The Validator Operating Model
In an audit, the work is done by a team of people in their twenties. The value is a partner's signature.
Nobody thinks that signature means the partner recreated the work. It means someone with judgment, standing, and something to lose has decided to stand behind it. That is what makes the output usable by a regulator, a board, an investor.
Last month I wrote that trust, not capability, is the binding constraint on AI in financial services. This is the second of the three pillars, and it is the one that decides whether anything reaches production.
When machines produce the work, production stops being scarce. The signature becomes the scarce thing.
Most firms are aiming their existing control model at this problem. Maker, checker, supervisor: the same task performed up to three times by people of increasing seniority. It was built for one failure mode and is now pointed at two others. Human error is random and tracks fatigue and volume. Code fails systematically, the same error in the same place every time. A language model does neither. Ask it the same question twice and you can get two different answers, one of which may be wrong for reasons you cannot inspect.
What replaces the pyramid is not a better pyramid. It is a different shape.
1. Fan out to makers that are genuinely different. Not one method run several times. One path is deterministic code, written by the model and tested once, wherever the answer is computable. Another is a model working the problem its own way. A third is a hybrid, where the model decomposes and the code executes. Arithmetic and reconciliation should almost never be a probabilistic step.
2. Check each output three ways. One checker examines the route and the sources. One asks whether the answer is plausible on its face. One reconciles it against an independent record. Three questions, three classes of error, and only the last one touches ground truth.
3. Bring them together and examine them against each other. Where the paths agree, and where they do not. Then top-down, does the whole hold together, and bottom-up, do the parts trace and sum. A detail checker cannot see incoherence in the aggregate. An aggregate check cannot see the one wrong line that nets out.
4. Score every entry. Not the model's opinion of itself, which is fluent and badly calibrated in the same direction. A number computed from structure: how many independent paths agreed, whether the deterministic path matched, whether the checkers disagreed with each other, how far the answer sits from what this entry type normally looks like.
5. Let the score route the human. High confidence is sampled. The middle is reviewed, with the context assembled in advance so the reviewer arrives at a decision rather than a research project. Low confidence escalates. Running the whole graph again is cheap enough now to be routine, but an item that will not converge does not need more compute. It needs a person.
6. Record the process, not just the answer. Asking a model to explain itself returns generated text about its reasoning, subject to the same failure as the answer, and a fluent explanation of a wrong answer is the most dangerous artifact this system can produce. The evidence has to be captured as the work happens: every output linked to its sources, its route, and every point where the layers disagreed. Then the validation is reviewable, not just the output, which is what lets the confidence scores be calibrated against what turned out to be true.
Three things decide whether that is an architecture or a diagram.
Independence has to be engineered, then verified. Makers that share a model, or read the same pre-processed input, are one maker wearing three coats, and their agreement is correlation dressed as confirmation. Aviation learned this with redundant flight computers built from a shared specification.
Confidence has to be earned. A score nobody calibrates against outcomes is decoration, and it sits directly upstream of where human attention gets spent.
Someone signs. The model does not. If no named person is accountable for a given output, the firm has not put AI into production. It has run an experiment using client assets.
None of this makes the output certain. That is the point. It makes the output inspectable, and it concentrates human judgment on the few places where independent methods could not agree.
The uncomfortable part is worth saying plainly. In this model far fewer people produce and far more people validate, and validation demands the judgment that used to be earned by producing for a decade. The apprenticeship that manufactured senior judgment is exactly the work being automated first. Firms that solve for throughput and ignore this will find, in about five years, that they have no one left who is qualified to sign.
That is the third pillar, and it is where I think most transformations will actually fail. It is next.
The work is getting cheaper. Standing behind it is not.