intervention or override and the resulting outcome. Changes to models, prompts, policies, and workflows should also be tested, approved, and recorded through a controlled process, with the ability to pause or roll back the use case if performance or customer outcomes begin to deteriorate.

Monitoring needs to consider business value and the quality of control together, bringing recovery, kept arrangements, cost and operational capacity alongside contact frequency, complaints, exceptions, escalations, overrides, model drift, and outcomes across different customer groups. These measures should feed into a regular review process so that emerging issues can be traced back to the data, model, policy, workflow, or training, allowing improvements to be made before isolated problems become repeated at scale.

Taken together, these controls are not intended to slow the adoption of AI but to give collections, risk, compliance, and technology teams the confidence to use it responsibly, knowing that they can explain how decisions are made, intervene when necessary and demonstrate that the resulting treatment remains appropriate for the customer.

How to Measure the ROI of AI in Collections Without Increasing Compliance Risk 

Without stating the obvious, the question is not whether value and control should be measured together, but how an organisation can determine whether AI is genuinely improving its decisions rather than simply helping it to make the same decisions faster.

Starting narrower than the technology allows still makes sense, although the test should not be limited to the easiest customers or most predictable situations. One portfolio, decision or journey can be used to examine how the AI responds to incomplete or conflicting information, changes in customer circumstances, repeated contact attempts, vulnerability indicators, and requests close to a policy boundary. Straightforward cases may demonstrate capability, but the less obvious ones reveal whether the controls work.

It can also be useful to compare the AI-supported journey with the best practical alternative available today, rather than the average performance of the existing operation, while measuring the value of restraint alongside recovery and efficiency. How often does the AI correctly decide that no contact is needed, that there is insufficient evidence to recommend a treatment or that a customer should move into a different journey? Avoiding unnecessary contact, protecting likely self-curer customers, and directing skilled support towards customers who need it may create more value than simply increasing the automation rate.

Exceptions can provide equally valuable evidence. Challenged recommendations, abandoned journeys, escalations and unexpected customer responses can expose gaps in the data, policy, model, or process that headline measures may miss, while testing whether the business can reconstruct a decision, change a guardrail and return customers to a proven route provides a practical measure of reversibility before scale.

This is where governance becomes a genuine enabler, giving operations, risk, compliance, and technology teams the evidence needed to understand where AI creates value, where its boundaries sit and whether it can be expanded transparently and with confidence.

Summary

As we all know AI offers significant potential across collections, from risk prediction and contact strategy through to agent support, digital engagement and compliance review, but the challenge right now is not simply proving that the technology can work; it is demonstrating that it can operate reliably within the data, policy, customer-treatment and operational controls that already govern the collections process.

This requires governance to cover the complete use case rather than the model in isolation, connecting every prediction to a permitted decision, an appropriate treatment, and a measurable customer outcome. It also means recognising that different forms of AI carry different levels of risk, with assistive AI, decision support, and bounded automation each requiring controls that reflect what the technology is allowed to influence and when human support should take over.

Used in this way, governance is not a constraint on progress but the foundation for responsible adoption, giving organisations the confidence to explain how AI is being used, understand where its boundaries sit, learn from exceptions and demonstrate that it is improving decisions rather than simply increasing the speed or volume of activity.

Recommendations for consideration

Define the collections problem, the customer outcome being sought and the specific decision the AI will influence before determining what should be automated. This creates a clearer basis for ownership, measurement and control while reducing the risk of applying technology simply because the capability exists.

Specify what the AI can do independently, where approval is required and which circumstances should trigger escalation or human support, applying these boundaries across prioritisation, channel selection, contact timing, agent recommendations, offer presentation and workflow routing rather than limiting them to digital journeys.

Include incomplete or conflicting information, changes in customer circumstances, vulnerability indicators, repeated contact attempts, and requests close to a policy boundary within the test population. Straightforward cases can demonstrate capability, but less predictable situations provide stronger evidence that the controls will continue to work under pressure.

Compare the AI-supported journey with the best practical alternative available today and consider how often the technology correctly decides that no contact is required, that there is insufficient evidence to recommend a treatment or that the customer should move into a different journey. Avoiding unnecessary activity and directing skilled support towards the customers who need it can be more valuable than achieving a higher automation rate.

Capture challenged recommendations, escalations, abandoned journeys, broken arrangements, complaints, and unexpected customer responses so they can be traced back to the data, model, policy, workflow, or training. Before scaling, confirm that the business can reconstruct a decision, change a guardrail, pause the use case, and return customers to a proven route if performance or outcomes deteriorate.

Taken together, these recommendations provide a practical route from demonstration to production, allowing AI to be expanded where it creates measurable value while ensuring that the resulting decisions remain transparent, explainable, and appropriate for the customer.

Talk to an EXUS expert to explore how your collections operation can adopt AI with the right governance, controls and human oversight in place, turning promising use cases into safe, effective, production-ready capabilities.