AI agents in lending and collections: what actually works in production

In 2026 nobody is debating whether a lender can put AI agents to work. The real question is which ones hold up in production and which ones never leave the demo. This year’s published evidence points in one direction — and it’s not the one in the pitch decks.

The agents that survive are short, with a human close by

The broadest study of real-world agents, Measuring Agents in Production (ICML 2026, practitioners across 26 domains), found that most production agents run ten steps or fewer before a human steps in, that nearly all use off-the-shelf commercial models without fine-tuning, and that they’re evaluated mainly through human review (arXiv 2512.04123).

For a lender, that means the agent that works isn’t the one that “approves loans on its own.” It’s the one that builds the file, reads the bank statements, flags what’s missing and hands committee a decision that’s ready to make.

Reliability isn’t keeping pace with capability

A Princeton team measured fifteen models on twelve reliability metrics — consistency, robustness, predictability and safety — and found that recent capability gains have only slightly improved reliability (arXiv 2602.16666).

In a regulated process, that matters more than any benchmark record. In lending, getting it right once isn’t enough: the agent has to get it right Monday and Friday, with the easy customer and the hard one. The metric worth demanding is consistency — solving the same task several times in a row, not “at least once” (τ-bench, arXiv 2406.12045).

The closest public case doesn’t sell autonomy. It sells method

Nubank published how it runs service agents across domains like card delivery, debt management and credit limits (arXiv 2606.08867). What’s interesting isn’t the bank’s size — it’s what they chose to show: A/B tests by domain, automated judges calibrated against humans, and measured results before scaling up. Not an agent deciding on its own.

What it means for your lending operation

1. Start where there’s the most paperwork and the least judgment. Origination and the digital file, payment reconciliation, collection reminders. That’s where an agent saves hours without making decisions that commit the institution.
2. Draw the line before you build. What the agent owns, what the system owns and what a person signs off. In collections, the agent sends reminders and logs the promise to pay; a person decides the restructuring.
3. Evaluate on your own cases. A hundred real conversations or files, graded pass/fail by someone who knows the process, beat any published score.
4. Measure by domain, not overall. An agent can be excellent at reading bank statements and mediocre at negotiating. Scale where it’s proven.

The bottom line

What the market pays for in 2026 is the “boring” part: measured reliability, continuous evaluation, scoped permissions and an accountable human. For a regulated lender, that boring part is exactly what your regulator will ask to see.

Want to know which part of your operation to tackle first? Tell Luz how your team works today.

The rules that apply: AI collections and CONDUSEF rules and AI credit origination.