An AI Agent Needs a Job Card, Not a Wish

“Help me run the business” sounds like a job description until you imagine saying it to a new employee on Monday morning and walking away.

Help with what? Starting when? Using which information? What can they change? When should they stop and call you?

An AI agent needs those answers too. Before you connect one to an inbox, calendar, customer list, or other business tool, give it one assignment small enough to explain on a job card. This guide gives you a six-part version you can write in plain language.

First, know what makes it an agent

An ordinary AI chat answers a question and waits for the next one. An AI agent can follow instructions, use approved tools, and carry a task through several steps toward a result. Depending on the setup, a person may start it, a schedule may start it, or an event such as a new form submission may start it.

OpenAI's current agent-building materials describe agents in terms of instructions, guardrails, and tools. Its workspace-agent guidance uses a similar three-part picture: a trigger, a process, and the systems the agent may use.

That distinction matters because access changes the stakes. Asking a chat to suggest a reply is different from giving a system permission to read an inbox, update a record, or send the reply.

The best first question is not, “How smart is this agent?” It is, “What exact job am I willing to let it practice?”

Pick work that has a shape

A useful first job comes back often and has an answer you can inspect. It should have a recognizable beginning, a short process, and a clear finish.

Good practice jobs might include preparing a morning list from an approved calendar, checking new request forms for missing fields, or turning a weekly set of authorized records into an internal summary.

Open-ended strategy is harder to judge. “Find ways to grow my company” can wander in a hundred directions. “Every Friday, summarize new quote requests and flag the ones missing a phone number or service address” has a shape. You can tell whether it ran, what it read, and whether the output is right.

Do not start with the job that can spend money, promise a price, cancel an appointment, or speak publicly in your name. Start where a mistake is easy to see and easy to correct.

Write the six-part job card

Take a sheet of paper and give the assignment six headings. None of them requires technical language.

1. Trigger: what starts the work?

Write the moment the agent should begin. “When I ask” is a valid trigger. So is “weekdays at 4 p.m.” or “when a completed website form arrives.”

Avoid “whenever necessary.” A vague trigger makes it hard to know whether the agent missed a run or started when it should not have.

2. Inputs: what may it read?

Name the specific source. “The approved estimate-request form” is clearer than “our customer information.” “Tomorrow's jobs on the service calendar” is clearer than “the calendar.”

Also name what is out of bounds. A job may need a business inbox folder without needing an owner's personal mail. It may need customer names and requested services without needing payment details.

Access should match the assignment, not the agent's maximum capability.

3. Process: what steps should it follow?

Write the routine as a short checklist. Use actions you could watch another person perform.

For example:

• Open the approved list of new requests.
• Check whether each request has a name, contact method, service location, requested work, and preferred timing.
• Copy supplied details into the internal summary.
• Mark missing information as missing; do not guess.
• Draft one question for each missing item.

Clear steps reduce the space where a smooth-sounding guess can sneak in. They also give you something concrete to improve after a bad test.

4. Allowed output: what may it produce or change?

Be precise about the finish line. The agent may create an internal draft, add a label, prepare a checklist, or update one approved field. Those are different permissions.

“Draft a reply for owner review” does not mean “send the reply.” “Recommend a time” does not mean “book the appointment.” “Flag a possible duplicate” does not mean “delete a customer.”

OpenAI's current guardrail guidance separates automatic checks from human review. It recommends pausing for human approval before side effects such as edits, cancellations, commands, or other sensitive actions. You do not need to use that exact software to borrow the principle: put the pause before the part that is difficult to undo.

5. Stop rules: when should it hand the job back?

Write down the conditions that need a person. Examples include conflicting records, missing required details, a request outside the service area, a price question, an angry message, or any instruction that falls outside the written job.

A stop is not a failure. It is the agent doing the job correctly by recognizing the edge of its authority.

Use simple wording: “If the form and calendar disagree, do not choose one. Show both and ask the owner.” That is far safer than hoping the agent understands what “use your judgment” means in your business.

6. Receipt: how will you know what happened?

Ask for a small record of the run. It might show the start time, sources checked, items reviewed, missing information found, drafts created, actions taken, and anything handed back to a person.

A receipt is not a victory speech. It is evidence you can compare with the original records. If the agent says it reviewed eight requests, you should be able to locate those eight requests. If it changed a field, the record should name the field and the before-and-after value.

NIST's voluntary Generative AI Profile is designed to help organizations include trustworthiness in the design, use, and evaluation of AI systems. For a small business, a plain run receipt is one practical way to make evaluation possible instead of judging the system by how confident its summary sounds.

A hypothetical bakery job card

Imagine a small bakery that takes weekend cake requests through a form. This is a planning example, not a report of a real business result.

The owner does not begin by asking an agent to accept orders. The first job is quieter: prepare a Friday readiness list.

• Trigger: Friday at 3 p.m., or when the owner runs it manually.
• Inputs: The approved cake-request form responses for the coming weekend and the bakery's pickup calendar. No payment information.
• Process: Match each request to a calendar entry; copy the requested date, pickup time, size, flavor, customer contact method, and exact notes; mark anything missing or inconsistent.
• Allowed output: Create an internal checklist and draft clarification questions. Do not send messages, confirm availability, quote a price, or change the calendar.
• Stop rules: Hand back duplicate names, conflicting pickup times, requests not found on the calendar, unclear notes, and any detail not present in the source.
• Receipt: List the records checked, missing fields, conflicts, and drafts prepared.

That assignment will not run the bakery. Good. It gives the owner a result that can be checked against two known sources. If the agent misses a conflict, the owner can improve the instructions without first apologizing to a customer.

Test in three stages

The first run should use invented or scrubbed practice records. Include an easy request, one with a missing phone number, one duplicate, and one conflict between the form and calendar. See whether the stop rules actually stop the work.

Next, let the agent shadow the real process with the smallest practical access. It prepares the list while the owner still does the normal check. Compare the two. Record what it missed, what it added, and which instruction was unclear.

Only then consider a limited real action. Move one step at a time: observe, draft, then act. Keep human approval in front of money, customer promises, outbound messages, account changes, and anything difficult to reverse.

Do not measure the pilot by whether the output sounds professional. Measure whether it used the correct records, preserved unknowns, followed the stop rules, and saved enough work to justify the checking time. A fast draft that needs a complete rewrite is not a successful shortcut.

The job card is also your troubleshooting guide

When a run goes wrong, return to the six headings. Did the wrong event trigger it? Was the source too broad? Was a step missing? Did “draft” accidentally become “send”? Was the stop rule vague? Did the receipt omit the evidence you needed?

That is much easier to fix than “the AI was weird.”

An agent will not make an undefined process clear for you. It may expose where the process was unclear all along. That can still be useful. The job card turns the confusion into a sentence you can rewrite.

Try this today: write one card without connecting anything

Set a timer for 15 minutes. Choose one recurring task that produces an internal result.

Write six lines: trigger, inputs, process, allowed output, stop rules, and receipt. Put a box around every action that changes a record, contacts a customer, spends money, or makes a promise. Write “owner approval required” beside each box.

Then test the instructions by reading them as a literal new helper. Could they tell when to start? Could they find the right source? Would they know when the job is complete? Would they know when to stop?

Do not connect a tool today. If the job cannot fit on the card, narrow it until it can.

The goal is not an agent that can do everything. It is one small assignment you can explain, inspect, and trust a little more after every careful run.

Rich Oneday

Rich writes about small business, creativity, and everyday technology. Clear explanations, useful examples, and a next step you can take today.