When Not to Use AI Agents: 5 Checks Before You Ship
Have you built an agent at work and then decided it should never ship? Many builders have. The demo works. Then someone asks what happens when it is wrong, what it can reach, and how you would know it helped. If nobody can answer, killing it is the right call.
Do not use an AI agent when the job is fixed rules, a single step, or a choice that cannot be undone without a person. Use one when the input is messy, the steps change, and a wrong answer is cheap to catch.
This post does not tell you agents are bad. It gives you five checks, and a free item on our shelf that runs each one. We did not run them. Kill rates: not measured.
Quick answer
| Sign it should not ship | What to check | Item to start with |
|---|---|---|
| Nobody wrote down how it fails | Real risks vs loud worries | Pre-Mortem |
| "It looked fine in the demo" | A rerunnable eval suite | DeepEval |
| It can reach data or tools it should not | Threats per component | Security Threat Model |
| No owner, no rollback | Go / No-Go per function | Launch Readiness |
| No one can say it beat the old way | A test with guardrails | A/B Test Planner |
How we picked
The top Google results for this search give good frameworks, but none hands you something to open today. We read the full file of each item below and kept ones that turn one check into a document you can show your team.
1. Write the failure story first
Most agents die quietly because no one says the risk out loud.
Pre-Mortem asks the model to imagine the launch in 14 days, then imagine it failed, and work backward. It sorts every risk into Tigers (real problems), Paper Tigers (overblown worries) and Elephants (things the team is not discussing). Tigers then get a tag: launch-blocking, fast-follow, or track. Every launch-blocking Tiger needs a mitigation, an owner and a date. If you cannot fill those in, you have your answer.
2. Prove it works on more than the demo
A demo is one happy path. Users bring the other hundred.
DeepEval adds an eval loop to an AI app. It builds a small pytest suite you can rerun without an agent, makes a dataset with deepeval generate (the file aims for about 30 to 50 goldens) and runs it with deepeval test run. It tells the model not to make up goldens by hand. You need Python 3.9+, pip install deepeval, and model credentials for the metrics, so expect API cost on your side. If scores stay low after a few rounds, the agent is not ready.
3. Map what it can touch
An agent with wide access is a security review waiting to fail.
Security Threat Model writes a STRIDE threat model: an asset list, trust boundaries, threats per component, a likelihood times impact score, and mitigations with owners. Every accepted risk needs a named owner and a review date. Note: the file marks itself deprecated in favor of a newer threat-model skill, but the method still holds. If the agent needs admin access to be useful, that is a strong sign to stop.
4. Ask for an explicit Go or No-Go
"Almost ready" is how bad launches happen.
Launch Readiness checks each function, from product and engineering to support and data, and returns Go, Conditional Go or No-Go with reasons. Its rules are strict. No function is "Ready" without evidence. Every blocker gets a single named owner, not "the team". The rollback plan only counts if it ran in staging. A No-Go here is the paper trail for a good decision.
5. Show it beats the old way
If you cannot prove a gain, the safe choice is to keep the old process.
A/B Test Planner writes the test before launch: a hypothesis with a direction, one primary metric, up to two guardrail metrics, sample size, and a rollback trigger. Its guide is blunt: an inconclusive result means do not ship and do not call it a win. If your traffic is too low to reach a result, it points you to interviews or a holdout instead.
When an agent is the wrong tool
- The steps fit on a whiteboard and the inputs never change. Use a script or plain automation.
- It is one call: classify, summarize, extract. Use a single prompt.
- A wrong action cannot be undone: refunds, mass email, deleting data. Let the agent draft, and let a person press the button.
- You have not mapped the process yet. Map it first.
Shelved an agent? Keep the eval set and threat model. Often a smaller tool fits: a prompt, a skill, or a draft-only flow with human approval.
Browse more on the Claude skills shelf.
FAQ
When should you not use AI agents? When the task is fixed rules, a single step, or a choice that cannot be undone without review.
When to use AI agents and when not to? Use them for messy, multi-step work where errors are cheap to catch. Skip them where rules or one prompt do the job.
Is it a failure to shelve an agent? No. A written No-Go with reasons saves the team from a worse launch.
How do I know an agent is ready to ship? It passes an eval suite beyond the demo, has an owner and a rollback plan, and beats the old way in a planned test.