Before the agent, look at the work.

Imagine a support leader looking at a growing queue. Some requests are repetitive. Some need judgement. A demo suggests an agent could handle the whole thing.

The next question is not “which model?” It is “what, exactly, are we about to delegate?”

This is a fictional situation, but a useful starting point. A team can understand its workflow better before it buys anything or connects an agent to a real system.

The Agent-Ready framework

I want this lab to make the operator’s questions visible. The Agent-Ready Audit turns ten of those questions into a short self-assessment. It is an original educational framework, not a certification or a validated prediction of project success.

Define the work

1. Outcome. Can you describe a successful result and check it? “Improve service” is a direction. “Route this request to the right team with the evidence attached” is a task you can test.

2. Pattern. Does enough of the work repeat? Document the ordinary path and the exceptions. A workflow does not need to be perfectly uniform, but it needs a boundary.

3. Data. Is the information accurate enough, accessible and permitted for this use? Access to a record is not permission to give its contents to any model or service.

Bound the authority

4. Authority. What may the agent suggest, prepare or commit? A draft refund and a paid refund are different actions. Name the approval point.

5. Recovery. How would you stop a run, contain damage and recover? Some actions cannot truly be undone. Keep those behind appropriate human approval rather than pretending that a rollback solves everything.

6. Evaluation. Do you have representative cases and pass criteria? Include missing information, conflicting instructions and failures—not only examples that make the demo look good.

7. Handoff. Where does uncertainty go? “Ask a human” is incomplete until the receiving person, the evidence and the response expectation are clear.

Own the operation

8. Visibility. Could someone reconstruct the input reference, action, approval and outcome? You need an operational record, not a store of private reasoning or unnecessary sensitive data.

9. Economics. What does the completed task cost, including retries and human review? Define a baseline and enforced spending, duration and retry limits.

10. Ownership. Who accepts responsibility for the workflow and has the authority to stop it? A sponsor is useful. An accountable operator is essential.

Why four questions are gates

Data permission, authority, recovery and ownership cannot be averaged away. An excellent test set does not give you permission to use restricted data. A cheap model does not make an irreversible action acceptable.

In the Audit, any of these four essential boundaries below the strongest answer leads to start with the foundations, whatever the total score. That is a deliberate conservative rule, not a statement that all other risks have been assessed.

How the score works

Each answer is worth 0, 1 or 2 points. “Not sure yet” scores 0 because missing evidence is something to resolve, not a pass. Ten answers give up to 20 points; multiplying by five produces a preparation score out of 100.

• Any essential gate below 2, or fewer than 10 total points: start with the foundations.
• All four gates at 2 and 10–15 points: keep a human reviewing actions while strengthening the remaining checks.
• All four gates at 2 and 16–20 points: explore a small, supervised pilot.

These thresholds are editorial choices. They have not been statistically validated, and a score is not a probability of success. Even 100/100 describes what someone selected, not what their systems have demonstrated.

Try this

Take the gaps to the people who own the workflow. Replace assumptions with evidence. Pick a narrow task, test with synthetic or appropriately permitted data, and agree when to stop.

You do not need a high score to learn something. A clear reason not to delegate an action yet is a useful outcome too.

Assess one workflow, or explore when an agent should ask.