Agentic KYB

Ask a model whether a business is real and it will answer. Proper nouns, near-duplicate company names, plausible street addresses, entity numbers with the right number of digits — this is the territory where a language model is most fluent and least reliable.

The failure is not that the model lies. It is that recall and retrieval look identical in the output. A fabricated registration number reads exactly like a real one, and so does a real one attached to the wrong company.

That makes the requirement narrower than “accuracy.” What a research run needs is provenance: for every fact in the answer, a source someone else could open.

Enigma serves resolved entities and their sources over MCP. The agent calls enigma.verify with a name; the response is a resolved entity plus the independent sources that corroborate it, at a stated tier.

The interesting part of that return is not the yes. It is the list underneath: which sources agreed, how many there were, and what kind of evidence each one is. That list is what the agent quotes, and what a reader can check afterwards.

It also fails visibly. When nothing corroborates the name, the agent gets an empty source list instead of a confident sentence — which is the one behaviour you cannot get from recall.

The decision gets looked at again: by a regulator, a fraud investigator, a chargeback dispute, or an acquirer's diligence, often years later and always by someone who was not there.

“The model concluded” is not a record. A cited answer is reviewable by that later reader, and it stays reviewable when the model has been replaced, the prompt has changed, and nobody remembers the run.

The working rule is simple enough to enforce in a review: let the agent do the reasoning and the writing. Never let it do the remembering.

Median response is around 1.5 seconds, which is what makes this more than a research tool. It is fast enough to sit inside an onboarding flow with no human in it, not just inside a chat session where a few seconds are invisible.

The same call serves a KYB pipeline clearing thousands of applications, an analyst working one business deeply in the console, and an agent chaining questions across a dozen entities. The answer does not change with the caller.

That is the quiet benefit. When the automated decision and the agent's write-up disagree, the disagreement is about interpretation, not about which version of the record each one happened to see.