A Four-Stage Loop, Not a Pipeline
Rather than treating AI as a single feature, a mature NPL operating model embeds it across a continuous cycle:
Decide. Portfolio and core-systems data feeds a decision engine that handles AI-driven segmentation, prioritization, strategy definition, and early-warning indicators — structured and governed, not opaque.
Execute. Decisions flow into loan management and work-out systems, then out across omni-channel communication: automated dialers, AI-supported human agents, virtual agents, SMS/Viber, and self-service portals.
Engage. Customer interactions run through a unified contact center where AI supports human agents rather than replacing them — predicting best-time-to-call, powering QA review, and handling routine conversations while complex cases stay with people.
Learn. Outcomes — cure rates, recovery, response, time-to-resolution — get analyzed by segment to see what’s working and what’s stalling, then feed straight back into the decision engine.
The key detail: outcomes continuously loop back into decisioning, so the model keeps refining itself rather than running static rules indefinitely.
Human in the Loop, No Black Box
It would be easy to let AI make every decision and have humans just execute. That’s the wrong model for NPL, given the regulatory scrutiny and human stakes involved. A better philosophy: AI, human in the loop, no black box.
In practice: AI-powered QA reviews 100% of interactions — not a sample — but findings go to human risk and quality teams. Business analysts, not IT, own and edit decision engine rules directly, with no release cycle in the way. And the daily feedback loop — execute, measure, analyze, adapt — closes within a single day, so yesterday’s outcomes shape today’s strategy immediately.
This is a meaningfully different posture than “autonomous AI.” Accountability for the rules stays with the institution’s own analysts.
Removing SQL as a Barrier
A quieter but equally practical application is conversational analytics. NPL analytics teams often sit on rich data that only a few SQL-literate analysts can query, leaving everyone else in a request queue.
A conversational layer lets business users ask questions in natural language and get back governed, schema-aware SQL, data, charts, or dashboards — with transparent SQL for auditability and persisted outputs for traceability. Governed output, not just generated text.
The Underlying Principle
None of this works as disconnected AI features bolted onto existing tools. The value comes from treating AI as infrastructure running through decisioning, execution, engagement, and learning simultaneously — with governance and human ownership built in at every stage.
The organizations getting this right aren’t asking “where can we add an AI feature?” They’re asking how to make their entire recovery process learn and adapt daily, while their own people stay in control of the strategy.
This article is based on a presentation delivered at the NPL Advisory Panel, July 2026.
The NPL Advisory Panel is an informal expert group set up by the European Commission’s Directorate-General for Financial Stability, Financial Services and Capital Markets Union (DG FISMA) to provide advice and expertise on non-performing loans across the EU.
Relational is a proud member of the Panel.
