Embedding agents in workflows

The pattern

You embed an agent into a workflow at one of three points:

1. Pre-action drafting — agent drafts the proposed action; a human reviews and submits.
2. In-flight enrichment — the workflow calls the agent for context (risk score, similar-case lookup, recommendation) at a decision point.
3. Post-action analysis — agent reviews completed actions and surfaces anomalies for follow-up.

For most novice deployments, (1) Pre-action drafting is the right entry point. It's the lowest-risk pattern: the agent does the work, the human commits it.

What "good" looks like

• The agent's output is shown explicitly to the human reviewer ("AI suggests: …"), not silently substituted for a draft they would have written.
• The reviewer can edit, accept, or reject with equal effort. Don't make rejection the painful path.
• The audit trail captures what the agent suggested, what the human approved, and the model version at the time.
• There's a clean off-switch — the workflow continues to work if the agent is disabled, just slower.

Common BPM patterns to embed agents in

• Invoice approval — agent drafts the GL coding; reviewer approves.
• Expense audit — agent flags policy violations; auditor decides.
• Performance review — agent drafts narrative; manager edits before submitting.
• Service case escalation — agent recommends route; service manager confirms.

Common mistakes

• Hiding the AI. If users don't know the agent is in the loop, they can't apply judgment to its output. Disclose it.
• Auto-approval. Skipping the human step for "low-risk" cases is where most agent-driven incidents start.
• No rollback path. Workflows that require the agent to function — instead of degrading gracefully — are operational risk.