Agentic AI does not read your content. It acts on it.

On 4 June 2026 the Digital Transformation Agency published an agentic AI addendum to its AI technical standard. It speaks to the people who build AI systems. It names website information as one of the things those systems take as an input.

That is your website, listed in a technical standard as a data source that software executes against.

A chatbot answers a question. An agent takes a step: it routes a case, lodges a form, or decides not to escalate. The difference is what each one does with a sentence that is not quite clear.

Someone reading an eligibility page at 11pm slows down when the wording gets slippery. They scroll back, ring the number in the morning, or give up. It costs them time and costs you a call. It stays a comprehension problem.

An agent does not slow down. It resolves the ambiguity and takes the next step. You find out which way it resolved after the action has happened.

The test your content has to pass is no longer whether it could be misread. It is whether it could be executed.

This is not a technical job

You do not need to know how an agent works to do anything about this. Nothing below needs a budget, a tool, or a conversation with an engineer.

It does need somebody to go looking, and that is the awkward part. The addendum puts accountability on the agency running the agent. In most agencies that is not the team that owns the website.

The accountability sits with people who cannot see the content problem, and the content sits with people nobody has asked.

Read every paragraph on its own

Take your 10 highest-volume task pages. Read each paragraph as though it arrived with nothing around it, because that is how it will arrive. Retrieval pulls a passage, not a page.

Mark every sentence that stops being true without its neighbours. The usual failure is a condition sitting in an introduction while the instruction sits in a list further down.

A person scrolls back. An agent does not know there is anything to scroll back to.

If a sentence is not true on its own, it is not ready to be acted on.

Turn your hedges into conditions

Find every 'generally', 'in most cases' and 'you may be eligible' on your eligibility and process pages. Government content hedges because hedging manages legal risk. That is a real constraint, not a writing failure.

Then sort them. Where the condition is knowable, write it as a condition: what has to be true, and what happens when it is not. Where it is not knowable, name what decides it and who makes that call.

A hedge is a request for a human to apply judgement, and an agent has none to apply. It resolves the sentence anyway, in whichever direction the surrounding text leans.

This is the plain language answer as well as the machine-readable one. Somebody was going to ask for it.

Decide which page is authoritative

Large government sites state their important rules more than once, in different words, on pages owned by different teams. Nobody notices, because both versions are defensible.

Find them. Decide which page is the source, and make the others link to it instead of restating it.

A person picks the version that looks official. An agent picks whichever one the retrieval scored higher, and it does not tell you there was a second one.

Two defensible versions of the same rule is not redundancy. It is a decision nobody made, left for a machine to make.

Write down the irreversible actions

On this point the addendum is firm. A person has to be able to step in on any action that is irreversible or high risk. Somebody has to decide which actions qualify. That decision gets made by reading the content that describes them.

Somewhere on your site there is a declaration you cannot withdraw once you submit it. Somewhere else there is an election that locks a payment rate for 12 months. That is where the human gate belongs.

The register of irreversible actions is a content artefact. Content people are the only ones who have read every page that describes one.

Nobody is going to hand a content team that job. Somebody has to offer it, before anyone designs the workflow around the assumption that the source material is unambiguous.

What good looks like

• Every paragraph on your top task pages is true without the paragraph above it.

• Every eligibility statement names either a condition or the person who decides.

• Each rule has one authoritative page, and the others link to it.

• You can produce a list of your irreversible published processes without doing new research.

• Somebody outside the content team has asked you for that list.

Most of what the addendum says about ambiguity is advice rather than obligation. Do not read that as time.

None of this is new content design. It is the same work, held to a standard where you cannot count on a charitable reader.

So here is the question for your next content meeting. If an agent acted on your busiest page tomorrow, which sentence would you least want it to act on?

Photo by Philip Oroni on Unsplash.