Rebuild the revenue data model so agents can act on it (L6)
Agents fail on messy data long before they fail on reasoning. This rebuilds definitions, permissions, and access paths so an agent reads and writes the same governed truth a human does. Warehouse plus CRM plus a governed access layer.
The steps
- 01
Agree one definition per revenue object
Tool: dbt
Pipeline, qualified opportunity, active customer, churn. One definition, one owner, one place it is calculated. Agents inherit whatever ambiguity you leave in place. Owner: RevOps plus data. Pitfall: two teams keeping their own version. The agent will pick one and nobody will know which. DoD: a published metric dictionary with named owners.
- 02
Give agents scoped credentials, not admin
Tool: Salesforce
Each agent gets its own service identity with the narrowest read and write scope that lets it do the job. Never share a human admin login. Owner: RevOps plus IT security. Pitfall: an integration user with full write access. One bad loop rewrites thousands of records. DoD: every agent has a named identity and a documented permission scope.
- 03
Log every agent write with a reversal path
Tool: Snowflake
Stamp actor, timestamp, prior value, and reason on agent-written fields, and keep a restore procedure you have actually run in a sandbox. Owner: data engineering. DoD: a tested rollback for at least one agent-written object.
- 04
Publish an access contract for new agents
Tool: Manual
Any new agent request goes through one intake: what data, what writes, what ceiling, who owns it. This is what stops shadow agents. Owner: RevOps. DoD: an intake form exists and the last three agents went through it.
Tools in this playbook
- dbt
- Salesforce
- Snowflake
- Manual
Next playbooks
Unfamiliar terms are defined in the AI and Revenue Dictionary. Related frameworks live in the framework library.