← All steps
Part B · Step 1

Discovery (workflow first, tool second)

Owner
Department Champion + the person requesting the agent.
Input
Someone in the company says "I want an AI agent for X."

Owner: Department Champion + the person requesting the agent. Input: Someone in the company says "I want an AI agent for X." Sub-steps:

  1. Find out what they actually do today. Sit with them. Watch the workflow. Take notes.
  2. Document the current workflow as-is. What systems? What inputs? What decisions? What edge cases? What's the failure mode of the workflow today?
  3. Identify the pain point. Is it volume? Boredom? Errors? Speed? Cost? Skill bottleneck? You need to know exactly what AI is supposed to relieve.
  4. Volume check. How many times a week / month does this happen? If <10/month, AI is probably not the answer (rules / templates / API integration may be).
  5. Apply the "is this even an AI problem?" filter:
    • Could a better API integration solve it? → not an AI problem.
    • Could a deterministic rule / template solve it? → not an AI problem.
    • Does it require probabilistic reasoning over unstructured / ambiguous input? → maybe an AI problem.
  6. Apply the "is this an agent or just a model call?" filter:
    • Does it need to chain actions across systems? → agent.
    • Does it need to use tools? → agent.
    • Is it a single prompt-response? → not an agent, just an LLM call.
  7. Briefly estimate value: how many hours saved / errors reduced / dollars saved per month.
  8. Briefly estimate risk: does it touch PII? Does it make decisions about people? Can it act autonomously?

Output / gate criteria:

  • A 1-paragraph workflow description.
  • A 1-line value statement.
  • A 1-line risk red-flag list.
  • A "this is genuinely a candidate for agentic AI" / "no, route elsewhere" recommendation from the Champion.

Decision branches:

  • Not a real AI use case → route to the appropriate team (process improvement, IT integration, training) and close.
  • Real AI use case → go to Step 2.

Skip-this-step risk: You end up automating a broken workflow at higher speed, or building an agent for something a $50 API call could fix.