An AI agent works best when the workflow is clear, the data is reachable, and the success metric is specific. Run a workflow through these four readiness groups before investing in a build. A failed checkbox isn't a blocker — it's a to-do list for what to fix first.
Workflow readiness
- The workflow happens often enough to matter
- The current process is documented or can be observed
- The painful steps are clear
- There is a known user or team for the agent
- The agent's output can be reviewed or measured
Common mistake: building for a workflow that's rare or one-off. If it doesn't happen often, the agent won't pay back the build.
Done right when: every box is checked. If the workflow is still vague, start with discovery instead of implementation — that's the fix, not a reason to push ahead.
Data readiness
- The needed data sources are known
- Data is reasonably fresh
- Access permissions can be defined
- Important documents or records have owners
- The agent can cite or trace where answers came from
An agent is only as useful as the context it can safely reach.
Common mistake: assuming data is reachable because it exists. "It's in the CRM somewhere" is not reachable, fresh, or owned.
Done right when: you can name the source for every answer the agent will give, and someone owns each one.
Action readiness
- You know whether the agent should answer, draft, or execute
- Risky actions have approval steps
- Errors have an escalation path
- The team knows what the agent is allowed to do
Common mistake: letting the agent execute before anyone's defined what "wrong" costs. Match the action level to the blast radius.
Done right when: every action the agent can take is one you'd be comfortable explaining after it goes wrong. If the workflow is business-critical, start with lower-risk actions (answer/draft) before execute.
Operations readiness
- Someone owns quality after launch
- Success metrics are defined
- Monitoring requirements are clear
- There is a plan for improvements after real usage
The agent does not need to be perfect on day one. It does need a named owner and a path to get better.
Common mistake: treating launch as the finish line. An agent with no owner degrades quietly as sources drift.
Done right when: a specific person owns post-launch quality, and you've defined what "better next month" measures.