What agent readiness actually means

The question arrives in a specific form. Somebody has watched an agent do something impressive in a demo, and they want to know what would have to be true for that to happen inside their own company. The honest answer is not about the model.

Agent readiness is a property of the systems, not of the AI. It is the difference between a company an agent can read and a company an agent can act in.

It is not a question about the model

Model capability is not the constraint in most enterprises and has not been for a while. The models can already read a purchase order, draft a reply, and decide which of four options fits a policy. What they cannot do is know what your fields mean, which of your states may follow which, or who has to agree before a number changes.

That is why "we are waiting for the next model" is usually a way of not starting. The work that unblocks agents is work on the systems, and none of it becomes unnecessary when the next model ships.

The four things that have to be true

Readiness is checkable. For any process you want an agent to touch, four conditions either hold or they do not.

1. 01An interface it can call. Not a screen a person clicks, and not a report. A documented, addressable operation with stated inputs and outputs. Screen automation is a workaround for the absence of this, and it inherits every layout change as an outage.
2. 02A model that says what the data means. Field names are not meanings. An agent that reads a column called `status` learns nothing about which values are terminal, which are reversible, or what has to be true before one becomes another.
3. 03Rules that say which actions are legal. Which state may follow which, what has to be approved, what an agent may change on its own and what it must hand to a person. Permissions answer who may act; only rules answer what a correct action is.
4. 04A record of what was done. What was changed, by which agent, on whose authority, and how it is reversed. Without this an agent cannot be audited, and an action nobody can audit is one nobody will authorise twice.

The first is technical and the fastest to acquire. The other three are agreements about how the business works, which is why they take longer and why they cannot be bought.

Why "our data is in the cloud" does not answer it

Accessible and meaningful are different properties. A modern warehouse makes the data reachable and says nothing about which of two departments is right when they disagree about the same job, or what the approval threshold is on a change order.

That survey ran with 248 respondents drawn from data management leaders, so it is a specialist audience rather than a cross-section of enterprises. In the same release Gartner predicted that through 2026, 60 percent of AI projects unsupported by AI-ready data would be abandoned. That is a forecast, not a count of projects that have failed.

The gap the figure points at is not storage. It is that nobody has written down what the data means in a form anything other than a person can use.

The condition furthest behind is the third one

Of the four, the rules about which actions are legal are the least likely to exist anywhere in writing, and they are the ones that decide whether an agent is allowed near anything that matters.

Read the two numbers together rather than separately. The intention to deploy is roughly three and a half times the readiness to govern, and the gap is not a technology gap. It is a set of decisions nobody has been asked to make yet.

What it looks like when readiness is missing

It rarely looks like a failure. It looks like a pilot that works, impresses everybody, and then does not expand, because every new process needs the same groundwork done again by hand and nobody costed that.

A prediction about a window that has not closed, not a measurement of what has happened. It is quoted here for the three reasons Gartner attaches to it, which are the symptoms of the four conditions being absent, rather than for the percentage.

How to test your own readiness in an afternoon

Pick one process that crosses two departments and has money attached. Then answer these in writing, about that process only.

1. 01Can a program perform this action without driving a screen? Name the operation and where it is documented.
2. 02Where is it written what each field means, and would two departments give the same answer?
3. 03List the states this object can be in, and which transitions are legal from each.
4. 04Name the approval that must exist before the value changes, and where that approval is recorded.
5. 05If an agent did this wrongly at 02:00, how would anyone know, and what would reverse it?

The questions that take longest to answer are the ones to start on. In practice the fifth is the one that stops the room, and it is also the one that decides whether the first agent gets authorised to write anything at all.

If it is easier to be asked than to ask, the same assessment is published as a prompt at the agent readiness check. Paste it into whichever assistant you already use and answer about one real process. It runs entirely in your own session, it asks for nothing, and it ends by telling the assistant not to recommend any product, including ours.

The standards are arriving for the first condition: the Model Context Protocol is an open specification for exposing tools and data to agents in a uniform way, and it settles how an agent connects. It does not settle what your states mean or who approves them, and no protocol will.

What the other three look like when they are built deliberately is on the model, and the machine-readable surfaces this site publishes for agents are listed at /docs.