Stage 1 — Read
The AI retrieves real records scoped to the current organization: counts, lists, statuses, history. The requirement here is grounding — every number comes from a query, never from the model's memory of the conversation.
A system that estimates 'about a dozen open jobs' has failed stage one, regardless of how well it writes.
Stage 2 — Decide
- Selecting the right record when a name is ambiguous.
- Choosing which rule applies: an entitlement, a policy, a threshold.
- Deciding to refuse, ask a clarifying question, or escalate to a person.
- Sequencing multiple actions where one depends on the result of the last.
Stage 3 — Create
Creation writes internal truth: a lead, an appointment, an estimate, a task, a note. The safeguards that matter are duplicate detection, required-field validation, correct ownership and tenant scoping.
Creation is reversible in principle, which is why it is a lower risk tier than execution — but a duplicated customer record still costs real money downstream.
Stage 4 — Execute
Execution produces effects the business cannot quietly undo: a message sent to a customer, a contract issued for signature, a payment processed, a crew dispatched.
This is where confirmation gates, external-send controls and audit evidence stop being good practice and become the price of entry.
How to score a capability honestly
- Score per action, not per product.
- Downgrade any stage you cannot demonstrate live on real data.
- Treat a manifest entry as evidence of intent, not of execution.
- Record the date and the tenant the demonstration ran in.