What does a readiness assessment for AI agents measure?

It measures whether a team understands the ideas a governed agent will put in front of them, scored concept by concept rather than as one pass mark. It shows where the team is strong, where the gaps are, and which agent to start with. It certifies nobody; it shows where training should begin.

What does it test?

The concepts every person who works with an agent meets on day one. In practice that means questions on:

• Why agents must be governed. What can go wrong when an agent acts without bounds.
• The decision chain. How each action meets several bounds at once, and why the lowest one wins.
• Depth and certification. What each depth allows, why depth is earned rather than configured, and what a lapsed certificate means.
• Verdicts. The difference between allowed, held, refused and withheld.
• Provenance. Telling sealed, signed, declared and external facts apart, and why an external claim cannot authorise a payment.
• The workspace and the record. Where held actions appear and how a decision is reconstructed.
• Ownership, data and trust. Who owns the data, the agent and the decision.

The best assessments set the questions in the team's own industry. A claims team answers questions about held claims; a payments team about held transfers. Generic questions measure general knowledge; industry questions measure whether people can apply it.

Why score by concept rather than overall?

Because the overall score hides the gap that matters. A team can average well while nearly everyone misreads provenance. That team will approve emailed instructions on its first week with an agent, whatever its average.

A report by concept shows each idea separately. If the team is strong on depth and weak on provenance, training starts with provenance. If it is strong on both but weak on the decision chain, people will misread refusals, so that comes first.

How does the report guide the first agent?

By matching the agent's demands to the team's strengths. Every agent leans on certain concepts more than others:

A team strong on verdicts but weak on provenance should start with a drafting agent on low-risk correspondence, not a document-matching agent. The report should say so, and name the first agent plainly.

What does it not measure?

Three things, each covered elsewhere. It does not measure whether an individual can decide held actions reliably; that takes a practicum of real decisions. It does not measure whether the agent is ready; that is certification on the business's own cases. And it does not measure whether systems, data and processes are ready; that is operational readiness. The general idea of a readiness assessment covers all three; for agents, the team assessment is the people part.

An example

Take a hypothetical bank's onboarding team of 22. The assessment shows strong scores on depth and verdicts, middling on the decision chain, and weak on provenance: most people treat an uploaded utility bill as equal to a registry record. The report recommends starting with an agent that drafts welcome letters and document requests at depth 2, while the team trains on provenance, before any agent that checks identity documents. Six weeks later, a reassessment shows provenance scores up, and the identity agent enters shadow. The full method for a bank is in how to run an AI agent readiness assessment in a bank; a fintech, with smaller teams and faster change, follows its own version. The readiness assessment report template shows how to lay out the results.

Where the Academy fits

Pryme Intelligence Academy offers a free readiness assessment for teams of up to 25. The team takes the Foundations test set in its own industry, and the report shows scores by concept, the gaps, and the first agent to start with.