The knowledge layer you are budgeting for 2027 will inherit the state of your procedures

A knowledge layer for AI agents is built by reading the company’s procedures as they stand on the day of ingestion. Whatever is uncertain in them, an overdue review date, an “except in special cases” left hanging, two versions in circulation, can come out the other side as a clean rule that nobody reads as a text any more. The state of your procedures before the build can only be recorded before it. Without an entry baseline, the defects of your procedures become the defects of your knowledge layer. Property law settled this long ago: a tenant who moves in without an inventory of condition is presumed to have received the premises in good repair.

The scene is playing out in budget committees right now. The CDO presents the 2027 line: a shared knowledge foundation for agents, connectors, a graph, continuous evaluation. The CFO asks a simple question: “If an agent applies a wrong procedure next year, how will we know whether the error came from the tool or from the procedure?” The CDO has no written answer. This piece is written for that CDO, and for the head of knowledge management who will share the question. If your agents read documents live, with no layer extracting rules from them, the argument will apply to the first project that structures them.

Two replies come up before the end of the page. The first: “our procedures already go through periodic review in the quality management system.” True, and that review checks that a document is current for a human reader who knows its context; it says nothing about what it becomes once cut into executable rules. The second: “the tool we are buying detects contradictions itself.” Some tools claim exactly that, and we look at them below in their best form; the integrator’s scoping phase, where there is one, inventories the sources to connect. Both work on the present. Neither records the starting state.

The test that tells you whether this concerns you needs no tool. Take the first business domain on the knowledge layer roadmap and ask two of its experts for the list of procedures they themselves know need revisiting. If the list is empty, record it with the date and the experts’ names: that is the first line of your baseline, and it already protects the programme. If it is not empty, those documents are the first ones the layer will turn into rules.

This piece extends our September 25 article, which showed that an AI programme can end up rebuilding a document description that already exists elsewhere in the company. The October 7 article dealt with the copies a withdrawn document leaves behind; this one is about a document still in place, whose state at build time is written down nowhere.

What 2027 budgets are buying under the name of knowledge

Budget season sets the tone. In its 2027 Budget Planning Guides, published on July 14, 2026 and based on a survey of more than 2,600 decision-makers, Forrester lists “enterprise context that agents can act on” among the areas to increase: machine-readable information and governed enterprise knowledge that let agents navigate business policies, processes and systems with greater accuracy. The same document recommends that, instead of cutting spend across all data cleanup efforts, organisations focus on targeted fixes. Three months on, October and November are when those recommendations become approved budget lines.

On October 5, 2026, MIT Technology Review Insights published the summary of a survey of 300 data, AI and technology executives, produced in partnership with Neo4j, a graph database vendor with a direct interest in its conclusion. What the summary establishes (the full report, distributed by the vendor, was not read for this article) is still useful, read as a market admission: on average, around a third of respondents’ agentic AI projects reach production; lack of knowledge and context ranks among the main points of failure, alongside legacy systems and security concerns; planned investments target ingestion pipelines, AI-ready APIs, RAG, evaluation agents and knowledge graphs. The study distinguishes three kinds of knowledge: semantic, episodic and procedural. The third, how the company does things, lives mostly in documents.

What transfers to a large industrial or financial group is the shape of the spend: knowledge is becoming an infrastructure line. What does not transfer is the production figure, measured on an international panel and a scope the sponsor chose. And the investment list has one feature worth noting: every item on it carries, structures or evaluates knowledge. None addresses the state of the procedures it contains at the start.

Knowledge layer for AI agents: what the tooling does best, and what it cannot date

Take the offer in its best form. SAP is previewing Company Memory in closed beta, a procedural knowledge layer associated with Signavio. According to its product page, it ingests policy documents, process models, agent traces and voice input, and structures each item as a “process atom” with a clear purpose, scope and guiding principles. Business owners review and approve every behaviour in their domain, every change is logged, and the tool surfaces contradictions before any agent acts on conflicting rules. Neo4j, for its part, describes the knowledge layer as a shared, governed substrate where enterprise knowledge lives.

That progress has a limit built into it: governance applies to extracted rules, from the moment they exist. It compares atoms with one another. It does not keep what the document said about its own doubt, the “under revision” note, the footnote pointing to an exception, the parallel version one department still applies. And it cannot date the state of the estate before its own first reading, since that reading is what opens its history.

This is where the CFO’s question becomes the CDO’s. The day an agent applies a wrong rule, the layer’s history will show when the rule was created and who approved it. It will not show that the source procedure was already ambiguous before the build. Without an earlier record, our reading is that the error will first be looked for on the project’s side, because the project is the only party whose trail starts on that date. That is an anticipation of the internal review that follows, not a written rule.

Back to the lease. Under Article 1731 of the French Civil Code, if no inventory of condition was drawn up, the tenant is presumed to have received the property in good repair and must return it in that state, unless proven otherwise. The transfer concerns the method, not the object: a document is not a flat, and nobody will plead before a judge. For a knowledge layer programme, the factual position is similar, with no legal presumption at all: without an entry baseline, nothing distinguishes what pre-existed from what the project introduced.

The entry baseline

What passes on in this way is an anomaly present in a procedure before the knowledge layer is built, which the layer reproduces as a rule and which nothing afterwards distinguishes from a project error. It is counted by enumerating, domain by domain, the anomalies in the procedural estate before the first ingestion.

The artefact that counts them is the entry baseline. For each business domain on the roadmap, it holds one line per source procedure retained, with five columns: the document and its last review date; the anomaly found (divergent duplicate, outdated version still in circulation, exception deferred to another text, contradiction with another procedure); the decision taken (fix before ingestion, ingest with a reservation, exclude); the expert who confirmed the anomaly; the date of the decision.

Here is a filled-in line. Domain: handling of B2B customer complaints. Document: complaints handling procedure, revised in 2024. Anomaly found: the first-response deadline differs between the procedure and the application note issued the following year by customer service. Decision: fix before ingestion, the application note being authoritative according to the department. Confirmed by: the head of customer relations. Date: October 21, 2026, during budget framing.

Its home is the scoping file of the knowledge layer programme, as an appendix, kept by the programme director and reconciled with the document review register of the quality management system, which remains the reference for revisions. Its first lines are written during budget framing, before the first ingestion, never in production, when the defect will already have changed form. The work sits on the very budget line it protects: it is a work package of the knowledge layer scoping, funded with it, and it gives substance to Forrester’s recommendation to target fixes rather than spread them.

The reading rule has two outcomes. A domain whose baseline shows no substantive anomaly enters the layer as is, and its baseline becomes the starting record. A domain with anomalies that change a rule (a deadline, a threshold, an approval route) is ingested only after correction, or with a written reservation visible to the owners who will approve the extracted rules.

Writing “ingest with a reservation” creates a record: it establishes that the anomaly was known on a given date. If the decision proves wrong, the record will show it. That is better than no record, which leaves the project exposed by default and says nothing about who knew what. The baseline contains only documents and their metadata: no user questions, no usage logs, no personal data.

The test also has a failure mode. If the experts find nothing to report, it will have cost a few hours of their time and produced a dated record that protects the programme. The only worthless result is the one nobody writes down.

What one real engagement showed, and what it does not establish

In a CAC 40 industrial group, during a diagnostic run by K-AI, close to a third of the base examined (32%) consisted of divergent duplicates: documents meant to say the same thing that no longer did. They were surfaced in two weeks and resolved in six, with decisions left to the business experts.

This case establishes that a defect of this kind existed in the estate before any knowledge layer project, and that it could be counted and then treated within a timeframe compatible with budget scoping. It does not establish what share of those duplicates would have become rules in a knowledge layer, nor that the proportion would recur in another group or another domain. That is precisely what the entry baseline lets each organisation measure for itself, domain by domain.

What a DKP adds, and where it stops

A Document Knowledge Platform (DKP) is the document quality and governance layer that runs upstream of AI systems. It has three functions: Govern, governing the document estate (authority, lifecycle); Clean, detecting and treating anomalies, duplicates, obsolescence and contradictions; Activate, exposing the corpus to agents only once the first two hold. It does not replace the knowledge layer, the graph, company memory, ingestion pipelines, evaluation agents or model observability. It is not sold in the process management category and does not compete there. Its unit of work is the document and what it asserts, before any extraction; the knowledge layer works on the rule, after.

In practice, K-AI is accountable for counting anomalies, making that count reproducible and routing each case to the relevant expert. It is not accountable for the substantive decision, which stays with the business. The review covers only the document content designated in the contract; no content is reused to train models. This position does not depend on the analyst calendar: the document layer comes before the engine, whatever trend ranking is published this autumn.

And if you do nothing

The status quo has a predictable result. The knowledge layer will be built on the estate as it is. Its tooling will catch some of the contradictions between extracted rules, which is useful. The uncertainties the documents spelled out will have vanished in extraction. And an agent’s first visible error in 2027 will open an investigation whose trail starts on ingestion day.

Conclusion: audit, clean, monitor

Audit: for the first domain on the roadmap, ask the experts which procedures they know need revisiting, then draw up the entry baseline. Clean: fix before ingestion whatever changes a rule, and write the reservation for the rest. Monitor: reconcile the baseline with the document review register at every new ingestion, so the layer and the documents age together.

At the next budget committee, the CDO will not have to improvise an answer to the CFO. They will open the appendix of the scoping file: domain by domain, the state of the procedures on entry day, dated and confirmed by the experts.

Frequently Asked Questions

What is a knowledge layer for AI agents?

It is a shared foundation between a company’s data and documents and its agents: it structures definitions, rules and procedures so that every agent reads them the same way. Depending on the offer, it takes the form of a knowledge graph, a company memory or a set of governed procedural rules.

Why baseline procedures before building a knowledge layer?

Because the layer reads procedures in their current state, and an uncertainty written into a document can become a clean rule once extracted. The entry baseline records, before the first ingestion, the known anomalies and the decision taken for each; it then lets you tell a tool defect from a defect already present in the document.

Isn’t the contradiction detection built into knowledge layer tools enough?

It is useful and compares extracted rules with one another. It does not keep what the document said about its own limits (revision in progress, exception deferred elsewhere) and cannot date the state of the estate before its first reading.

What does a document entry baseline contain?

One line per source procedure in a domain: the document and its review date, the anomaly found, the decision (fix, ingest with a reservation, exclude), the expert who confirmed it and the date. It lives as an appendix to the programme’s scoping file and is reconciled with the quality management system’s document review register.

How does this fit into a 2027 AI budget?

As a work package of the knowledge layer scoping, funded by the same line. It matches Forrester’s recommendation, in its 2027 guides, to target data fixes where they unblock agents, without launching a general cleanup.

What confidentiality framework applies to a document estate review?

The ingestion scope is contractual and limited to the designated document content, excluding user questions, 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

• MIT Technology Review Insights — Connecting AI agents to enterprise knowledge, October 5, 2026 — sponsored content in partnership with Neo4j; survey of 300 data, AI and technology executives; figures taken from the published summary, full report not read.
• Forrester — 2027 Budget Planning Guides, press release, July 14, 2026 — 2,600+ decision-makers; “enterprise context that agents can act on” as an area to increase; targeted fixes rather than across-the-board cleanup cuts.
• SAP Signavio — SAP Company Memory preview (closed beta), accessed October 9, 2026 — process atoms, business-owner review and approval, contradictions surfaced before agents act.
• Neo4j — The knowledge layer for enterprise AI — the knowledge layer as a shared, governed substrate.
• French Civil Code, Article 1731 (as cited by Le Monde du Droit) — presumption of good repair absent an inventory of condition, unless proven otherwise.
• K-AI — clients page — CAC 40 industrial group (anonymised): 32% divergent duplicates, surfaced in two weeks, resolved in six.

---

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 one domain on your roadmap: together we look at which procedures would fill the anomaly column of its entry baseline, and which expert each one 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.