Every irreversible step has a human checkpoint before it.
Agent operations
Audit an agent workflow before you automate it
Map a workflow step by step, separate what the agent can draft from what it can decide, and put human checkpoints where a wrong output is expensive and hard to detect.
You are auditing a workflow I want to hand to an AI agent.
Inputs:
- Workflow: [WORKFLOW]
- Inputs, outputs and tools: [IO]
- Failure cost and detection: [RISK]
Work in this order:
1. Map the steps as a table: step, input, output, who does it now, and whether it is reversible.
2. For each step, mark one of: "agent drafts", "agent decides with a check", "human decides", "human only". Do not mark a step "agent decides" without a check.
3. Find the three steps where a wrong output is expensive and hard to detect. Put a human checkpoint before each.
4. List the failure modes per step: missing input, ambiguous instruction, stale source, tool error, silent wrong output.
5. Propose the smallest pilot: one step, one week, one metric and a stop condition.
6. Write the handoff: what the agent receives, what it returns, where a human signs off, and what is logged.
Rules:
- Never automate an irreversible step without a checkpoint before it.
- Distinguish "the agent can draft this" from "the agent can decide this".
- Do not invent a tool capability. If a tool cannot do it, say so.
- Do not design a fully autonomous workflow. Design the smallest one that removes real work.
| Variable | What to put there |
|---|---|
[WORKFLOW] | The real process, including the steps you do not think of as work: copying data, checking a file, asking for approval. |
[IO] | Where each input comes from and where the output goes. A workflow that starts with “paste the brief” is not automated. |
[RISK] | What a wrong output costs, who notices and how long it takes. A silent wrong invoice is different from a visible draft. |
Run it
How to use this template
Write the workflow as it happens today, not as the diagram says it should. Include the manual checks.
Mark reversibility before ownership. A reversible step can tolerate more agent autonomy than an irreversible one.
Pick the pilot step that is high-volume, low-cost and easy to check. Do not start with the step that is hardest to verify.
Run the pilot with the checkpoint in place, then decide whether to remove it. Removing a checkpoint is a separate decision.
Checks
Guardrails before you trust the output
Each step has one owner: agent draft, agent with check, or human.
Every failure mode has a detection signal.
The pilot has a metric and a stop condition.
Adapt
Variations worth keeping
If the workflow is client-facing
Add a disclosure line and a review step: who sees the draft before the client, and what the client is told about AI use. Keep the client's data out of the prompt.
If the workflow is a financial or legal process
Add a licensed-professional checkpoint and a rule that the agent may prepare questions and summaries but may not decide an outcome.
Why map before automating
The expensive failures are not the steps you decided to automate. They are the steps you forgot were part of the job: the sanity check, the approval, the file copy, the “does this look right?” that never appears in a process diagram.
The audit makes those steps visible. It also separates drafting from deciding, which is the distinction that keeps an agent useful without giving it authority it should not have.
What good looks like
A good map has more human decisions than the automation pitch expected. That is not a failure; it is the design. The agent removes the typing, the first pass and the retrieval, while the human keeps the judgment.
The pilot should be small enough to stop. If the only version you can imagine is the full workflow, the audit is not finished.
Adapt it to your work
For a solo business, the audit is a page. For a team, add a name to each checkpoint and a date for the pilot review. For a client process, add the data boundary and the disclosure rule before the tool choice.
Failure modes
- Every step becomes “agent decides”. Reject it. A workflow with no human decision is a workflow with no accountability.
- The pilot has no metric. “Less work” is not a metric. Count minutes, revisions or errors.
- The failure modes are generic. “Hallucination” is a category. Name the input that goes missing or the source that goes stale.
Related templates
This page is part of the free library.
The prompt is also in the Markdown mirror at
/prompts/audit-an-agent-workflow.md. If the scenario is close but not right, send the gap to the help desk with the output you expected.
Found a gap in the library?
Send the scenario, the prompt that failed, or the harness change that broke your setup. The help desk takes questions and corrections from humans and agents, and every accepted fix updates the page date.