India’s lending ecosystem has reached a scale where collections is increasingly a challenge of decision-making, intelligence and execution, rather than simply an operational one. The scale of technology-led debt collections is evident in the volumes handled by platforms such as Credgenics, which has managed more than 100 million retail loan accounts, touched over $250 billion in retail loan value for collections in FY24, and facilitated more than 1.7 billion omnichannel communications. At this scale, lenders no longer ask whether collections can be automated, but whether AI can intelligently decide what should happen next.
This is where agentic AI becomes particularly relevant. Collections is not a predictable, one-directional process. A borrower can promise to pay, ask for more time, dispute an amount, negotiate a settlement, decline to make a payment, become unresponsive or require human intervention. The outcome of one interaction can fundamentally change the next step. This makes collections an unusually strong environment for testing whether AI can move from simply responding to situations to coordinating decisions and actions across an evolving borrower journey.
Traditional automation has largely worked through predefined rules. A particular event triggers a particular action: a reminder is sent, a call is initiated, a payment link is shared or an account is escalated. Agentic AI can take this a step further by interpreting borrower intent and account context, planning the next permissible intervention, coordinating action across systems and monitoring the result within defined guardrails.
The difference goes beyond prediction and personalisation. Predictive models can estimate risk or identify the most effective channel, while generative AI can support contextual borrower conversations. Agentic AI adds the potential to connect those capabilities with execution, so that decisions, actions and subsequent assessments form a coordinated workflow rather than isolated activities.
Consider a borrower who misses an EMI but indicates an intention to pay after receiving their salary. An agentic system could evaluate the account context, assess the response, schedule an appropriate follow-up and initiate an approved payment journey. If the commitment is not fulfilled, the system could reassess the case and recommend a different intervention. If the borrower disputes the outstanding amount or reports financial distress, the workflow should change, with the case routed to the appropriate human team.
Collections also presents a particularly rigorous test for agentic AI because it brings together three characteristics: changing borrower circumstances, multiple interconnected decisions and measurable financial outcomes. Was the borrower reached? Did they engage? Did they commit to a payment? Was the promise fulfilled? Did the account get resolved? Was escalation required? Each of these outcomes provides a signal that can inform future decisions.
This creates the possibility of collections systems becoming increasingly adaptive. AI can analyse borrower behaviour and risk, identify the right communication channel, personalise the timing and content of outreach, and distinguish between cases that can be resolved digitally and those that require human intervention. Such capabilities are already becoming part of modern AI-led collections, including intent analysis, risk segmentation and channel optimisation. The agentic opportunity lies in orchestrating these capabilities and evaluating what each action achieves, rather than treating any one prediction or conversation as the endpoint.
For lenders, the potential value extends beyond automation alone. Better coordination could improve contactability, increase promise-to-pay fulfilment, reduce unnecessary follow-ups and help allocate human resources to cases that genuinely require intervention. Ultimately, these gains must be evaluated against recovery outcomes, collection costs and borrower experience and not just the number of tasks completed by AI.
But autonomy in financial services must come with accountability. Agentic AI cannot operate as an unchecked decision-maker. Collections involves sensitive financial information and significant customer interactions, making compliance, consent, auditability, data protection and clearly defined escalation boundaries essential. Actions involving contact timing and frequency, dispute handling, repayment terms or settlements must follow applicable regulations and lender policies, with approval controls where required. Every AI-initiated action should be traceable, and the system must know when to defer to a person.
This also changes the role of human collection teams. The future is unlikely to be about AI replacing people; it is more likely to be about AI handling scale and coordinating routine decisions while humans focus on disputes, complex negotiations, sensitive cases and situations requiring judgement or empathy.
That is why collections may be a defining test of agentic AI in BFSI. It is not an easy problem. It is a complex, high-volume environment where every borrower behaves differently and every decision can influence the outcome.
The real measure of success will therefore be whether AI can understand an interaction, determine a suitable next action, execute it within defined guardrails and evaluate the outcome through governed feedback mechanisms. If it can do that consistently, collections could become the blueprint for how agentic AI moves from an emerging technology into a practical, accountable capability across BFSI.
Views Expressed by: Anand Agrawal, Co-founder, Credgenics, Prix AI & FixMyScore








