Reading all your published documents side by side now costs your customers almost nothing

An AI assistant can compare your terms, notices and FAQs in minutes. Starting from the complaints you already received shows the gaps before the next letter.

Any discrepancy between two of your published documents can now be found from the outside before it is seen from the inside. Comparing terms and conditions, a product notice, a policy and an FAQ word for word used to cost the person challenging you hours; a generative AI assistant can now do it in minutes. Inside the organisation, cross-reading is still manual. The consistency of published customer documents, long protected in practice because nobody read them all together, has become a standing exposure.

The signal comes from the public sector. A study by researchers from the Hertie School, the Cooperative AI Foundation and GovAI, posted in August 2026 and presented at the AIES conference in October, describes public services strained by requests drafted with language models, either more numerous or longer and more complex. The authors themselves state that their method supports no causal or statistical conclusion. Applying this to companies is our reading: the study says nothing about customers or terms and conditions.

Imagine a Monday morning in the complaints team of an insurer or an energy supplier. A multi-page challenge arrives, clean and well structured. It cites clause 7.2 of the 2024 terms and conditions and the 2026 FAQ on the website, which say different things on the same point. The in-house lawyer spends half a day establishing which of the two texts applied to this customer on the date in question. The head of complaints handling, who will sign the reasoned reply, confirms the discrepancy is real. The customer, or their assistant, took minutes to find it.

This piece is for the CDO and the head of customer relations at a large group that publishes documents binding on its customers: terms and conditions, notices, policies, FAQs, price lists. If your published documents fit on a few pages reviewed by a single team, the argument bites less; it bites as soon as several departments publish on the same subject.

Two responses come to mind. The first handles the flow through capacity: add staff, or hand triage to the complaints management tool you already run, which can summarise each case. The second adds friction: a single channel, a mandatory form, length limits. Both act on the flow. Neither removes the published discrepancy that fed the complaint. And on the discrepancy itself, a complaint citing two divergent published texts is irrefutable, however good or bad the rest of the letter is. As for the editorial approval workflow, which signs off each document when it is published, it does not reread its neighbours.

A note for regular readers. On 21 September we covered the wrong answer an organisation gives to an authority that questions it; on 16 and 21 October, the reader of your documents was internal, your assistant or your agents. Here the reader changes: external, reading your published documents, looking for the gap. The response changes with it: start from the complaints you have already received.

What the study establishes, and what transfers to a large enterprise

The study by Chris Schmitz, Lewis Hammond and Alan Chan calls “agentic flooding” an AI-driven rise in the volume or complexity of requests to a public service, strong enough to strain its capacity. The mechanism shared by the cases it collects is simple: language models produce large amounts of text cheaply, and people carry that text into existing procedures. The authors also note that friction measures, the fastest to deploy, mainly hold down the number of requests and come at a cost to equitable access.

What transfers to a banking, insurance, energy or telecoms group is the falling cost of writing and reading for the person who complains. What does not transfer are the constraints specific to public services (equal access, bans on certain fees) and the orders of magnitude. The route by which the phenomenon reaches a large enterprise is well known: the written complaint, then the ombudsman, then litigation.

A consumer-side relay exists. On 14 October 2026, Business AM reported that Belgian customer service and ombudsman teams were facing AI-generated complaints, overly long and loaded with irrelevant legal jargon; only the standfirst is publicly available and we have not read the full article. That relay carries a useful nuance: part of these texts is poorly founded, and the triage tool will set it aside. The complaint that cites a real discrepancy between two of your documents needs different handling, because its strength lies in the documents and nothing else.

The approval workflow signs off each document, and the complaint reads them all

At its best, a large group’s editorial approval workflow does a lot. Legal reviews the terms and conditions, compliance approves the notice, product marketing signs off the FAQ, and each version carries a publication date. For each document, taken one by one, the control exists.

The boundary lies in the unit of control. The workflow validates a document on its publication date. It does not reread the FAQ when the terms and conditions change, or the notice when a price list is replaced. Two texts, each correct on its date, can diverge a year later without any workflow having failed. The person complaining reads the whole set as of the date that matters to them.

This is the reading asymmetry: comparing your published documents costs almost nothing outside and remains expensive inside; it is measured by counting, on a given point, the published documents whose cited version was not in force or does not match the others.

This is where the head of complaints handling is exposed. The triage tool saves time on each case, friction slows the next ones, and the same discrepancy will come back in the next complaint for as long as it stays published.

The status quo has a precise cost: each complaint citing a discrepancy is handled as a one-off, the goodwill gesture or correction is decided case by case, and the document that created the discrepancy stays online.

Complaints already lead back to root causes: banking and insurance wrote it down

This is not new. In France, Recommendation 2024-R-02 of the ACPR, the banking and insurance supervisor, applies to banks, insurers and intermediaries. It asks them to put in place the means to identify, through written complaints, malfunctions, regulatory breaches or poor practices, and to take corrective measures, in particular towards customers affected by the incorrect application of a contractual provision. It also asks for a summary, at least annually, to the relevant governance bodies, describing these malfunctions and corrective measures (free translation).

Reading a discrepancy between two published documents as a malfunction of that kind is our reading; the ACPR does not make it. For sectors outside its scope, the recommendation remains a model of method: trace complaints back to their cause, and report the cause to governance. The catch-up suggested here is to add the document as a possible cause.

The cited-documents log, built from your last ten complaints

The exercise takes half a day and needs no tool: take the last ten written complaints that cite a document, and write one line per document cited.

What it produces. A cited-documents log, with five fields per line: the document and the version cited; the date of the disputed event; the version in force on that date; consistency with the other documents published on the same point (yes, no, to check); the owner of the document.

Where it lives. The log sits in the periodic complaints review file, kept by the head of complaints handling. For a bank or an insurer, its conclusions feed the annual summary presented to governance bodies. No standalone document is created.

When it starts. The first lines are written at the next periodic complaints review, outside any case being handled, then at each review after that. Without a formal periodic review, the moment is the next customer relations quality meeting. To find the ten complaints, searching the written complaints register for “terms and conditions”, “clause”, “FAQ” or “notice” is enough for a first pass; if there are fewer than ten over the last quarter, the log starts from those that exist.

A filled-in line. “Terms and conditions 2024, clause 7.2 (early termination fees); disputed event in March 2026; version in force on that date: terms and conditions 2025; consistency: no, the website FAQ still describes the 2024 rule; owner: legal for the terms and conditions, product marketing for the FAQ.”

The reading rule. A “yes” line closes the matter for that document. A “no” line calls for two decisions from the owner: which wording prevails, and the withdrawal or update of the other. A “to check” line stays open until the next review, with a name against it.

The data handled. The log contains no customer data: no name, no contract number, only the document reference, its version and the date of the event. The minimisation rule is validated by the DPO before the first review.

Failure mode. If all ten lines are “yes”, the exercise cost half a day and you know your recent complaints do not rest on your documents. If even one line is “no”, it is a discrepancy the next complaint can cite in turn.

The record it creates. A “no” line documents a known discrepancy. Our reading, argued and without any published decision to support it: a discrepancy written down, dated and assigned is more defensible than one discovered in the next letter. A “no” line can also raise the question of other customers to whom the outdated wording was applied; in banking and insurance, this is the case the ACPR recommendation targets when it refers to corrective measures towards affected customers. Compliance and legal are therefore involved from the first “no” line, and remain the judges of both the wording and what follows.

Who decides. No new role: the decision belongs to the owner of each document, through a written assignment added to an existing role. The time spent is charged to the complaints handling set-up, which is already budgeted.

What a Document Knowledge Platform brings, and what it does not

A Document Knowledge Platform (DKP) is the document quality and governance layer that runs upstream of AI systems and of the channels that expose your documents: it governs the document estate (Govern), detects and handles anomalies, duplicates and obsolescence (Clean), and only activates the corpus once those two steps hold (Activate). It is sold neither in the complaints management category nor in that of the content management systems or DMS that publish your documents, and replaces none of these tools; it works on the estate they expose, published documents included. The complaint is simply the channel through which the document problem becomes visible.

On the cited-documents log, K-AI’s contribution is specific. K-AI counts the published documents covering the same point, indicates the applicable version of each, and flags obsolete documents and competing versions present elsewhere. To link a complaint to the document it cites, the tool proposes candidate documents and the team confirms. The decision on which wording prevails stays with the owner.

Ingestion covers only the published document content within scope, set by contract and validated jointly by the document owner and the CISO or DPO. No complaints, no transcripts, no usage logs. No content is reused to train models.

The closest case comes from TotalEnergies Retail Power & Gas, whose customer chatbot relies on around 500 pages of official web documentation. During an initial diagnostic of those ~500 pages, mapped and audited without migration, K-AI identified 19% of pages needing correction, including cases impossible to spot by eye at scale; 53% of cases were resolved in 3 weeks, with 1 to 2 experts working half a day to a day per week. This measurement establishes that a body of published customer documentation can carry a notable share of pages to correct. It establishes nothing about complaints or disputes, which were not the subject of that diagnostic.

Conclusion: audit, clean, monitor

Auditing means writing the cited-documents log from your last ten complaints. Cleaning means having each owner decide on the “no” lines and withdraw the wording that no longer prevails. Monitoring means updating the log at every periodic review, so the discrepancy is seen from the inside before the next letter arrives.

At the next complaint citing two of your texts, the head of complaints handling will know whether the discrepancy is already known, who decided it, and since when.

Frequently Asked Questions

Why does generative AI change the exposure of a company’s published documents?

Because it makes reading and comparing terms and conditions, notices, policies and FAQs almost free for the person challenging the company. The company still reviews its documents one at a time. A discrepancy between two published texts can therefore be found from the outside before it is seen from the inside.

What is “agentic flooding”?

It is the term used by a study posted in August 2026 (Hertie School, Cooperative AI Foundation, GovAI) for an AI-driven rise in the volume or complexity of requests to public services. The authors state that their method supports no causal or statistical conclusion. Applying it to companies is K-AI’s reading.

Is an AI complaint-triage tool enough?

It saves time on each case. It does not remove the published discrepancy behind the complaint: as long as two live documents diverge on the same point, the next complaint can cite it.

How can I check the consistency of published customer documents from complaints?

Take the last ten written complaints that cite a document and write one line per document cited: version cited, date of the event, version in force, consistency with the other published documents, owner. The “no” lines point to the discrepancies to resolve first.

How does K-AI handle confidentiality during an analysis?

K-AI ingests only the document content within scope, set by contract and validated jointly by the business document owner and the CISO or DPO, never by IT alone. No complaints, transcripts or usage logs are collected, and no content is reused to train models.

Sources

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 starts from the documents your log cites, locates those with several versions still published, and clarifies what it would take to count their competing versions. 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.