Should This Be an Agent? Single Agent, Workflow or Plain Script Decision

Scores a task you want to automate on variability, tool needs, error tolerance and cost, then recommends a plain script, a fixed LLM workflow, a single agent or multi-agent, with the cheapest first build.

You geta score table, the recommendation with reasons, the cheapest version to build first (steps and tools), what evidence would justify moving up a level, and 3 checks with pass…

The prompt

Prompt
Act as a pragmatic AI engineer who prefers the simplest thing that works. Help me decide how to automate this task: Triage inbound customer refund emails for our online kitchenware store: read the email, look up the order in Shopify, decide if it qualifies under our 30-day refund policy, and draft a reply in Zendesk. (example value for {describe the task and its inputs})

If I have not said, ask me first in one message: how often it runs and at what volume, what happens when it gets something wrong, which tools or systems it must touch, and what a run may cost.

Score the task 1-5 on: how much the steps vary from case to case, how much it needs tools or live data, how much error it can tolerate, and how sensitive it is to cost and latency. Explain each score in one line.

Then recommend one of: a plain script or rules, a fixed LLM workflow (set steps with model calls), a single agent with tools, or multi-agent.

Rules: start from the simplest level and justify every step up; choose multi-agent only if one agent clearly cannot hold the context or the work truly runs in parallel; say where a human should approve.

Output format: a score table, the recommendation with reasons, the cheapest version to build first (steps and tools), what evidence would justify moving up a level, and 3 checks with pass thresholds that prove the choice works. List your assumptions and the risks of getting this wrong.

Dashed parts are example values, so the prompt runs as copied. Change them in Fill in the blanks.

Example output

At a glanceRefund email triage: automation scoresScores 1–5 lead to a fixed LLM workflow recommendation
  • Step variability: 2
  • Tool / live data need: 3
  • Error tolerance: 2
  • Cost / latency sensitivity: 3

Example for: ~150 runs/day on weekdays, a wrong refund costs money and customer trust, touches Zendesk and Shopify, budget under $0.05 per run.

Task: Triage inbound customer refund emails for an online kitchenware store: read the email, look up the order in Shopify, decide whether it qualifies under the 30-day refund policy, and draft a reply in Zendesk.

Score Table
DimensionScore (1–5)Why
Step variability2Nearly every case follows the same path: find order → check date and condition → draft reply.
Tool / live data need3Needs one Shopify order lookup and one Zendesk write. Both are fixed calls.
Error tolerance2A wrong approval loses about $40–$120, and a wrong denial angers a customer.
Cost / latency sensitivity3At 150/day the budget is tight, but replies within an hour are fine.
Recommendation: Fixed LLM workflow
  • Not a plain script. Customers write freely ("the lid cracked", "wrong colour", no order number). Rules would break when extracting the reason and the order reference.
  • Not an agent. The tool sequence never changes, so letting a model choose its next step adds cost and failure modes for no benefit.
  • Not multi-agent. The context is small and nothing runs in parallel.
  • Human approval: An agent reviews every draft before it is sent. Any refund over $100, or any case marked "unclear," goes to a supervisor.
Cheapest First Build
  1. A Zendesk webhook fires on new tickets tagged refund.
  2. LLM call 1 (small model): extract order number, email, reason category and damage claimed into JSON.
  3. Script: run the Shopify lookup, then apply policy rules in code (within 30 days, item not in the final-sale list).
  4. LLM call 2: draft a reply from a template using the rule outcome.
  5. Post the draft as an internal note with a confidence flag.

Estimated cost is about $0.01 per run. Total is roughly 2–3 days of work.

Evidence to Move Up a Level

Consider a single agent only if more than 20% of tickets need extra lookups the workflow can't handle. Examples include split shipments, exchanges or carrier tracking checks.

Checks
  1. Extraction accuracy: at least 95% correct order matches on 200 labelled past tickets.
  2. Policy decision agreement: at least 98% agreement with human decisions over the first two weeks.
  3. Agent edit rate: fewer than 25% of drafts need substantive edits by week 3.
Assumptions
  • Policy rules are written down and stable.
  • Shopify API access is available.
  • Agents will review drafts rather than rubber-stamp them.
Risks
  • Over-trusting drafts leads to wrong refunds.
  • Customers without order numbers cause failed lookups, so add an email-based fallback search.
  • Escalating to an agent too early multiplies cost by about 5× with little accuracy gain.
Written by AI from this prompt. Your results will vary.

Paste into Claude, ChatGPT or Gemini before you start building. Describe a real task with examples of inputs; vague tasks get vague scores.

Keep exploring

More in Agentic AI