Cognatum

Financial services

MNPI rules follow the person, not the answer

An information barrier is keyed to a named person and a named document. A generated answer is neither, so the one record that makes assembled research defensible, the trail of what was read, never gets written.

Cognatum Team · Nov 12, 2026 · 6 min read

An analyst asks the assistant a short question about a company her firm covers. Four sentences come back. They read like analysis.

Served to one approved entry
AI assistants & agents
Proposal tools
Internal search & chat
Customer portals
Compliance & audit

Cognatum governs the entry

source · version · approver · permissions

The answer is good. It is also assembled. Behind it sit a research note, a slide from a deal folder, a credit memo and an old email thread.

She cannot see which of those it read. Neither can anyone else, now or a year from now.

Her firm has spent decades building a wall to keep certain documents away from certain people. The wall was built around people and around documents. An answer is neither of those things.

What makes information nonpublic

Material nonpublic information is the pairing of two tests. Information is material if a reasonable investor would want it before making a decision. It is nonpublic until it has been made known to the market at large.

Both halves have to hold. Firms do not leave that to judgment alone. They build an information barrier, still widely called a firewall, and they staff it.

What the barrier controls

The published description of a working firewall is a list of controls over people and paper. Interdepartmental messages route through a clearance area in compliance or legal.

Employee trading is checked against watch, restricted and rumor lists. The procedures that limit the flow between departments are documented, and so are the actions taken to enforce them.

One rule underpins the rest. In a firm with several businesses, a person should sit on only one side of the wall at a time. Everything is keyed to a named individual and a named document.

Nonpublic is a question about time

The second test is not a property of a document. It is a property of a moment. Information is nonpublic until it has been disseminated, and disseminated means made known.

The guidance is careful about this. A firm needs a reasonable expectation that people have received the information. A room full of analysts is not the market. The passage of time can also make once important information immaterial.

The window nobody records

So the real question is never just what the document said. It is what the document said at the hour somebody acted on it, and whether it had reached the market by then.

Answer that for a document and you look at a version history. Answer it for a generated reply and you have nothing to look at. The reply does not record which version of which file it drew on, or when.

The mosaic theory runs on receipts

There is a long standing principle that protects honest research. An analyst may piece together public facts and nonmaterial private scraps, reach a conclusion that would have been inside information if somebody had simply told her, and act on it.

That is the mosaic theory, and it is what makes a research desk lawful rather than reckless. But it carries a condition that is easy to skip past.

Save and document all your research

The guidance says analysts should save and document all their research. Evidence of what the analyst knew, public and nonmaterial, strengthens the claim that the conclusion came from proper methods rather than from inside knowledge.

Read that again as a records requirement. The defense is not the reasoning. The defense is the trail of inputs behind the reasoning.

A mosaic with no bibliography

A retrieval assistant builds a mosaic for a living. It gathers from many places, weighs what it finds and returns one confident paragraph. That is the same move an analyst makes.

What it does not do is keep the receipt. The output arrives without the list of what went into it, which removes the one artifact that would have made the method defensible.

The regulator already described the use case

This is not a hypothetical. A 2024 FINRA notice on generative AI lists the uses it expects, and the first ones named are internal.

The notice says such tools may let an associated person easily locate and query a firm's policies, procedures or forms, and generate summaries derived from the firm's research reports. Research reports are precisely what the wall is built around.

Reasonably designed, against what

The underlying statute is short. Every registered adviser must establish, maintain and enforce written policies reasonably designed to prevent the misuse of material nonpublic information by the adviser or anyone associated with it.

Reasonably designed is measured against how the firm actually works. A control written for documents and desks is not obviously reasonable against a tool that reads across all of them and leaves no record of what it touched.

What the answer would have to carry

None of this argues for shutting the assistant off. It argues that the answer needs the same furniture the document already has.

A named approver. A date. A source. Those three fields turn a reply into something a reviewer can examine instead of a claim that has to be taken on trust.

Where the knowledge was at the time

The harder field is the one the nonpublic test demands. A record of where the knowledge stood when the answer was given, not where it stands today.

The security standards bodies have a word for this. Provenance is defined as the chronology of the origin, development, ownership, location and changes to a system or its data. That chronology is what lets anyone reconstruct the moment.

Two memos, one of them cleared

Conflict is the other gap. Research on large language models names inter context conflict, the case where two retrieved sources disagree, as a distinct failure mode.

In a bank that case is routine. One memo was cleared for release and a second was not. A system that silently picks a winner has made a disclosure decision nobody signed off on.

The wall is not the weak point

The firewall was a good piece of engineering and it still works on what it was designed for. The gap is that it governs people and files, and the thing now circulating is a derived answer that is neither.

Still wrong, with a source attached

Grounding an assistant in internal material does not close that gap. A preregistered study of commercial legal research tools found that systems built on retrieval still produced false information between 17 and 33 percent of the time, against vendor claims of eliminating it.

So the work is unglamorous and it sits upstream. Approve the knowledge once, with a person's name against it. Record what was true when the answer went out. Show the disagreement instead of resolving it quietly. Your company's knowledge isn't missing. It's unusable. Cognatum changes that.

Common questions

Questions this raises.

What is material nonpublic information?

Material nonpublic information is information that a reasonable investor would want before making an investment decision and that has not yet been made known to the market in general. Both tests have to be met. Once the information has been disseminated widely enough that people can reasonably be expected to have received it, it is public and the restriction lifts.

Do information barriers cover an AI assistant?

A barrier controls which people and which systems can reach which documents, so access rules can be applied to an assistant like any other reader. What a barrier does not cover is the answer the assistant produces. That answer is a new piece of writing, held by whoever asked, carrying no classification, no approver and no record of the material behind it.

Why does the mosaic theory matter here?

The mosaic theory lets an analyst combine public facts and nonmaterial private details into a conclusion and act on it. The accepted way to show the conclusion was reached properly is to keep and document the research behind it. A retrieval assistant performs the same assembly but keeps no comparable record, so the usual evidence is missing.

Does grounding an assistant in internal documents solve the problem?

Grounding narrows where an answer comes from, which helps. It does not establish that the source was approved, current, or consistent with the next document along. A preregistered study of commercial legal research tools built on retrieval found false information in 17 to 33 percent of cases despite vendor claims to the contrary.

What record would make an AI answer reviewable?

Four things. A named person who approved the underlying knowledge, the date of that approval, the source it came from, and a point in time record of what that source said when the answer was given. With those, a reviewer can reconstruct the moment. Without them, the answer can only be trusted or distrusted.

Knowledge governed. Intelligence everywhere.

See it on your own content, in your own environment.