Enterprise-Wide AI Agent Deployment: When One Document Contradiction Becomes a Serial Defect

Cisco opened MyAgent to 90,000 employees. At pilot scale the user is the anomaly detector. At 100% of the workforce, the detector is gone.

On 27 August 2026, Cisco announced the rollout of MyAgent to all of its roughly 90,000 employees: a personal agent, built on the company’s internal Circuit platform, able to call on more than 800 subagents and to run supervised task sequences across messaging, project tracking and the company’s document workspaces.

This is one of the first agent deployments covering an entire workforce, documented publicly by the company itself. What transfers here is not the architecture of a technology vendor building its own platform. It is the move from a few hundred pilot users to 100% of employees, a shift that will reach large non-tech organisations through the far more ordinary route of the productivity suite they have already bought. The thesis of this article is one sentence: that shift changes the nature of document risk, because a contradiction read once by a human becomes a contradiction replayed continuously by subagents that never hesitate.

This is written for a specific reader: the CDO or CTO of a large organisation who has already approved an internal assistant in principle and now faces the decision to open it to every business function. If your deployment is still confined to one pilot department, the argument will hold in six months; it does not bite yet. At pilot scale, the user is the anomaly detector: they know their domain, they see a doubtful answer, they flag it. That is the shift worth holding on to. Opening the agent to everyone at once does not degrade the corpus, it removes the detector.

Two instinctive answers arrive before the corpus question does, and they are worth naming immediately: the in-house document policy, with its classification plan and named owners, and the tool already purchased that appears to cover the ground — data catalog, document management system, data governance platform. We return to both below, because neither answers the question raised here.

One test takes half a day and requires no tooling. Take the five questions your future agent will receive most often — an expense policy point, a commercial discount rule, a purchase approval threshold. Put each one to two different departments and ask for the document of record. The number of times you get back two distinct documents, both in force, is your starting point.

Three recent articles on this blog sit in adjacent territory, and the distinction is worth drawing now. The 14 August piece covered agentic pilots that never reach production; this one covers the symmetrical case, the pilot that reaches production and generalises. The 31 August piece covered the trace an agent leaves when it arbitrates. The 2 September piece covered what an agentic task costs at each run. The subject here is distinct from all three: the effect of deployment scale on the severity of one and the same document defect, independent of its unit cost and of its traceability.

What the Cisco deployment demonstrates, and what it does not

Let us credit the setup in its strongest form, because that is the only way to locate its boundary. MyAgent does not run in a vacuum: it sits on a governed enterprise platform, with approved models, connectors that carry each employee’s permissions, and human supervision announced over task sequences. It is one of the most complete arrangements publicly described to date on the access side.

That architecture answers two questions: who may invoke what, and under whose supervision. It does not address a third, which is ours: among the documents those connectors legitimately expose, which ones contradict each other, which ones were superseded without being withdrawn, which ones no longer have an identifiable owner. A permissioned connector verifies that an employee may read a procedure. It does not verify that the procedure still carries authority.

The order of magnitude of the estate is well established: analyst literature places the unstructured share of an organisation’s information assets in a 70% to 90% range. A 2026 Hyland study conducted with Harvard Business Review Analytic Services, relayed by CDO Magazine, finds that 65% of surveyed executives consider their structured data AI-ready, against 39% for their unstructured data. That gap has been documented for years. What changes in 2026 is its exposure: a deployment covering 100% of the workforce tests it at every execution, across every function at once.

What scale changes about one and the same document defect

The phenomenon has a name you can use in a steering committee: the serial defect, meaning a single corpus contradiction that reproduces identically at every execution because nothing in the chain perceives it as a contradiction. The mechanics are straightforward. As long as a human is doing the reading, two valid but divergent documents produce one questionable answer, at one moment, in front of someone who can hesitate, ask for confirmation, escalate. When hundreds of subagents read the same repository continuously to prepare decisions, the same divergence produces a run of mutually consistent, never-flagged deviations, because no agent perceives having arbitrated anything: it ranked, it kept the best-scored source, it moved on.

This can be counted, and counting is the only move that yields a defensible number. On the repository of a CAC 40 industrial group we audited, 32% of the base consisted of divergent duplicates: two versions of the same content, both accessible, saying two different things. They were surfaced in two weeks and resolved in six, by the business experts themselves.

The exact reach of that measurement is worth stating. It establishes prevalence: in a real large-enterprise document estate, the share of content that contradicts itself is measured in tens of percentage points, and none of those cases was visible without counting them. It says nothing about a rate of wrong decisions on the agent side, and multiplying one by the other would be a mistake: how often an agent actually lands on a divergent pair depends on the questions it receives and the scope it reads. That link remains a working hypothesis, and testing it is elementary: across the hundred most frequent requests sent to your assistant, check how many concern a subject where your corpus already holds two answers in force.

Back to the two instinctive answers. The in-house policy describes the desired state of the estate at a theoretical moment, where the question concerns its real state on a given date. The tools already in place — data catalog, governance platform, document management system — inventory assets, schemas and rights; they do not read the content of two procedures to establish that they say opposite things.

What a Document Knowledge Platform governs upstream of deployment

That is the work a Document Knowledge Platform (DKP) covers: to govern, clean and activate the unstructured document estate, meaning to establish which documents carry authority, who owns them, which version is in force, and where the corpus contradicts itself. The discipline runs upstream of the agentic layer. A DKP is distinct from an enterprise search engine, an agent platform and a data catalog, and has no ambition to become one: it prepares the ground those components run on.

A word on vocabulary, to avoid a category misunderstanding. The term used above is borrowed from industrial quality language because it describes the mechanics well. A DKP is nonetheless sold as a document quality layer upstream of AI, not as a quality management system or a compliance module. The vocabulary is a channel of visibility onto a document problem; it does not define the category.

One scoping point matters here, because it determines how the file is handled by the CISO and the DPO. The analysis covers document content — procedures, contracts, internal policies, technical and customer documentation. It ingests no assistant conversation transcripts, no usage logs, no application telemetry. The ingestion perimeter is contractual, and analysed content is not reused to train models.

A final point on sequencing, because it is the most common ordering error we see in the field. Cleaning is run as a continuous operation rather than as a one-time prerequisite: a corpus degrades as it lives, and the pace of document production is accelerating precisely because assistants generate documents too. The sequence that holds is this one: audit before opening access, clean by order of criticality, then monitor continuously for contradictions reappearing.

What the regulatory framework already requires, and what has been deferred

A calendar note, re-verified as of today. The EU AI Act was amended by Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026 and in force since 27 July 2026: high-risk systems under Annex III move to 2 December 2027, those under Annex I to 2 August 2028.

That deferral releases nothing on the perimeter currently in force. Transparency obligations have applied since 2 August 2026, and the allocation of roles is unchanged: designing logging capability falls to the system provider, while retaining those logs and exercising effective human oversight falls to the deployer, meaning the organisation opening the agent to its entire workforce. No vendor can certify on your behalf that the documents its agent read carried authority on the day it decided.

Conclusion: audit, clean, monitor

Enterprise-wide agent deployment is a sound project, and there is no reason to delay it. There is a reason to prepare it differently. At the scale of one pilot department, corpus quality remains a continuous improvement topic. At the scale of the whole workforce, it becomes an operating condition, on the same footing as permissions and availability.

Go back to the half-day test: five questions, two departments, one document of record. If you get more than one divergent answer out of five, you have just measured a sample. The next question is how many comparable divergences sit in the full corpus your agent will read, and the only honest way to answer it is to count them.

Frequently Asked Questions

Our agent platform already enforces each user’s permissions. Isn’t that enough?

Permissions determine what an agent is allowed to read. They do not determine whether what it reads is current, unique and authoritative. Two contradictory documents, both perfectly authorised, remain two contradictory documents.

We already run a data catalog and a data governance programme. Where is the difference?

A data catalog inventories assets, schemas and accountabilities. A Document Knowledge Platform reads document content to establish which items contradict each other, which were superseded without withdrawal, and which have no owner left. The two are complementary, on two different objects.

Do we have to clean the whole corpus before opening the agent to every function?

No, and that would be the surest way never to deploy. The useful sequence is to audit the scope the agent will actually read, treat findings by order of criticality, then monitor for contradictions reappearing over time.

What happens to confidentiality during a diagnostic?

The ingestion perimeter is contractually defined and limited to document content: no conversation transcripts, no usage logs, no telemetry. Analysed content is not reused to train models. Scoping is validated jointly with the business Document Owner and the CISO or DPO, never with IT alone.

How long before we get a first order of magnitude?

On a CAC 40 industrial group’s repository, divergent duplicates were surfaced in two weeks and resolved in six by business experts. The entry diagnostic itself runs in ten business days on a defined perimeter.

Sources

Where to Go From Here

K-AI Corpus Diagnostic — 10 business days. We count the contradictions and divergent duplicates in your corpus and hand you a report of the 20 most critical anomalies, each with the business owner it should be routed to. Money-back guarantee if no meaningful anomaly is found. To count the divergences in your corpus before opening the agent to every function, 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.