The Agent Job Card: an AI agent workflow template with three filled-in examples

A custom AI system sounds big until you see one agent's job written down. Then it looks like a work order: what starts the job, what you need, what gets done, who signs off, and where it's recorded. The Agent Job Card is that work order for an AI agent. Every agent in a system gets one, and you start with one or two cards, not twenty.

The five lines, in order

1. Trigger. The event that starts the job, like an estimate with no reply for three days.
2. Access. Only what that job needs to read, and nothing else.
3. Result. What the agent produces, like a drafted follow-up that mentions the job the customer asked about.
4. Approval. A person in your office approves, edits or rejects the draft before anything goes out.
5. Record. A dashboard or log that shows what ran, what got approved and what's still waiting.

Why one card per agent

It's tempting to picture one big robot running your company. A real system is a few small agents, each doing one job inside the software you already run. That's easier to trust, easier to fix and easier to switch off.

It also matches where small businesses are. In the Federal Reserve Banks' 2026 survey of small employer firms, 46% were using AI but only 7% had it fully integrated into their operations, and 46% of current users named accuracy as a challenge. One card, one job, with a person on the approval line, is how you get from trying AI to relying on it.

Who this is for and what you need first

This is for an owner or ops manager scoping a custom AI system, with a developer or on their own. You need:

• The job, in one sentence.
• The systems it touches (field software, books, phones, CRM).
• The person who does the job today. They usually know the trigger and the result better than anyone.

How to fill in each line

Trigger. Write it as an event your software can see, with a number. "No reply to an estimate after 3 days" works. "When a customer seems interested" doesn't. If you can't say what the software would check, the job isn't ready.

Access. List the records and the permission for each: read or write. Most agents only need to read a few records and write a note or a draft. Connect through each system's own app access with chosen scopes; ServiceTitan, for example, has admins create an API application with the scopes it needs, using machine-to-machine sign-in, and HubSpot gives each private app its own token and scopes. You'd think more access makes a smarter agent. Every extra permission is one more thing that can go wrong.

Result. Describe the output exactly: a text, an email, an invoice draft, a note on a record, a line in a report. Say what it must mention and what it must never say.

Approval. Name the person and their backup. Say what they check against (the estimate, the invoice, the schedule). Anything about money, prices, refunds or promises always needs a yes.

Record. Say where you'll see each run: a dashboard, a log, a note in the CRM. If you can't look back at yesterday, you can't trust tomorrow.

The blank card

Copy this for each job.

Line	What to write
Job name
1 Trigger	The event and its number
2 Access	Each record, read or write
3 Result	The exact output; must mention; must never say
4 Approval	Who approves, their backup, what they check against
5 Record	Where each run shows up
Owner	Who in the office owns this card
Review date	When you'll read the record and decide what changes

Example card 1: estimate follow-up

Line	Filled in
Job name	Estimate follow-up
1 Trigger	Estimate sent, no reply after 3 days, no follow-up logged
2 Access	Read the estimate and the customer's name and phone. Write a draft text.
3 Result	One short text from the salesperson that mentions the job the customer asked about. Never mentions a discount or a new price.
4 Approval	The salesperson who wrote the estimate approves or edits. Backup: office manager.
5 Record	Dashboard shows each draft, who approved it, and whether the customer replied.

Example card 2: missed-call text-back

Line	Filled in
Job name	Missed-call text-back
1 Trigger	Missed inbound call with no callback within 10 minutes during business hours
2 Access	Read the call log and match the number to a customer. Write a draft text.
3 Result	A text in the company's name asking what they need and the address. Never gives a price or an arrival time. A gas smell or no-heat emergency goes to the on-call tech instead.
4 Approval	Dispatcher approves until the wording is locked; after that, the plain version can go on its own. Emergencies always go to a person.
5 Record	Each missed call, the text, and whether it was booked.

Example card 3: late-invoice nudge

Line	Filled in
Job name	Late-invoice nudge
1 Trigger	Invoice 7 days past due, no payment, no open dispute note
2 Access	Read invoices and payments in the books and the job's status. Write a draft email.
3 Result	A polite reminder with the invoice number, amount from the books and the payment link. Never mentions late fees, liens or collections.
4 Approval	Whoever does billing checks the amount against the books and approves.
5 Record	Each reminder, its approval and the payment date when it arrives.

These are examples, not results. Your triggers and wording will differ.

How cards become a system

Start with one or two cards, picked because the job crosses systems or eats the most office time. Run them for a month, read the record every week and change the cards when something's off. Then add the next one.

That's how we build at Cortex: an audit finds the first one or two workflows, each agent gets only the access it needs, nothing gets replaced, and every build includes a private dashboard where the record lives. How we work lays out the five steps, and human-in-the-loop shows the approval queue. Very few small firms run anything like this yet; the Census Bureau found 19.8% of US businesses were using AI at all as of May 2026.

What to watch out for

• Vague triggers. "When it makes sense" is not a trigger. Give it an event and a number.
• Access creep. A card's access list should only shrink or stay the same unless you change the job.
• No owner. A card nobody owns stops getting reviewed, and an unreviewed agent drifts.

How to tell it worked

After a month, read each card's record. You should see how many times it triggered, how many drafts were approved as written, how many were edited and why, and what happened afterward. If the approver edits most drafts, fix the Result line before you add another card.

For the bigger picture, read our guide to a custom AI workflow system. When you want the first two cards filled in for your company, your free AI Operations Audit does that.