I haven’t posted here for a while, but I’ve been thinking about where agentic AI would genuinely add value within collections.
A lot can already be achieved with decisioning, workflow, self-service and guidance for collections staff. So if you’re considering adding agentic functionality, what problem is it solving that those capabilities don’t already solve?
One scenario that comes to mind is where work spans multiple systems or stages, and something needs to pull together the context, coordinate the next steps and keep the case moving. But I also suspect some use cases currently being described as “agentic” could be handled pretty well using existing automation methods.
So I’m interested in how others are drawing that distinction. What’s the first collections use case you’d choose for an agent, and why wouldn’t you solve it with standard decisioning or workflow?
While its not agentic AI, Voice AI has tremendous value in the collections space. Regulators require conversation tracking to ensure the correct disclosures are shared on the call. Therefore Voice AI can serve as a verification and audit tool during the call. I’ll continue to think about agentic AI use cases in the space.
I think the distinction comes down to deterministic vs. non-deterministic work.
If a process can be fully defined through business rules, decisioning, workflow orchestration, and integrations, I’d generally avoid introducing an agent. Traditional automation is often more predictable, explainable, and easier to govern.
Where I see Agentic AI delivering genuine value is when the work involves research, reasoning, and synthesizing information across multiple systems and large volumes of unstructured data—activities that are difficult to model using explicit rules alone.
Some examples include:
Identifying related or duplicate fraud and collections cases using semantic similarity rather than exact matching. An agent can analyze collector notes, customer narratives, emails, documents, disputes, and historical cases to uncover connections and patterns that traditional matching approaches may miss.
Researching customer context across multiple systems and assembling a consolidated view before recommending the next best action. This becomes particularly valuable when information is fragmented across applications, knowledge repositories, and case histories.
Dispute, claim, or hardship assessment, where the agent must evaluate case data, policies, procedures, prior outcomes, and supporting evidence before recommending a course of action.
That said, I don’t see Agentic AI replacing workflow or decisioning—quite the opposite. Workflow remains the governance and orchestration layer, while the agent augments the process by performing the research, analysis, and reasoning that would otherwise require significant manual effort from case workers.
In many organizations, investigators and collection specialists spend hours—or even days—reviewing multiple systems, documents, notes, and knowledge sources to validate information and determine the appropriate next steps. Agentic AI can significantly reduce that effort by bringing together relevant context, identifying meaningful insights, and presenting explainable recommendations. The human remains in control of the decision, but the time spent on manual research and data gathering is dramatically reduced.
Thanks @Ramesh. I agree, although I think there is another distinction here….
Using AI to match similar cases, summarise case histories or suggest a next step doesn’t necessarily make it agentic…. All of that could still happen within an existing workflow.
I feel it becomes agentic when the AI can work out what it needs to look at next, change direction as it learns more, and then do something with the result.
So how much freedom does it really need? In your examples, would you let the agent make the recommendation, or use it to pull the evidence together and leave decisioning to do the rest?
Thanks @Aaron. I agree… Voice AI could be huge in collections, particularly when it’s combined with CDH during the call, (not just to check compliance but) to help decide what should happen next.
Yep… disclosures still need to be fixed and controlled. But Voice AI could also spot that something’s been missed, flag possible signs of vulnerability / confusion, and bring what the customer has actually said in to the case.
If it can then prompt / trigger the right next step while the call is still happening, perhaps that is where it starts to become agentic rather than simply Voice AI.
So… what would be the first in-call action you’d trust it to take?
I think you are spot on in the sense that the core collections strategy work is primarily a decisioning use case. It allows for transitioning from relatively basic time-based collection strategies (send everyone this email after 4 weeks in arrears) to strategies that better balance customer experience, hardship, customer value and collections exposure and likelihood to repay, and not just be better at identifying who to target and how to target.
The agentic and GenAI use case come around the corner when we are in more ‘high touch’ type situations. This is not just when engaging a customer over the phone, it could also be some self-service service a customer can use to find better solutions for their problem, or creating a full picture when a customer is entering into arrears or has made it past early arrears. All kinds of research agent and document analysis type use cases could be relevant here to create a full picture and provide (candidate) solutions.
