How to map workflows before you build AI agents

Direct answer

Map the real workflow before you write prompts or wire tools. A useful map names every step, the tool it lives in, who owns it today, what "done" means, and the cost if that step fails. Without that map, agents automate myths.

Why mapping comes first

Teams often start with a model or a chatbot UI. Production agents start with work: intake, triage, draft, approve, send, log. If those steps are fuzzy, the agent will be fuzzy too.

A minimal mapping session (half day)

1. Pick one high-volume workflow (not ten).
2. Walk a real recent case end to end - not the happy path slide.
3. Capture for each step: actor, tool, inputs, outputs, exceptions.
4. Mark irreversible actions (send, pay, change permissions, public post).
5. Score failure cost: low / medium / high for customer, money, trust.
6. Decide which steps an agent may draft, execute, or never touch.

Template fields (copy this)

Common mapping failures

• Mapping the org chart instead of the work
• Skipping exceptions ("VIP clients", "legal holds")
• Assuming one tool when work hops across five
• Forgetting the human who currently holds judgment

After the map

Only then choose agent scope: one production path, clear gates, and a two-to-four week measurement window (cycle time, error rate, rework). See also human approval gates.

How Dali runs discovery

Dali starts with process discovery inside tools the team already uses, then designs agent paths with approval boundaries. Start from solutions.