The first demo of an AI agent is usually impressive. The second week in production is where teams learn what the demo hid: the agent is only as reliable as the records, knowledge and permissions it can reach. Before designing topics and actions, look hard at the foundation.
Start with a narrow, valuable job
Pick one job with a clear definition of done, such as answering order status questions, triaging inbound cases or preparing a service summary. Narrow scope makes it possible to measure accuracy and to explain the agent's behavior to the people who supervise it.
Audit the records the agent will read
List every object and field the agent needs, then check completeness and consistency. Missing contact roles, free-text status fields and duplicate accounts produce confident but wrong answers. Fixing a handful of high-value fields beats a broad cleanup project.
Unify without copying everything
Data 360 brings customer data together from Salesforce and other systems, and Zero Copy federation lets agents use data that stays in platforms such as Snowflake or Databricks. With Informatica now part of Salesforce, data quality, cataloging and master data management fit into the same foundation.
Treat knowledge as a product
Articles written years ago for internal staff rarely suit an AI that will quote them to customers. Assign owners, retire outdated content and structure articles so each one answers one question well.
Design permissions deliberately
An agent should see only what the person it serves is allowed to see. Review the running user, sharing model and field-level security for every action. Least privilege is a design input, not an afterthought.
Measure before and after
Capture baseline metrics (handle time, resolution rate, satisfaction) before launch. Without a baseline, it is impossible to show whether the agent is helping or simply moving work around.