Industries / Financial services
Forty agent pilots, and nothing in production that touches a real decision.
For heads of AI, chief risk officers and model risk teams at banks, NBFCs and insurers building AI decisioning in lending, underwriting and collections — where the blocker on autonomy was never capability. It was accountability.
The model risk committee
The demo was excellent. The agent read the portfolio, found the delinquency spike, and wrote a paragraph explaining it that was clear, specific and confident.
It was also wrong, and there was no way to tell that from the output. The wrong answer and the right answer are rendered in the same font, with the same certainty, at the same speed.
So the committee asked the only question that matters: when it is wrong, how will we know, and what will it have done by then? You did not have an answer. The pilot is still a pilot.
Why this one is different
In lending, a confidently wrong answer isn't an embarrassment. It's a provision.
Silent
A hallucinated root cause produces a mistargeted intervention. It fails quietly, over a quarter, and looks like bad luck.
Written
Your supervisor will ask for the reasoning behind an automated action. A prompt log is not reasoning. It's a transcript.
Compounding
An agent that misreads sourcing quality as a collections problem keeps buying the bad cohort. The error funds itself.
The industry response has been to keep a human in the loop on everything, which means you have not built an agent — you have built an expensive suggestion box, and you are paying for both the model and the reviewer.
What Hetu does about it
The model is never allowed to compute. It only narrates what was proved.
It tells you before it's wrong
“How do we know when it's wrong?”
Every conclusion carries a label: confirmed, high, medium — two candidates, or hypothesis. The label is computed from explained variance and interval overlap, not from the model's self-report.
It refuses
“What if the evidence is thin?”
Below the sample-size floor, the conclusion is withheld and the thin metric is named. A regulatory or macro event overlapping the window suppresses attribution entirely.
Structurally, no
“Can it hallucinate a number?”
Figures come from a deterministic rule engine or a fitted causal model. The narration call is separate, receives only the verified object, and is rejected if its output contains a number or a causal claim not in the payload.
Only what it has proved
“What can it do unsupervised?”
The autonomy envelope starts closed. Decision types move inside as measured outcomes demonstrate accuracy, and fall back out when they don't. Irreversible, fat-tailed actions never auto-execute.
One immutable file
“What do we show the regulator?”
Every decision generated, approved, deferred, executed and measured — who, when, what, why, on what evidence, and what happened. Refusals are logged as first-class records, not as gaps.
Your experts seed it
“Do we have to rebuild the graph?”
The causal graph and the WHY traversals are configuration, hand-seeded by the people who already know your failure patterns. Adding one is an insert, not a deployment.
Running today
An NBFC origination pipeline, from field sourcing to 90 days past due.
Not a benchmark. A production deployment where the target variable is deliberately the hard one: not a disbursed loan, but a loan still performing at 90 DPD. Anything less lets a system claim success for originating bad credit quickly.
The graph extends past the sale.
Sourcing → underwriting → disbursement → repayment. That is the only way “fewer loans disbursing” and “more disbursed loans going bad” can be told apart — they look identical in a throughput chart and demand opposite interventions.
Confounders are declared, not discovered.
Policy changes and macro events sit in a maintained log. When one overlaps an anomaly window, the system says so instead of attributing around it.
Accuracy is validated on a labelled holdout,
stratified by confidence label, against root causes a human analyst determined by file review.
The graph
It extends past the sale, because that is where the two readings separate.
A throughput chart cannot distinguish fewer loans disbursing from more disbursed loans going bad. They demand opposite interventions.
Target
Not a disbursed loan. A loan still performing at 90 days past due — anything less rewards originating bad credit quickly.
Divergence
One anomaly, two candidate readings. Tier 1 separates them on vintage and lag before anything is attributed.
Lag
Sourcing quality shows up in repayment sixty to ninety days later, so the temporal check does real work here.
Performing rate −14%, one zone
Sourcing quality at two field officers, 60 days prior
Confirmed
Throughput down, applications up
Early delinquency dominant · 44%, CI 37–51
High
Back-book DPD rising, policy change same window
Attribution suppressed. Two hypotheses, each with a falsification test.
Refused
The third row is the one to take to your committee. A system that will say “I cannot attribute this, and here is what would settle it” is the only kind that can be trusted to act on the rows where it does not say that.
Deployment
It runs where your data already is.
Design partners
Bring the decision your committee refused to let an agent make.
We start with one: a decision your team already makes manually, where you can articulate the failure patterns and where being wrong is expensive. We encode the constraint framework, seed the graph with your experts, and show you the first refusal.