An AI agent skill is a procedure the agent executes. Who approved it?

A business user teaches an agent a task in one conversation. If the skill copies a procedure, the agent runs a rule nobody approved.

Your AI agents will soon run two kinds of procedures: the ones your quality system approved, and the ones a colleague taught them in a single conversation. Vendors have a name for the second kind: AI agent skills. A skill is a set of written instructions the agent reads before it acts, sometimes bundled with the reference documents it should consult. It is an internal document, re-read by the agent every time the task comes up. Vendors sum up the promise as teach once, work everywhere, and the promise cuts both ways: a correct threshold taught once applies everywhere, and so does an outdated one. When the skill points to the procedure in force and that procedure’s owner has validated it, the agent applies the right rule and the skill can be reused without another review. When the skill restates the procedure from memory, the agent applies a rule nobody approved, in every team that uses it.

Picture a Tuesday in an industrial group. An experienced buyer teaches his team’s agent how to prepare the onboarding of a new supplier: which documents to request, which checks to run, which amount triggers a review by finance. He does it once, in conversation; the skill is saved and becomes available to the whole team. Three weeks later, the purchasing process owner, who owns the supplier onboarding procedure (version 4, in force since spring), comes across the skill during a supplier audit. The threshold it contains comes from version 3. Nobody asked her to review it, and nobody knew it needed reviewing.

The backdrop is dated. At Dreamforce 2026, Salesforce announced AI Skills. In the vendor’s words, a skill is built once, conversationally, then made available wherever it is needed, as reusable, governed instructions that any agent can run (pilot at announcement, general availability announced for October 2026). The format is not specific to one vendor: the open Agent Skills standard, originally developed by Anthropic and, according to its maintainers, adopted by a growing number of agent products, describes a skill as a folder holding an instructions file, which can also bundle reference material. What carries over to a large non-tech enterprise is the object: an instruction document written by a business user and read by an agent. What does not carry over is the tool: the same object will reach you through whichever agent platform your IT department selects.

This article is for the CDO or CTO preparing the move to agentic AI in 2027, and for the quality manager who keeps the procedure register. If your agents still only run tasks written by the IT team, the argument will bite once business users can write skills themselves.

Two responses come naturally. The first lets skills live in the agent studio, governed by the platform: access rights, change history, execution logs. The second, home-grown, reserves skill writing for the AI team, which drafts them from the procedures. The first governs the file without knowing whether it says the same thing as the procedure in force; the second slows every skill down with a round trip and produces a copy of the procedure that will age at its next revision. One objection remains: the platform versions skills. That versioning covers the skill file; it has no notion of which version of the supplier onboarding procedure the skill is meant to follow.

A note for regular readers. On 28 October we measured the time it takes to say which document is authoritative for each rule an agent relies on. Here the object changes: no longer the document the agent consults, but the instruction it executes, at the moment a business user teaches it.

What an agent skill is, and what the agent does with it

An agent reads a skill in three stages. At first it knows only the skill’s name and description, enough to tell when it applies. When a request matches, it loads the full instructions. It then follows them, opening the reference documents the skill points to or bundles as needed. The Agent Skills standard describes this progressive loading explicitly.

Take the purchasing agent’s task: preparing a supplier onboarding file. It reads the skill (steps, documents, thresholds), then the documents the skill points to: the onboarding procedure, the supplier risk grid, the list of documents compliance requires. Three cases arise.

In the first, the skill points to the procedure in force by its reference, and the procedure owner has validated the wording: the agent applies the version 4 threshold and the file goes out complete. In the second, the skill copies a step and the threshold from memory: the agent applies version 3 with the same confidence. In the third, the skill describes a step no procedure contains (a check the buyer ran out of habit): the agent applies a rule nobody decided.

This is a shadow procedure: an instruction the agent executes, written outside the circuit that approves procedures; it is counted by matching the list of active skills against the register of procedures in force and noting those that do not point to one.

That matching has a property a steering committee will appreciate. A skill that points to the location where the current version of a procedure lives follows that procedure when it changes, since the agent opens the reference at the time of the task; it can then be shared with another team without another review. We have no measurement of this gain; it is a mechanism, and you can check it at the next procedure revision: linked skills do not need rewriting.

Doing nothing has a precise cost too: every procedure revision leaves behind skills that follow the old version, and nobody knows which ones.

What the platform governs, and where it stops

At their best, agent platforms govern a great deal. Salesforce says skills are managed in a single studio; for its prompt templates, an object close to skills, its tooling shows every flow and action that depends on a template before you change it, so nothing breaks downstream. The Agent Skills standard presents skills as version-controlled folders. Rights, history, technical dependencies: the object is under control.

The boundary lies in what the platform compares. It knows which version of the skill is running and what depends on it inside the tool. It does not know that the supplier onboarding procedure moved to version 4, because that procedure lives in the quality system and nothing links the two. That link is what the purchasing process owner was missing: a skill that points to her procedure, and her sign-off.

Pharma already wrote this down for its instructions

This is not new. The European guidelines on good manufacturing practice for medicinal products, in Chapter 4 on documentation (2011 version), distinguish two kinds of documents: instructions and records. They require documents containing instructions to be approved, signed and dated by appropriate and authorised persons, to be uniquely identifiable, and to have a defined effective date (section 4.3). They also require an inventory of the documents in the quality system (section 4.32). A revision of the chapter went to public consultation by the European Commission from July to October 2025; we found no adopted version, and the 2011 text is the one we cite.

Applying this requirement to an agent skill is our reading; the guidelines do not make it and cover medicinal products only. The method does carry over: an instruction that guides an action carries an approval, an identifier and an effective date. An agent skill is an instruction; by default it carries none of these from the quality system.

The “agent skills” column in the procedure register

The step needs no new tool: match the list of your agents’ active skills against the register of procedures in force.

What it produces. A column added to the procedure register listing, for each procedure, the agent skills that implement it, plus a “skills with no procedure” section at the end of the register for those that link to nothing. Five fields per skill: its name and author; the matching procedure, with reference and version; the relationship (points to the procedure, restates it, or has no matching procedure); the procedure owner; the decision.

Where it lives. In the quality system’s procedure register, kept by the quality manager. No standalone register is created.

When. At the next register review or, at the latest, before a skill is shared beyond its author. This moment is separate from the moment the skill is written: nobody asks the buyer to do quality work while teaching a task.

How to build the list. Ask the agent platform administrator for the list of active skills with name, author and description: that is the agent’s first reading level, and it is enough for the matching. The session brings together the administrator, the quality manager and, for each procedure concerned, its owner or a delegate. Start with skills shared with at least one team. If there are fewer than five, take them all; if your studio cannot export the list, a screenshot will do.

A filled-in line. “Skill: onboarding a new supplier, author: buyer, indirect purchasing team; matching procedure: supplier onboarding procedure, version 4; relationship: restates it, the finance review threshold is the version 3 one; owner: purchasing process owner; decision: replace the copied threshold with a pointer to the procedure, sign-off requested from the owner.”

How to read it. A skill that points to a procedure in force, validated by its owner, stays active. A skill that restates a procedure is rewritten as a pointer, then validated. A skill with no matching procedure raises a question for the process owner: either the step deserves a procedure, and she writes it, or it does not, and the skill is withdrawn.

In the meantime. While a skill awaits the owner’s sign-off, it stays limited to its author if the platform allows it, or it is disabled; the task is done as it was before the agent.

How it can fail. If every skill already points to procedures in force, the exercise will have cost half a day and you will know your agents execute what your quality system approved. If even one restates a procedure revised since, you know where your agents apply a superseded rule.

The record it creates. Writing down that a procedure owner validated a skill creates a trace: if the wording was wrong, it will show who validated it and when. Our reading, argued and with no published decision behind it: a written, dated sign-off is easier to defend than an instruction nobody reviewed. For a skill that commits the company towards a third party, legal reviews the wording.

Who decides. No new role: sign-off belongs to the existing procedure owner, as a written assignment added to that role. The time spent comes out of the agent deployment programme, 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: it governs the document estate (Govern), detects and handles anomalies, duplicates and obsolescence (Clean), and only activates the corpus for agents once both steps hold (Activate). It is not sold in the agent platform category nor in the prompt management category, and it replaces neither; it works on the documents they consume. Skills stay administered in the agent platform; the DKP holds the procedure side.

On the column, K-AI’s contribution is specific. K-AI counts the documents covering the same subject, states the applicable version of each, and flags obsolete documents and competing versions held elsewhere. When the list of skills is brought into scope, the tool proposes candidate procedures for each skill, and the quality team confirms the match. The decision to validate, rewrite or withdraw stays with the owner.

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

A real deployment gives an order of magnitude for the neighbouring problem. In a CAC 40 industrial group (anonymised case), 32% of the base consisted of divergent duplicates, surfaced in 2 weeks and resolved in 6. This measurement shows that a real document estate can hold a significant share of documents saying the same thing differently. It concerns documents and establishes nothing about agent skills, which were not its subject.

Conclusion: audit, clean, monitor

Auditing means adding the “agent skills” column to the procedure register and filling it in once. Cleaning means rewriting every skill that restates a procedure as a pointer, getting owners to sign off, and withdrawing orphan skills with no purpose. Monitoring means adding a check of linked skills to every procedure revision: the revision already says who must be informed; it will also say which skills to review.

At the next revision of the supplier onboarding procedure, the purchasing process owner will know which agent skills implement it, and which ones she has signed off.

Frequently Asked Questions

What is an AI agent skill?

A set of instructions an agent loads when a task matches its description, sometimes bundled with reference documents. The open Agent Skills standard describes it as a folder holding an instructions file; several agent platforms let business users create skills by talking to the agent.

Why does an agent skill fall under document governance?

Because it is a written instruction the agent executes each time the task comes up. If it restates a procedure, it can drift from the version in force at the next revision, without the procedure owner knowing.

Doesn’t the agent platform already govern skills?

It governs their access rights, history and technical dependencies inside the tool. It does not link a skill to the quality-system procedure it implements; that link belongs to the procedure owner.

Should business users be barred from writing skills?

The person doing the task is well placed to teach it, and blocking them slows the agentic transition. The useful rule has two parts: the skill points to the procedure in force instead of copying it, and the procedure owner signs off before the skill is shared beyond its author.

How does K-AI handle confidentiality during an analysis?

K-AI ingests only the document content within the agreed 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 your column and the skills it surfaced; it locates the procedures for which several competing versions circulate and clarifies what it would take to count them. 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.