AI governance
Your AI answer has no chain of custody
A forensics team can name every person who held a piece of evidence. Your assistant cannot name anyone who held the knowledge behind this morning's answer. Custody is solved for samples, timber and build artifacts. Nobody wrote a form for knowledge.
Cognatum Team · Nov 2, 2026 · 6 min read
A forensics team can tell you who held a piece of evidence at every point in its life. A name, a date, a time, a reason for each move, and a signature at each handoff. If one link is missing, a court can refuse the evidence. Not because anyone showed it was tampered with. Because nobody can show it was not.
Cognatum governs the entry
source · version · approver · permissions
Your assistant answered a policy question this morning. Ask who held that knowledge between the approved document and the words on the screen. There is no list. There was never a form.
What a chain of custody asks for
NIST defines it plainly. It is a process that tracks the movement of evidence through its collection, safeguarding and analysis lifecycle. It does that by documenting each person who handled the evidence, the date and time it was collected or transferred, and the purpose for the transfer.
The legal description is stricter still. An identifiable person must always have custody of the item. The record covers the conditions it was gathered under, the identity of every handler, how long each one held it, how it was stored, and how it passed to the next custodian. Signatures at every step.
Three questions and a signature
Who held it. When. Why it moved. Then a name on the line. That is the whole instrument. It is not a label on a box. It is a list of names in order, with no gaps.
The test is authentication, not storage
Federal Rule of Evidence 901 sets the bar. To authenticate an item, the party offering it must produce evidence sufficient to support a finding that the item is what they claim it is.
Read that against an AI answer. The answer claims to be your company's position on a question. What would you produce to support that finding?
Rule 901 lists ten examples. One is worth dwelling on. Example nine covers evidence about a process or system, and it asks for evidence describing the process and showing that it produces an accurate result.
Describing the system is the easy half
Any vendor can describe the pipeline. Showing it produces an accurate result is a claim about the material going in, not about the machinery. Nobody has signed for the material.
Retrieval has handlers, not custodians
Follow one answer backwards. A connector copies a document out of a repository. A splitter cuts it into chunks. An embedding model turns each chunk into numbers. A ranker picks a handful by similarity. An assembler drops them into a prompt. A model writes prose around them.
Six handlers. Each one changed the material. None of them signed for it.
A handler who leaves no record is the exact thing the doctrine was built to catch. In a lab it would end the case. In a retrieval pipeline it is the normal operating state, and it ships to every desk in the building.
The working copy problem
Digital forensics has a rule about copies. You do not work on the original. You preserve it as a sealed master and you analyse a duplicate, so the master can always settle a dispute about what changed.
Retrieval works on copies by design. A chunk is a copy. A vector is a copy of a copy. The difference is that there is no sealed master on the other side of the comparison. Nothing records which version of the document the chunk came from, or whether that version is still in force.
What the telemetry standard writes down
The OpenTelemetry attribute registry is the common vocabulary for AI telemetry. Its field for retrieved documents says each document object should carry at least two properties. An identifier, and a relevance score.
Two fields. A custody form asks for six. There is no handler, no version, no approver, no effective date, and no place to put one.
So the gap is not that teams forget to read the log. You cannot read a transfer log that was never written.
Regulation writes the field once
The EU AI Act comes close in one place. Article 12 requires high risk systems to allow automatic recording of events across their lifetime. For one Annex III use case, remote biometric identification, it then names the minimum fields.
Among them is the reference database against which input data has been checked, and the identity of the people who verified the result. That is a custody field and a custodian, written into AI law.
Written once, scoped narrowly
It reaches one narrow use case. Note also that the Commission's own page carries a disclaimer, because the published text reflects amendments made by the Digital Omnibus on AI and is not yet fully updated. An assistant answering from your procedures sits outside that scope either way.
Custody is solved for everything except knowledge
ISO published a general standard for this in 2020. ISO 22095 covers chain of custody terminology and models, with an amendment issued in 2026 and a revision now in development. Supply chains have models for proving an unbroken line.
Software has them too. Supply chain security practice extends custody to build artifacts, scan results and access logs, and asks for acknowledgment from both the sending and the receiving party at every handoff.
Timber has it. Drug samples have it. Container images have it. The institutional knowledge your teams answer customers from does not.
What a custody record for knowledge holds
It is not complicated, and it is not a log file bolted on afterwards. Four fields travel with every entry in the knowledge base.
The entry itself. The version. The named person who approved it. The date they did. Then one more thing a log cannot reconstruct, which is where the knowledge stood at the moment the answer was given.
Where the human signature goes
The Cognatum Knowledge Loop is eight steps in four phases. AI runs seven of them. It finds material, structures it, cleans it, enriches it and serves it. It does not decide what is true.
Step seven hands the item to a person. They can rework the wording before they approve it, and their name stays on it. Where two documents disagree, both are surfaced with the point of conflict named, rather than one being quietly chosen.
When a source changes, the system tells you what now looks out of date. It does not rewrite your approved answer on your behalf, and it is not a claim that everything is current. It is a notification with a name attached.
Three numbers a reviewer can read
Duplicates. Conflicts. Gaps. Those are countable, and a count is what a reviewer asks for when they want to know whether the knowledge behind an answer can be trusted.
A test you can run this week
Pick a procedure your team replaced this year. Ask the assistant the question that procedure answers. Then ask it for the entry, the version, the approver and the date.
If it cannot produce those four, you do not have an accuracy problem to debug. You have a custody gap, and the answer already went out. Your company's knowledge isn't missing. It's unusable. Cognatum changes that.
Sources
- Chain of custody (en.wikipedia.org)
- Chain of custody, CSRC Glossary (csrc.nist.gov)
- Federal Rule of Evidence 901, Authenticating or Identifying Evidence (law.cornell.edu)
- ISO 22095:2020, Chain of custody, general terminology and models (iso.org)
- Digital Forensics and the Chain of Custody (online.champlain.edu)
- Chain of Custody in software supply chain security (kusari.dev)
- GenAI attribute registry, semantic conventions (opentelemetry.io)
- EU AI Act Article 12, Record-keeping (ai-act-service-desk.ec.europa.eu)
- The Cognatum Knowledge Loop (cognatum.ai)
Common questions
Questions this raises.
What is a chain of custody for an AI answer?
It is a record of every entry the answer drew on, each one carrying its version, the named person who approved it and the date. It also records where that knowledge stood at the moment the answer was given, so the answer can be reconstructed later rather than guessed at.
Isn't a citation enough?
A citation names a source. It does not say which version of that source was read, whether anyone approved it, or whether it had been replaced by the time the answer went out. A link to a current page is not evidence about the page as it stood last March.
Does logging the retrieved documents solve it?
Not on its own. The common telemetry vocabulary records a document identifier and a relevance score. Those are facts about the retrieval, not about the document's approval, version or effective date, so the log cannot answer a custody question that was never captured.
Does this make us compliant with the EU AI Act or the NIST AI Risk Management Framework?
No. No software confers compliance with either. What governed knowledge does is supply the evidence a reviewer asks for, including named approvers, versions, dates and point in time provenance. The compliance judgment stays with your organization.
Does Cognatum keep our documents up to date automatically?
No, and that distinction matters. It watches the sources you point it at, including structured sources, and notifies you when something tied to them looks out of date. A person decides what the new answer should say and approves it.