Legal hold preserves documents, not answers

A legal hold arrives and the organization moves quickly. Counsel identifies custodians, suspends routine deletion, sends the notice and collects the acknowledgments. Mailboxes, drives and collaboration sites are frozen in place. The documents stop moving.

The AI assistant does not stop. It has been answering questions drawn from those same documents for eighteen months, and it keeps answering while the hold is in force. Nobody has asked what it told people, which version of which document it drew from, or whether the person who acted on an answer could reproduce it today.

What a legal hold actually freezes

Rule 37(e) of the Federal Rules of Civil Procedure deals with electronically stored information that should have been preserved in the anticipation or conduct of litigation and is lost because a party failed to take reasonable steps to preserve it. The committee note is explicit that the rule rests on the common law duty to preserve and does not attempt to create a new one. The object of that duty is information, and in practice what an organization freezes is repositories and custodians.

What the hold tooling does

The tooling follows the same shape. Microsoft Purview creates eDiscovery holds across mailboxes, SharePoint sites, OneDrive accounts and Teams content, preserving items in place so they survive deletion by a user. Every product in the category does a version of the same thing. It works out where relevant material sits, then stops that material from changing or disappearing.

The answer is not the record

An AI assistant does not hand back a document. It assembles a response from several sources at a particular moment, under a particular retrieval configuration, using whichever version of each source was indexed that day. Preserve every one of those source documents perfectly and you still cannot reconstruct the response. The inputs survive. The act of answering does not.

Three things a preserved document will not tell you

Which version was retrieved. A document preserved today is the document as it stands today. A hold placed in March says nothing about what the assistant read in January.

Who had approved it, and when. Preservation captures the file. It does not capture whether a named person had signed off on the content, or whether that sign off was still valid on the day the answer went out.

Whether a conflicting source existed at the same moment. Two documents can sit in the preserved set saying opposite things. The preserved set does not record which one the assistant used, or that a disagreement existed at all.

Regulators have started asking for the operating record

Article 12 of the EU AI Act requires high risk AI systems to technically allow for the automatic recording of events over the lifetime of the system. That is a requirement about the record of operation, not about the underlying documents. FINRA made a related move in Regulatory Notice 24-09, reminding member firms that its rules are technology neutral and continue to apply when generative tools are used, and that supervisory procedures should address technology governance, model risk management, data privacy and integrity, and the reliability and accuracy of the model.

Preserved and reliable are different questions

A preserved corpus is not the same as a correct answer. The first preregistered evaluation of commercial AI legal research tools found that products from LexisNexis and Thomson Reuters hallucinated between 17 and 33 percent of the time, despite vendor claims of eliminating hallucinations. Those tools retrieve from curated, well maintained legal databases. Grounding an answer in preserved material does not make the answer right, and it does not make the answer reproducible.

The same gap shows up at the other end

Disposal runs in the opposite direction and has the identical blind spot. ISO 15489-1 sets out records management as a lifecycle in which records are created, captured and eventually disposed of under a schedule. When a record is disposed of on time and by the book, the source disappears. Any answer already derived from it stays in circulation, cited to something that no longer exists. Preservation and disposal are both policies about documents. Neither of them governs the answers built on top of those documents.

Point in time, not just point of storage

This is not a storage problem. Organizations are good at keeping things. The gap is that nothing sits above the repositories recording what the organization knew, and had approved, at the moment a question was answered.

Cognatum is built to be that layer. Information stays where it lives, in the departmental repositories teams already use, and Cognatum works bidirectionally with those systems rather than asking anyone to migrate. Every entry it serves carries a named approver, a date and a source. It also keeps track of where the knowledge was when the answer was given, which is provenance at a point in time rather than a pointer to a file that has since moved on.

What happens when a source changes

When a source document, a table or a stored procedure changes, Cognatum notifies the people who own the entries that depend on it. It does not quietly rewrite them, and it is not self correcting. Where two sources disagree, it presents both and shows where the disagreement sits rather than picking a winner. A named human approves, and that approval is the gate the whole loop turns on.

What to do before the next hold lands

Ask one narrow question inside your own organization. If a regulator or opposing counsel asked what your AI assistant told an employee on a given day last quarter, what would you produce?

If the answer is a folder of documents, the honest position is that you preserved the inputs and cannot show the output. That may well be defensible. It is better to work it out now than to discover it under a preservation order.

The unglamorous part

The practical work is unglamorous. Know which repositories your assistant reads. Know who approved the content in them and when. Know what happens to an answer once its source is superseded. None of this requires consolidating your systems, and consolidation was never the fix.

Your company's knowledge isn't missing. It's unusable. Cognatum changes that. The Cognatum Knowledge Loop sets out the eight steps and the single human gate in full.