When AI drafts become
company knowledge

The copied paragraph

A project manager asks an AI assistant to summarise six customer interviews. The answer is fluent and plausible. One paragraph is copied into a project update, then into a product brief. A month later it appears in a steering paper as an established customer requirement.

Nobody intended to falsify the record. The original interviews are still available. Yet the organisation now treats an interpretation as a fact because the interpretation travelled farther than its evidence.

This is the point at which an AI quality problem becomes a management problem. The important event is not that a model produced text. It is that the organisation admitted the text into its working memory.

The HBR signal

Matthias Holweg and Thomas H. Davenport raised a related warning in their June 2026 Harvard Business Review article, “Don’t Let AI Slop Muck Up Your Company’s Processes”. Its accessible summary describes decay in the accuracy and quality of organisational knowledge as AI output enters company processes. ¹

That warning is useful, but “slop” can make the problem sound like a matter of bad writing or careless employees. The sharper management question is about status:

When does an AI draft become company knowledge?

An output gains institutional authority when it is placed somewhere people trust, repeated by people with standing, or used to justify action. A polished answer in a private chat is provisional. The same answer in a policy, customer record, knowledge base, audit paper or board pack can shape later decisions long after its origin has disappeared.

The control point is therefore admission, not generation.

Three ways a draft acquires authority

Trusted location

People infer reliability from where information appears. A paragraph in a team chat invites challenge. The same paragraph in an approved procedure looks settled. A CRM field, policy library or management dashboard confers status even when nobody has verified the underlying claim.

The system of record can turn an uncertain output into an apparently certain input.

Repetition

Repeated claims become easier to accept. AI makes repetition cheap. One generated summary can populate a presentation, a project plan, a customer note and a later prompt. Each appearance may look like independent confirmation even though all four descend from the same unverified answer.

The resulting problem is not merely duplication. It is false corroboration.

Action

Information acquires practical authority when somebody acts on it. A generated market summary used for background reading is different from the same summary used to change a price, reject a supplier or alter a safety procedure.

The closer an output sits to a consequential decision, the stronger the admission test should be.

Records rules are moving towards the same boundary

The latest official guidance makes this issue more concrete. On 21 August 2026, the US National Archives and Records Administration issued guidance explaining that AI inputs, outputs, data, audit trails and software can all fall within federal records requirements. The guidance does not make every prompt or answer a permanent record. It asks agencies to apply existing records law to the role the material plays. ²

Australian National Archives guidance draws an equally useful distinction. Minor AI drafts and reference-only outputs may be disposable. Outputs relied on for decisions, used as major drafts, or accepted as final business records may need to be retained with relevant prompts, inputs and metadata. Where unreviewed AI content is accepted as a final record, the guidance recommends marking its AI-generated and unconfirmed status. ³

These are public-sector rules, not a universal prescription for private companies. Their management logic travels well: significance depends less on who or what produced a document than on how the organisation uses it.

Provenance is necessary but not sufficient

Labelling an output “AI-generated” helps, but it does not answer the questions a later user needs.

They need to know:

• which sources grounded the claim;
• whether a human checked those sources;
• which parts are fact, interpretation, estimate or recommendation;
• who accepted the output for organisational use;
• where the output may be relied upon;
• when it should be reviewed or retired.

A model name and timestamp are technical provenance. Company knowledge also needs managerial provenance.

That means naming an owner who is willing and able to defend the admission decision. “Reviewed by a human” is too vague. The reviewer should have the competence and authority appropriate to the claim. A legal interpretation needs a different owner from a product description. A forecast needs a different standard from meeting notes.

Preserve the evidence behind the answer

NIST’s Generative AI Profile identifies confabulation and information-integrity risks across the AI lifecycle. A 2025 Nature Machine Intelligence study tested 24 language models on 13,000 questions and found material weaknesses in distinguishing belief, knowledge and fact. These findings do not mean that every AI answer is unreliable. They mean that fluency does not establish epistemic status. ^{4 5}

Organisations should therefore preserve access to the human-produced evidence behind consequential claims. The point is not to archive every chat. It is to stop a summary from replacing its sources.

Research on model collapse offers a useful but bounded analogy. A 2024 Nature paper found that models trained recursively on generated data can lose information about the original distribution. An organisation is not a language model, and copying AI text into a knowledge base is not model training. The transferable lesson is narrower: repeated derivation can obscure the original evidence and reduce variation. If managers cannot return to the source, later review becomes comparison between summaries rather than reconsideration of reality. ⁶

Create an admission boundary

Most organisations already have places where provisional material becomes authoritative: policy approval, CRM validation, document control, board-paper sign-off, release management and records capture. AI governance should use those boundaries rather than create a separate bureaucracy around every prompt.

Before generated content enters a trusted repository or supports a consequential action, require a short admission record:

On a narrow screen, each field is shown as a separate card.

This is not a demand for perfect certainty. It is a way to keep uncertainty attached to the information instead of losing it during reuse.

Do not turn every draft into a compliance object

An indiscriminate retention rule would create a second problem. Employees would fill repositories with minor prompts, abandoned drafts and duplicated output. Important evidence would become harder to find, while routine experimentation became slower.

The better threshold is use and consequence.

A disposable brainstorming list does not need the same control as a supplier assessment. A rewritten email does not need the same trace as a policy recommendation. A generated transcript used only as an aide-memoire is different from one filed as the official account of a meeting.

Managers should define a small number of admission classes and attach controls to them. For example:

On a narrow screen, each class is shown as a separate card.

The labels matter only if repositories, workflows and behaviour preserve them.

Ask vendors about knowledge admission

Enterprise AI products increasingly promise to draft, summarise and write directly into business systems. Buyers should look beyond answer quality and ask what happens next.

Can the product distinguish provisional output from approved knowledge? Does it preserve source links after content is copied into another system? Can a reviewer see which claims were generated, retrieved or supplied by a user? Can the organisation attach an owner, scope and review date? Can superseded content be found and retired? Can downstream users tell whether several apparently independent documents came from one generated source?

A product that generates excellent material but erases its status creates a hidden information-management cost.

The management decision

The common control “check AI output before use” is too weak because “use” covers everything from private thinking to a binding decision. The meaningful boundary is the moment an output acquires organisational authority.

At that boundary, the organisation should be able to answer three questions:

• What evidence supports this?
• Who accepted it for this use?
• When must it be reconsidered?

AI makes drafts abundant. It also makes it easier for a draft to travel farther than its evidence. The answer is not to distrust every generated sentence. It is to decide deliberately which sentences the company is prepared to remember.

The control point is admission, not generation.

Executive view

• An AI output becomes company knowledge through location, repetition or use in action, not because the model produced it.
• Trusted repositories can confer authority on content whose evidence and uncertainty have been stripped away.
• Technical provenance should be supplemented with a named owner, permitted use and review date.
• Controls should be proportionate. Working material does not need the same treatment as a decision record.
• The practical control point is the admission boundary between provisional output and organisational knowledge.