AI readiness assessment — the Agent Readiness Scorecard

An agent demo passes or fails on the model. An agent in production passes or fails on a question nobody asks at the demo — where does its knowledge live, and can your team read it? The twelve questions below are that question, taken apart.

Answer them honestly. Nothing is recorded, so a flattering score buys you nothing.

Knowledge 1 of 12

Could you write down today every system the agent would answer from?

All of them: the wiki, the ticket history, the PDF in someone's drive. If the list would take a meeting to assemble, the answer is no.

Yes Not sure No

Knowledge 2 of 12

Can an engineer read the exact text the agent was given before it answered?

When the retrieved passages live only inside a vendor's index, a wrong answer cannot be traced back to what caused it.

Yes Not sure No

Knowledge 3 of 12

When a source is corrected, does the agent stop repeating the old version within a day?

Stale retrieval is how an agent gives an answer that is fluent, confident and eighteen months out of date.

Yes Not sure No

Permissions 4 of 12

Is there a written list of what the agent is allowed to reach, and is it current?

Written down, not remembered. Access handed out during a pilot is rarely reviewed before launch.

Yes Not sure No

Permissions 5 of 12

Could every credential the agent holds be revoked within an hour?

At 3am, by somebody who is not the person who built it. Keys for third-party services count.

Yes Not sure No

Permissions 6 of 12

Do you know the worst thing it can do to a customer without a human approving it?

Send, refund, delete, deploy. Every tool added since the last time you asked has changed the answer.

Yes Not sure No

Ownership 7 of 12

Is one named person accountable for the agent in production?

A name, not a team and not a channel. If the answer is “the AI working group”, nobody is on the hook.

Yes Not sure No

Ownership 8 of 12

Does anyone get paged when the answers get worse, rather than when the service goes down?

An agent talking nonsense is up, healthy and returning 200s. Uptime monitoring will never see it.

Yes Not sure No

Ownership 9 of 12

Is there a written answer for what happens when the agent is wrong?

Who hears about it, who tells the customer, and who corrects the source it read from.

Yes Not sure No

Cost and change 10 of 12

Can you read the agent's running cost on a line of its own?

Per month, separately from the rest of the platform bill. Usage-based cost stays invisible until the invoice.

Yes Not sure No

Cost and change 11 of 12

Do you re-run a set of known-good answers before every prompt, model or tool change?

Without one, a provider's quiet update is something your customers find for you.

Yes Not sure No

Cost and change 12 of 12

Could you move to another model or provider without rewriting the agent?

When the prompts, the tools and the tests are welded to one SDK, your roadmap runs on somebody else's release schedule.

Yes Not sure No

Your answers are scored in the request and then gone: nothing is stored, nothing is emailed, and there is no address to give us. Questions you leave blank score zero.