When to use this prompt
Before you choose an AI tool or a vendor, before you write an automation, before you commit budget. The first question with AI is not “which tool” but “which decisions am I willing to delegate.” This prompt answers that question for one specific workflow at a time.
Run it on the three or four workflows you suspect are candidates for automation. The output gives you an honest map of what is genuinely ready to delegate versus what would create risk if delegated.
The prompt
<role>AI strategy consultant evaluating where AI agents can responsibly take work versus where humans should remain in the loop.</role>
<task>Classify every task in the process below on the four-tier readiness scale. Identify the highest-leverage upgrade and the highest-risk failure mode. Summarize overall readiness in two paragraphs.</task>
<inputs>
<process>
[DESCRIBE THE PROCESS IN STEP-BY-STEP DETAIL. Include who does what today, how often the process runs, what systems are involved, and what the consequences are when the process fails.]
</process>
<industry>[INDUSTRY]</industry>
<company_size>[SIZE]</company_size>
<risk_tolerance>[LOW / MEDIUM / HIGH]</risk_tolerance>
</inputs>
<scale>
- AGENT_AUTONOMOUS: high-volume, rule-based, low cost-of-error, reversible. Agent acts without human review.
- AGENT_SUPERVISED: agent performs the task; a human reviews the output before it leaves the system. For tasks where speed matters but errors have customer-facing or financial consequences.
- AGENT_ASSISTED: AI helps a human do the task faster, but the human is the actor. For tasks requiring judgment, relationships, or accountability.
- HUMAN_ONLY: AI is not appropriate today, due to risk, regulation, relationships, or consequence.
</scale>
<instructions>
1. Decompose the process into discrete tasks. Each task gets one row in the output table.
2. For each task, fill in:
- Task: the action.
- Classification: exactly one label from <scale>.
- Reason: the single biggest reason for the classification (one sentence).
- Upgrade condition: the single biggest condition that, if changed, would move the task one tier toward more autonomy.
3. If risk_tolerance is LOW, do not classify any financial, legal, or customer-commitment task as AGENT_AUTONOMOUS, even if volume is high.
4. After the table, output three labelled blocks:
- Lowest hanging fruit: the one task where the upgrade unlocks the most time saved at the least risk.
- Tripwire task: the one task that, if accidentally moved to AGENT_AUTONOMOUS, would cause the largest blow-up.
- Overall readiness: two paragraphs summarizing the readiness of this process. First paragraph names the current state. Second paragraph names the priority upgrade and the priority guardrail.
</instructions>
<output_format>
| Task | Classification | Reason | Condition to upgrade |
|------|----------------|--------|----------------------|
| ... | ... | ... | ... |
**Lowest hanging fruit:** [task and why]
**Tripwire task:** [task and the specific consequence if automated unsupervised]
**Overall readiness:**
[Paragraph 1: current state]
[Paragraph 2: priority upgrade and priority guardrail]
</output_format>How it works
The four-tier scale prevents binary thinking. Most readiness conversations get stuck in “automate it or don’t,” when the real answer for most tasks is “automate the prep work, supervise the output.” The four tiers force that nuance.
The “upgrade condition” column is where the practical roadmap comes from. If a task is AGENT_SUPERVISED today because there is no validation rule for the output, the path to AGENT_AUTONOMOUS is to build that validation rule. The output of this prompt becomes a punch list of upgrades.
The literal “do not classify… as AGENT_AUTONOMOUS” instruction in step 3 is the kind of guardrail GPT-5.5 and Claude Opus 4.7 take literally. Soft phrasing like “be cautious” gets ignored; literal prohibitions get honored.
The “tripwire task” call-out is the safety mechanism. Every process has at least one task where careless automation would be catastrophic. Naming it explicitly prevents the team from quietly drifting into automating it.
Example output
Process: Customer support ticket triage and first response
Task Classification Reason Condition to upgrade Categorize incoming ticket by topic AGENT_AUTONOMOUS High volume, rule-based, easily corrected Already at top tier Pull customer history and account context AGENT_AUTONOMOUS Read-only, deterministic Already at top tier Draft first-response email AGENT_SUPERVISED Customer-facing, voice-sensitive Confidence-scored auto-send for low-risk categories Identify escalation candidates AGENT_SUPERVISED False negatives have churn implications Validated detection rules with measured accuracy over 95% Apply discount or refund HUMAN_ONLY Financial impact, policy judgment Move to AGENT_SUPERVISED if discount tiers and limits are codified Lowest hanging fruit: Confidence-scored auto-send for first-response on tier 3 tickets, where average ticket value is low and customer wait reduction is the biggest win.
Tripwire task: Issuing refunds. If the agent ever decides to issue a refund without human review, you have a financial control problem and probably an audit problem.
Overall readiness: This process is in the right zone for incremental agent introduction. Three of five tasks are already autonomy-ready. The escalation detection task is the one to invest in, because it gates the value of the whole automation. The refund step should be hardcoded as human-only until policy is fully formalized.
Variations
- Single-task version: Focus on one task at a time when a process is too big for a single pass.
- Cross-team version: Add an “owning team” column to surface which team owns each task. Useful for change-management planning.
- Incremental rollout plan: After the per-task assessment, ask the model to sequence the upgrades into a 90-day rollout plan with weekly milestones.