Experts fix their documents when the AI gets an answer wrong citing them. Only if they find out

A user flagging a wrong answer on an AI assistant is the moment when the link between an error and the passage it cited is easiest to establish. That link is fragile. It depends on what the tool retains, on the retention periods your own policies set, and on the memory of the user who clicked. Once it is lost, all that remains is the review campaign, which subject-matter experts fit in after their real work. Preparing documents for AI therefore depends on a condition that is rarely tooled: a path that carries each wrong answer back to the person able to fix its source.

This article is written for the CDO, CTO or head of knowledge management of a large group whose internal assistant already answers users from the document estate: procedures, legal notes, commercial terms, quality standards. If your assistant is still a closed pilot with no way to flag an answer, the argument will apply the day it opens up.

The person this text is uncomfortable for is the expert who wrote the source. Take the lawyer in the contracts team who wrote, in 2023, the note setting out termination terms for business contracts. The scene is the assistant’s steering committee, one quarter after launch. Someone puts up an example of a wrong answer given to a salesperson, citation included. The lawyer recognises her paragraph, and the rule she has since replaced with a newer note. She asks how long the team has known. The silence that follows has a simple explanation: the report landed in the AI team’s feedback dashboard, the team adjusted an instruction, and nobody thought to pass it on to her.

The step this article proposes costs nothing. Take last month’s negative reports from your assistant, if it collects them, and ask one question of each: has the cited document changed since? If none has, the signal stops at the AI team. If some have, and the report is what triggered the change, an informal loop already exists, and it needs writing down before it leaves with the person who keeps it alive. The output is an artefact described below, the source referral.

Two answers come back before the end of the read: “we have launched a document quality project”, meaning a review campaign plus awareness sessions for the business; and “our tooling already covers it”, whether that is the assistant’s feedback dashboard or the document management system’s periodic review workflow. Both are useful and share one trait: they ask the expert for work on top of their own, at a moment someone else chose.

On 28 September we asked who would sign off the correction of a document an agent cites. This article takes the next, narrower question: assuming someone can fix the document, when do they actually do it, and what leads them to?

What Gartner measured, and how far the measure reaches

On 21 September 2026, Gartner published the results of a survey of 223 data and analytics leaders conducted in March: cultural resistance outweighs funding constraints as the primary reason governance initiatives fail, 60% versus 40%. The firm concludes that AI-ready data also requires AI-ready stakeholders. Two of its recommendations matter here: establish shared accountability across business and technology stakeholders instead of treating governance as an IT responsibility, and embed governance into day-to-day operations.

The survey covers data and analytics governance, not the document estate. We do not extrapolate its figure. What it contributes here is its recommendation: governance that arrives on top of the job holds poorly, governance that arrives inside the job has a chance.

For a document estate, “culture” is a word that does not say what to do on Monday. The observable behaviour is simpler. An expert seldom rereads a note out of duty, halfway down a list of documents to review before quarter end. They fix it when it shows them up in front of someone and they are shown exactly where. That is our reading, informed by diagnostic work and not demonstrated by a study: correction follows the incident when the incident reaches the author.

The position we hold is independent of this release. The document layer comes before the engine, and document governance is continuous, whatever analyst artefact is published this quarter.

What assistant feedback fixes, and what it leaves online

Existing tooling deserves to be described at its best. Enterprise assistants collect negative feedback on an answer, often with a category and the sources used. The best dashboards let the AI team group that feedback, derive retrieval or instruction adjustments, and check their effect. It is real work, and it improves the assistant.

Its boundary is precise: these tools change how the assistant reads the corpus, and leave the corpus as it is. When the source is wrong, the AI team has three options: exclude the document, add an instruction that works around it, or wait. None of them touches the lawyer’s note. It stays online, cited by the other tools that read it, and read as is by employees who open it directly. The assistant improves on the surface while the corpus stays wrong.

The document management system’s periodic review has the opposite boundary. It does reach the author, but on a fixed date, unconnected to any incident, and in the form of a list. That is exactly the “on top of the job” governance the survey describes.

The alternative heard most often is the campaign: a document quality project, a sponsor, a list, a deadline. This is the boundary where the lawyer comes back in. During the last campaign she received a list of documents to confirm before quarter close, approved the ones she knew by heart and put the rest off. The 2023 note was on the list. Nothing told her at the time that it was a problem. Three months later an assistant cited it wrongly: the reason to reread it finally existed, and it never reached her.

Another discipline settled this question in writing long ago. Quality management, in ISO 9001, organises correction around the nonconformity: react, determine the cause, check whether similar nonconformities exist elsewhere, review the effectiveness of the action and keep a record. A wrong answer from an assistant is a nonconformity still waiting for its corrective action path back to the document. The catch-up is modest, and the transfer concerns the method, not the object: this is not about certifying a corpus, and the quality function is not the intended buyer.

Fixing a document after a wrong AI answer: the source referral

The step in the introduction produces an artefact, the source referral. It is the only new name in this article.

What it contains. One line per report of a wrong answer: the date, the question asked, rephrased without any user identifier or customer data, the answer given, the passage cited (document and version), the author or authoring team of that passage, the follow-up and its date. The follow-up takes one of three values: document corrected, document confirmed accurate (the error then came from how the assistant read it, and the line goes back to the AI team), or passed to another team.

Where it lives. The queue where the assistant’s reports already arrive, kept by the team that runs it: its ticketing tool, its feedback dashboard or, failing that, a shared file. The referral adds two columns to what exists, “author” and “follow-up”, plus a hand-off through the channel the author already uses. The author column is filled from what the document itself carries: its cover page, its title block or the metadata in the document management system.

When its first lines get written. At the next report of a wrong answer, as it happens. Never during a review campaign, which is precisely when the author has no reason to look.

A completed line. Date: a Tuesday in October. Question: what are the termination terms for a business contract? Answer given: the terms from the old note. Passage cited: legal note “Termination of business contracts”, 2023 version, notice-period paragraph. Author: legal department, contracts team. Follow-up: passed to the lawyer, pending.

Its reading rule. After a month, two outcomes. If most lines show “document corrected” or “confirmed accurate”, the loop works and the referral becomes part of the assistant’s support procedure. If most remain “pending” or have no identifiable author, the transmission path is what is missing: the document does not say who wrote it, or the hand-off lands nowhere. The fix then targets that path before it targets the documents. Who signs off when nobody can be identified is the question we covered on 28 September.

How it fails. If the exercise works, authors also receive reports about accurate documents that the assistant misread. That is its cost: a little expert time to confirm, which the “confirmed accurate” value makes useful by sending the case back to the AI team. If the assistant collects no reports at all, the exercise costs nothing and returns a finding: your programme has no way back.

The record it creates. A referral line establishes that on a given date the company knew a document had produced a wrong answer. If the correction lags, the record shows it. Our reading, with no published decision to support it, is that a report trail followed by a dated follow-up is preferable to no trail, which protects nothing and prevents correction. The legal weight of such a record depends on your context: your legal department should weigh in before it becomes a group rule, and your DPO before any user question appears in it.

The hand-off creates no new role. It is added, in writing, to the assistant’s support procedure and rests on the assistant’s operating budget rather than a new budget line. The trade-off is a dependency: if the assistant is replaced, the referral must be carried over to the next tool, or the loop disappears with it.

What a real deployment showed, and what it does not establish

At TotalEnergies Retail Power & Gas, whose experience report is public and named, roughly 500 pages of official web documentation fed the customer chatbot in production. During a first diagnostic carried out without migration, 19% of the pages turned out to need correction, including cases that were hard to spot by eye at that scale. 53% of the cases were resolved in three weeks, prioritising the most critical ones, with one to two business experts engaged half a day to a day per week. Thomas Bensoussan, Head of Digital Products, highlights two effects: surfacing conflicts that are hard to spot by eye, and indicating which expert each case should go to. Business experts remained the decision makers throughout.

The boundary of this evidence has to be stated. This experience report establishes a modest point that matters here: when each case reaches an identified expert, with the passage in question, correction is counted in half-days per week. It concerns a diagnostic and documents no feedback loop. It says nothing about whether experts are more likely to correct after a wrong answer than during a campaign, and we do not ask it to carry that thesis. The source referral aims to reproduce, as cases arise, the condition the diagnostic brought together: the right case, in front of the right expert, with the passage in view.

What the DKP adds to the referral, and where it stops

A Document Knowledge Platform (DKP) is the document quality and governance layer that runs upstream of AI systems: it governs the estate (Govern), detects and treats anomalies, duplicates, obsolescence and contradictions (Clean), and only activates the corpus for assistants once those two steps hold (Activate). It replaces neither the assistant, nor its feedback dashboard, nor the document management system, nor enterprise search. It is not sold as LLM observability or model evaluation, nor as a support tool, and does not belong in a tender for those categories. Reports are simply the channel through which the document problem becomes visible.

Its contribution to the referral sits in the step the quality standard calls “determine whether similar nonconformities exist”. A report points to one passage. The DKP counts the other documents in the estate that carry the same statement or its opposite, which neither the assistant nor the document management system produces. That is its own unit of work: the contradiction counted across documents, beyond the one just cited. Without it, the lawyer fixes her note, and the next day the assistant cites a sales sheet that repeated the old rule.

The split of responsibilities deserves to be written down. K-AI is accountable for counting anomalies, for their reproducibility and for routing each case to the relevant expert, with a prepared diagnostic. The decision stays with the business: the lawyer says which rule is right. Technically, the analysis covers only the designated document content, within a contractual ingestion scope. Reports, user questions and assistant logs stay in your tools, because the DKP does not need them to count contradictions. No data is reused to train models.

And if you do nothing

The status quo has a predictable outcome. The AI team will keep compensating with instructions and exclusions, each becoming a dependency to carry over at the next change of tool. The next document quality project will send experts a new list, which they will handle like the last one.

Conclusion: audit, clean, monitor

The sequence stays the same. Audit: take last month’s reports and ask the single question, has the cited document changed since? Clean: open the source referral, pass each line to its author, and deal with them on the other documents carrying the same contradiction. Monitor: make the referral a written step of the support procedure, so that every wrong answer travels back to its source instead of stopping at the AI team.

At the next steering committee, the lawyer from the contracts team will not discover her note on screen. She will have received the line on Tuesday, with the passage in question, and the question put to her will be the only one that concerns her: which rule is right.

Frequently Asked Questions

Our AI assistant already has a feedback button. What is missing?

The button collects the feedback and sends it to the team that runs the assistant, which can adjust retrieval or instructions. What is most often missing is the follow-through to the document: the hand-off to the author of the cited passage, and a record of what they did with it. The source referral adds those two columns to the existing queue.

How do you tell whether a wrong answer comes from the document or the assistant?

Only by looking at the cited passage. If the passage is wrong or outdated, the document is at fault; if it is accurate, the assistant misread it. That is why the referral includes the “confirmed accurate” value, which sends the case back to the AI team.

Isn’t a document review campaign enough?

A campaign is useful for working through a backlog, and it remains necessary at the start. It reaches the expert on a date someone else chose, unconnected to any incident, which is why it often leaves aside the documents that cause problems. The referral handles the flow, at the moment a document has actually produced an error.

Should user questions be passed on to experts?

Only in rephrased form, with no user identifier or customer data, and after the DPO has been consulted. The expert needs the cited passage and the nature of the error; the identity of whoever asked is of no use to them.

What confidentiality framework applies to a review of the document estate?

The ingestion scope is contractual and limited to designated document content, excluding transcripts, usage logs and telemetry. No data is reused to train models. The scope is validated jointly by the business Document Owner and the CISO or DPO, never by IT alone.

Sources

• Gartner — Gartner Predicts 60% of Organizations That Ignore Data Governance Culture Challenges Will Fail to Govern AI Successfully by 2027, Mumbai, 21 September 2026 — March 2026 survey of 223 D&A leaders; cultural resistance 60% vs funding 40%; “AI-ready stakeholders”; recommendations on shared accountability and embedding governance in day-to-day operations.
• ISO 9001:2015, clause 10.2 — Nonconformity and corrective action (overview) — react, determine causes, check for similar nonconformities, review effectiveness, retain documented information.
• K-AI — clients page and TotalEnergies Retail Power & Gas experience report — ~500 pages, 19% to correct, 53% of cases resolved in three weeks, routing of each case to the right expert.
• K-AI — TotalEnergies RETEX (PDF)

---

Where to Go From Here

K-AI Corpus Diagnostic — 10 business days on your document estate, full report of the 20 most critical anomalies, money-back guarantee if no meaningful anomaly is found. A one-hour conversation can start from a few lines of your source referral, or from your latest reports: together we look, for each one, at whether other documents carry the same statement or its opposite, and which expert each case should go to. Reach the K-AI team: contact@k-ai.ai. The scope of every diagnostic is validated jointly by the business Document Owner and the CISO/DPO, never by IT alone.

K-AI already works with CMA CGM, Veolia, PwC, BNP Paribas, TotalEnergies and CEVA Logistics. Partners: AWS, Snowflake, Microsoft, Wavestone, Devoteam.