Data readiness
The grounding data assessed honestly: duplicates, stale records, fields that three teams fill in differently, and knowledge articles contradicting each other.
This is where most of the effort goes and where most pilots did not. An agent grounded in inconsistent data does not fail loudly; it answers confidently and wrongly.
A grounding set you would stand behind.
Grounding and retrieval
What the agent can see, from which objects and knowledge sources, with retrieval scoped so answers come from the records that should inform them.
Over-broad grounding is the usual cause of an agent being technically right and practically useless.
Answers traceable to a source.
Permissions and disclosure
The agent's access modelled deliberately, tested against roles, and reviewed as a security change rather than as a configuration detail.
An agent inherits the permission model it runs under. Years of accumulated access debt becomes visible the moment it starts answering questions.
An agent that cannot say what it should not.
Topics, actions and handoff
What the agent handles, what it does, and the point at which it stops and hands to a person, with that boundary chosen deliberately rather than discovered.
The handoff is a design decision. An agent that will not let go is worse than no agent.
A boundary everyone agreed on.
Testing behaviour
Tested against real cases, including the ones it should decline, the ambiguous ones and the ones where the right answer is to escalate.
Testing only the happy path tells you what the demo does, not what the deployment will do.
Behaviour you have evidence for.
Consumption and cost
Usage measured at pilot scale and projected before rollout, so the economics are understood while the decision is still reversible.
Consumption-based pricing rewards measuring first. A pilot that nobody costed is a rollout nobody can budget.
A number you can take to finance.