Standards
Objective evidence: an AI answer has none
ISO 9000 defines objective evidence as data supporting the existence or verity of something, and audit standards rank it by reliability. An AI answer sits outside that hierarchy entirely: no artifact, no version, no approver, no way back.
Cognatum Team · Oct 14, 2026 · 5 min read
An auditor has one move, and it is always the same move. Show me.
Cognatum governs the entry
source · version · approver · permissions
A claim on its own counts for nothing. What settles a claim is evidence, and the quality world has a precise term for it. ISO 9000 defines objective evidence as data supporting the existence or verity of something. It can be obtained through observation, measurement, test or by other means.
Your staff now put audit questions to an AI assistant first. The answer comes back in seconds, in clean prose, and with nothing behind it that an auditor could take away.
What the standards mean by evidence
The ISO definition carries a second note, and it matters more than the headline. For the purpose of an audit, objective evidence generally consists of records, statements of fact, or other information that is relevant to the audit criteria and verifiable.
Verifiable is the load-bearing word. Someone other than you has to be able to go to the same place, look at the same thing, and land in the same spot.
How auditors actually gather it
ANAB, the US accreditation body, sets out how auditors gather it in practice: by observing processes, listening, interviewing personnel, and reviewing documents and records. Every one of those leaves a trail back to something fixed.
That is the whole point of the word objective. It does not mean neutral in tone. It means the thing stands up without you in the room.
Finance already wrote the hierarchy down
US audit standards go further and rank evidence by reliability. PCAOB standard AS 1105 says audit evidence is all the information used by the auditor in reaching the conclusions behind the opinion. To be appropriate, it has to be both relevant and reliable.
The reliability test is a short list, and one line on it lands squarely on AI answers. Original documents are more reliable than photocopies, or than documents that have been filmed, digitized or otherwise converted into electronic form. For converted material, the standard says reliability depends on the controls over the conversion.
A conversion nobody controls
An AI answer is a conversion. Text was chunked, embedded, retrieved and rewritten into a sentence. Ask who controls that conversion and the room goes quiet.
Inquiry is the weakest thing in the book
AS 1105 also covers inquiry, which is seeking information from knowledgeable people inside or outside the company. It attaches a blunt note. Inquiry of company personnel, by itself, does not provide sufficient audit evidence.
Asking an assistant is inquiry. It is inquiry of something that cannot be put under oath, cannot be trained, and cannot be held to account. It reads like a conclusion and it ranks below a photocopy.
Where the answer leaves the chain
Four things are missing, and not one of them is a model problem.
No artifact
An answer leaves nothing to file. Ask the same question twice and you get two wordings. There is no exhibit, so there is nothing to attach to a finding.
No source identity
Which document, which revision, which effective date. A file name is not a version, and a confident paraphrase is not a quotation.
No authorization
AS 1105 offers inspection of records for evidence of authorization as a test of controls. An answer carries no authorization at all. Nobody signed it.
No way back
Weeks later you cannot reconstruct what the answer rested on. The underlying material has moved, and nothing recorded where it stood on the day the question was asked.
The standards moved this year, and the answer will not say so
Here is the awkward part. The definition everybody quotes comes from ISO 9000:2015. ISO has now published ISO 9000:2026, which updates the definitions and terminology to line up with the revised ISO 9001. ISO 19011, the guidance on auditing management systems, has a 2026 edition too.
An assistant that read the 2015 text will answer from the 2015 text. It will sound exactly as sure of itself as it did last year. It is not wrong about what it read. It is wrong about what is current, and it has no way to tell you which.
Logs record the system, not the approval
The usual answer to all this is logging. The EU AI Act requires it. High-risk systems have to allow the automatic recording of events over the lifetime of the system, so that risks can be identified and operation can be monitored.
For remote biometric identification the law goes down to field level. It names the period of each use, the reference database the input was checked against, the input data that produced a match, and the people involved in verifying the result.
What those fields do not say
Read that list again. It is a record of what the system did. It says nothing about whether the material the system answered from was approved, current, or contradicted by another document down the hall.
Grounding the model does not close the gap either. The first preregistered study of commercial legal research tools found that the retrieval-based products from LexisNexis and Thomson Reuters hallucinated between 17 and 33 percent of the time, while their vendors described hallucination as eliminated.
What would make an answer worth showing
Not a better model. An entry behind the answer with the properties evidence has: a named approver, a date, a source, and a record that outlives the conversation.
That is the job of the Cognatum Knowledge Loop. AI does the finding, the gathering and the surfacing, across the systems the knowledge already lives in. Then it routes the item to a person, and Approve is where a human decides what is true.
Where two documents disagree, both go in front of that person with the disagreement named, rather than one being quietly picked. When a source changes, Cognatum notifies the people who depend on it. It does not rewrite the entry on its own.
A test you can run this week
Take a question your next audit will ask. Put it to the assistant your team already uses. Then ask for the approved entry behind the answer, the person who approved it, and the date.
If nobody can produce those three inside a minute, you did not have evidence. You had a sentence.
Your company's knowledge isn't missing. It's unusable. Cognatum changes that.
Sources
- Evaluating objective evidence for internal audits (anab.ansi.org)
- ISO glossary of terms: objective evidence (pillarmanagement.com)
- PCAOB AS 1105: Audit Evidence (pcaobus.org)
- ISO 9000:2015, quality management systems fundamentals and vocabulary (iso.org)
- ISO 9000:2026, the new edition of the vocabulary standard (iso.org)
- ISO 19011:2026, guidelines for auditing management systems (iso.org)
- EU AI Act Article 12, record-keeping (ai-act-service-desk.ec.europa.eu)
- Hallucination-Free? Assessing the reliability of leading AI legal research tools (arxiv.org)
- The Cognatum Knowledge Loop (cognatum.ai)
Common questions
Questions this raises.
What is objective evidence?
ISO 9000 defines objective evidence as data supporting the existence or verity of something, obtained through observation, measurement, test or other means. For the purpose of an audit it generally consists of records, statements of fact, or other information that is relevant to the audit criteria and verifiable. The test is not whether a statement sounds right. It is whether somebody else can go back to the same source and reach the same conclusion.
Is an AI-generated answer objective evidence?
No. An answer is not a record, it carries no approver and no date, and it cannot be reproduced word for word on request. PCAOB standard AS 1105 treats inquiry of company personnel, on its own, as insufficient audit evidence, and it ranks converted copies of documents below originals, with reliability resting on the controls over the conversion. An AI answer is both of those at once: an inquiry, answered from a converted copy.
Do AI audit trails and logs count as objective evidence?
They are evidence of what the system did, which is a different thing. EU AI Act Article 12 requires high-risk systems to record events automatically over their lifetime, and for remote biometric identification it specifies fields such as the period of use, the reference database and the input data that matched. None of those fields records whether the source material was approved, current, or contradicted by another document.
Does grounding or retrieval make an answer verifiable?
It narrows where the words came from without settling whether they were right. The first preregistered evaluation of commercial legal research tools found retrieval-based products from LexisNexis and Thomson Reuters hallucinating between 17 and 33 percent of the time, against vendor claims that hallucination had been eliminated. Grounding points at a document. It does not tell you the document was the approved current one.
What would an auditor need to see to accept an answer produced with AI help?
The entry behind it, and the entry's own paperwork: a named approver, an approval date, the source it came from, and a record of what that source said at the time the answer was given. If two sources conflicted, the auditor needs to see that the conflict was put to a person rather than resolved silently. That is governance of the knowledge, not of the model.