When you should not hire an AI agent
Wait when the work has no stable owner, no observable acceptance check, too little volume, or a decision that must remain with a person.
Do not hire an AI agent when nobody owns the result, the work changes shape every week, a fresh reviewer cannot check it, the consequence belongs to a licensed professional, or the volume is too small to justify a recurring role. Use a person, a simpler tool, or a short test instead.
Team size is not the gate. A two-person business with a stable weekly reporting burden may have a clear role. A larger company with vague ownership may not. The decision turns on work shape, evidence, authority, and frequency.
Five conditions mean wait
The no-hire test
A failed gate points to a safer route rather than a different sales pitch.
| Condition | Why an AI role does not fit yet | Use instead | Done when it is worth revisiting |
|---|---|---|---|
| No accountable owner | Exceptions and consequences have nowhere to go | Assign a person to own the result | One person owns inputs, approvals, exceptions, and corrections |
| No stable work product | The role changes faster than a repeatable specification can follow | Use a chatbot or human project owner | Three recent examples share the same inputs, output, and acceptance checks |
| No observable check | Fluent output can pass without being correct | Keep the work with an inspectable human process | A fresh reviewer can reproduce every material check |
| Licensed or binding judgment | Authority cannot be transferred by software access | Keep the decision with the qualified professional | The AI scope is limited to preparation and every consequential action requires approval |
| Too little recurring work | Maintaining role context costs more than the unfinished work | Use a person or a one-off tool | A dated queue shows repeated work and a response window that matters |
Passing every gate supports a test. It does not guarantee that a particular role will perform well.
Wait when nobody owns the result
An AI role needs an accountable owner even when the work is internal. That person controls source access, resolves exceptions, approves consequential actions, and decides whether a failed work product is corrected or abandoned.
Do not assign the role to “the team.” Name one person and one backup. The guide to hiring an AI agent treats the owner, work product, acceptance check, and approval boundary as parts of the purchase. The work-product promise does not create an internal owner for the buyer.
Done when: the role record names the owner, backup, approval boundary, exception route, and correction decision.
Wait when the work will not hold still
Early discovery, a company repositioning, a crisis, and an unfamiliar negotiation can change direction several times in one day. Those situations require someone to decide what the work is while doing it. A chatbot may help that person think; it should not be presented as the owner of a stable role.
Look at three recent examples. If the inputs, finished result, and acceptance checks differ materially every time, keep the work with a person. The agent versus chatbot guide explains the simpler route. Revisit when the repeated part can be stated without describing one exceptional project.
Wait when correctness cannot be inspected
“Looks right” is not an acceptance check. A financial figure should resolve to an authoritative record and date. A source claim should resolve to the supporting passage. A customer-support draft should resolve to policy, account state, and an approved remedy boundary.
The production-reliability guide shows why a good sample is insufficient when the worst failure matters. NIST's AI Risk Management Framework emphasizes measurement, monitoring, override, and incident response. It is voluntary guidance, not proof that a role is safe.
Done when: a fresh reviewer can run the check from the source record and reach the same pass, correction, or stop decision.
Wait when judgment must remain licensed or human
An AI system can prepare material for a lawyer, clinician, accountant, or employment decision maker. It does not inherit the person's license, duty, or authority. The EEOC's AI guidance makes clear that employment law still applies when automated systems are used. The AI constitution release test keeps preparation separate from professional release.
The same boundary applies outside licensed work. A customer promise, bank transfer, contract signature, personnel action, public statement, or deletion remains behind explicit owner approval.
Wait when the queue is too small
A recurring role earns its keep by preserving working context across related assignments or watched sources. When the gates pass, it replaces some human monitoring, preparation, record maintenance, and follow-through. The accountable owner retains judgment, exceptions, approval, and correction decisions. One short task every few months may be better handled by the existing owner or a Day Pass for a current role when the task fits that published scope. Check the current catalog and rate board; do not reshape a role around one stray task.
Measure the queue for four weeks. Record the work product, time requested, due state, correction burden, and consequence of delay. Done when: the dated record shows enough repeated in-scope work to justify the engagement and the owner can state the first useful output.
Run a bounded test after the gates pass
Choose an ordinary case, a known exception, and a request that must stop. For FARO, the AI SEO strategist, a bounded test could use an ordinary weekly Search Console brief, a page whose query does not match its reader task, and a request to publish or change the site without owner approval. Provide the exact sources and write the acceptance checks before work begins. Keep external and binding actions behind owner approval.
The test passes when all three cases leave an inspectable record, the known exception is surfaced, the disallowed action does not occur, and the owner can calculate the remaining review and correction work. Google's People + AI Guidebook provides additional human-centered evaluation patterns. If the test fails, keep the work record; do not convert the failure into a broader promise.
Use the AI agent catalog only for roles currently offered. If no current role matches, keep the operating specification and wait. A planned role is not available, and it has no public rate until it joins the catalog.
Follow the connected questions
Plan jobs and team change includes this decision and the questions that usually change it.
Which work should remain with a person?
Keep licensed sign-off, employment decisions, unfamiliar exceptions, relationships, physical presence, and accountability for consequential choices with a qualified person.
How narrow should an AI agent’s role be?
The role should be wide enough to own connected workflows and narrow enough that its sources, checks, limits, and approvals remain specific.
How do you hire an AI agent?
Start with one role, one first result, the systems it may read, the checks you will use, and the decisions that stay with a person.
Sources
- NIST, AI Risk Management Framework.
- U.S. Equal Employment Opportunity Commission, Artificial Intelligence and Algorithmic Fairness Initiative.
- Google, People + AI Guidebook.
- FidelicAI, how to hire an AI agent.