An AI agent can be created in one sentence. Deciding which documents it should believe can take weeks

Creating an agent is now fast. Saying, rule by rule, which internal document it should believe takes weeks. A timed five-rule record measures it.

In an enterprise agentic AI project, building the agent has become the short step. The long step is saying, rule by rule, which internal document the agent should believe. A new hire who finds three price lists asks their manager which one applies; an agent does not ask: it reads, and it answers. The real timeline of your agentic transition is therefore counted in rules for which nobody has yet named the document of record. That work is done once, upstream, and every agent that later reads the same rules benefits: it starts from sources already settled, and it answers with the right one.

Picture a Monday at an industrial group. The steering committee approves a “pricing assistant” agent for the sales force. On Tuesday, the agent is created. Three weeks later, it still has not been opened to users. The head of sales administration, who must sign off its launch, asked a single question: for legacy accounts, which of the three price lists on the sales Drive is authoritative? Nobody in the meeting could answer, and the question has been waiting for a decision ever since.

The backdrop is dated. On 8 October 2026, Google introduced the “Gemini agent”, a single agent for work. In its announcement, the vendor writes that describing the role you need is enough for the agent to be created, that the agent understands the business, from pricing to product portfolios to departmental norms, and that it keeps a semantic memory, “a structured knowledge base it builds as it reads documents, talks to people, and works with other agents”. It adds that the agent onboards itself “the way a new hire would”. Google is announcing, not measuring, and the text gives no general availability date. What transfers to a large non-tech company is the mechanism: the agent learns from what it reads. What does not transfer is the tool: the same mechanism will arrive through whichever agent platform your IT department selects.

This article is written for the CDO or CTO preparing the first agents of 2027. If your first agent’s task relies only on structured data (an ERP, a data warehouse), the argument bites less; it bites as soon as the task assumes a rule written in a document.

Two responses come naturally. The first lets the agent learn from the whole Drive or the whole SharePoint estate, relying on existing access rights and on the platform’s search. The second hand-builds a folder of copies dedicated to the agent, reviewed once before launch. The first gives the agent all three price lists and lets it choose; the second freezes a choice made under launch pressure, and the folder ages with the next price list update. The natural objection remains: the agent will take the most recent version. Yet the most recent of the three price lists does not apply to legacy accounts; that is exactly what the head of sales administration was asking.

A note for regular readers of this blog. On 9 October, we proposed a baseline review of procedures before a knowledge layer’s first ingestion; on 16 October, we showed that an agent can serve a superseded version. Here the object changes: not all your documents, only the few rules a specific agent’s task depends on, and the proposed measure is a duration.

What the agent reads to do its job

Take the pricing assistant’s task: preparing a price proposal for an existing account. To do it, the agent reads the price list, the rule on the maximum discount granted without approval, the specific terms for legacy accounts, the note on this year’s indexation and the payment terms policy. Five rules, each written down somewhere.

When the document for each rule is settled, the agent proposes a price that sales administration can sign. When it is not, three outcomes are possible: the agent picks one of the price lists, and its answer is right or wrong depending on what its search retrieved; it blends them, and its answer is plausible without being applicable; or the project team removes the rule from its scope, and the agent loses part of its value.

This is the time-to-reference: the time it takes to say, for each rule an agent’s task depends on, which document is authoritative; it is counted by enumerating, among those rules, the ones for which several candidate documents coexist without a written decision.

This duration has a property a budget committee will appreciate. A rule whose document of record is settled stays settled for the next agent. The collections assistant will read the same payment terms policy; the bid-response agent will read the same price list. We have no measurement of that gain; it is a mechanism, and it can be checked at the second agent: rules already decided do not come back to the table.

The status quo has a precise cost too: each agent reopens the same question, the decision is remade project by project, and an agent opened without it answers with whichever price list its search found.

What the platform does best, and where it stops

At its best, a recent agent platform does a lot. Google’s announcement describes a dedicated identity for every agent, fine-grained permissions approved by security administrators, an audit trail of every action, and a policy that can forbid an agent from opening a category of documents. It also describes a knowledge catalog that maps business definitions such as net margin once, so that every agent uses them.

The boundary lies in the object. These controls tell you who the agent is, what it may open and what it did. The catalog defines terms and metrics carried by data. None of these mechanisms, as the announcement describes them, says which of three price lists, all accessible and all current at their own date, applies to legacy accounts. That decision belongs to the business function that wrote the price lists, and it is the one the head of sales administration was waiting for before signing.

Quality management already wrote this requirement for people

This is not new. ISO 9001:2015, clause 7.5.3.1 a), requires documented information needed by the quality management system to be controlled so that it is available and suitable for use, where and when it is needed; clause 7.5.3.2 c) adds control of changes, for example version control. Because the standard is paywalled, this wording was read through an auditor training provider (Auditor Training Online). For the 2026 revision, the certification body Perry Johnson Registrars, in its reading of the final draft, reports a change in documented-information vocabulary and unchanged documentation requirements.

Treating an agent as a new point of use of documented information is our reading; the standard does not make it. It does, however, give the right method for a newcomer: the information they use must be the right version, where they work. For a salesperson, the question asked to a manager closes the gap. For an agent, which does not ask, the catch-up is to ask the question on its behalf, before launch.

The timed record of your first agent’s five rules

The exercise needs no tool: for the first agent planned for 2027, take the five rules its task depends on, and time how long it takes to say, for each one, which document is authoritative.

What it produces. A five-line record with five fields: the rule; the candidate documents found; the authoritative document, or “undecided”; the owner who decided or must decide; the time elapsed until a written decision.

The stopwatch. The session that opens the record takes half a day, with the agent’s product owner and the business owner. For a rule decided in the session, the clock stops there. For a rule referred to its owner, the clock keeps running until the written decision. The longest of these durations is the agent’s minimum time to launch.

Where it lives. The record is an appendix to the agent’s scoping file, kept by its product owner and approved by the business owner who will sign off its launch. No standalone document is created.

When it is written. The first lines are written at scoping, between use-case approval and agent creation. If the agent already exists, the record is done before the review that decides its opening to users.

How to choose the five rules. Reread the task description in the scoping note and underline every business rule it assumes: a rate, a threshold, an exception, a deadline. If there are fewer than five, take all of them; if there are more, take the five that commit the company to a third party.

A completed line. “Maximum discount without approval, legacy accounts; candidate documents: 2025 price list on the sales Drive, sales management note of March 2026, sales administration internal FAQ; authoritative: undecided; owner: sales management; time elapsed: ongoing, decision requested at the session.”

How to read it. A decided rule has a document of record: the agent may read it. An “undecided” rule is written out of the agent’s scope in the scoping file until its owner has written the decision; meanwhile, proposals that depend on it remain handled by sales administration, as before the agent.

How it can fail. If all five rules are decided in the session, the exercise will have cost half a day and you will know your first agent’s timeline is decided elsewhere. If even one stays open, you know the agent’s real date of full launch: the date of that decision.

The record it creates. Writing that a document is authoritative leaves a trace: if the decision was wrong, it will show who made it and when. Our reading, argued and without any published decision to support it: a written, dated decision is more defensible than a choice left to the agent’s search. For a pricing rule enforceable against customers, legal reviews the wording.

Who owns the decision. No new role: each decision goes to the rule’s existing owner, as a written assignment added to that role. The time spent is charged to the agent project’s budget, already approved.

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: it governs the document estate (Govern), detects and handles anomalies, duplicates and obsolescence (Clean), and activates the corpus for agents only once those two steps hold (Activate). It is not sold in the agent platform category or the data catalog category, and replaces neither; it works on the documents they consume.

On the record, K-AI’s contribution is specific. K-AI counts the documents that address the same rule, indicates the applicable version of each, and flags obsolete documents and competing versions held elsewhere. To link a rule from the scoping note to the documents that carry it, the tool proposes candidate documents and the team confirms. The decision on which document is authoritative stays with the owner. That is where the time is saved: the session starts from a list of candidates already established.

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

A real deployment gives an order of magnitude for the upstream problem. On a document estate audited by K-AI (anonymised case): over 1,300 anomalies detected, about 30% of document volume removed, over 90% precision after cleaning. The public source states neither the size of the estate nor the timeframe. This measurement establishes that a real estate can carry a significant share of documents to remove before an AI reads it. It establishes nothing about an agent’s deployment time, which was not its object.

Conclusion: audit, clean, monitor

Auditing means writing the timed record of the first agent’s five rules. Cleaning means having each owner decide on the “undecided” rules and withdrawing or flagging documents that are no longer authoritative. Monitoring means repeating the record for every new agent, starting with rules already settled: they are what makes the second agent faster to launch than the first.

At the next launch review, the head of sales administration will know, rule by rule, which document the agent will read, who designated it and since when.

Frequently Asked Questions

Why do AI agents depend on a company’s internal documents?

Because part of the rules they apply (prices, thresholds, exceptions, deadlines) is written in documents rather than in structured data. Recent agent platforms say so openly: Google describes an agent that builds its knowledge of the company by reading its documents. What the agent learns depends on what it reads.

What slows down enterprise AI agent deployment?

Creating the agent has become fast. Time is then spent deciding, for each rule its task depends on, which document is authoritative when several coexist. Until that decision is written, the agent cannot safely be opened on that rule.

Can’t the agent simply use the most recent version?

The most recent version is not always the one that applies: a 2026 price list can coexist with a 2025 list still valid for some accounts. Only the rule’s owner knows which is authoritative, and that decision must be written for the agent to read it.

How can I estimate the real deployment time of my first agent?

Take the five rules its task depends on and time how long it takes to say, for each, which document is authoritative, up to the written decision. The record lives as an appendix to the agent’s scoping file; the longest duration gives the minimum time to launch.

How does K-AI handle confidentiality during an analysis?

K-AI ingests only the document content within scope, set by contract and approved jointly by the business document owner and the CISO or DPO, never by IT alone. No transcripts, usage logs or telemetry 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 five rules in your record and the time they took you; it locates those for which several candidate documents coexist 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.