Before You Hand a Process to an AI Agent, Write Down These Seven Things

Every business has a job that "everybody just knows how to do." Confirming tomorrow's appointments. Chasing a customer for a missing form. Sorting the morning's emails into the right piles. It works because the person doing it carries the rules in their head.

That is exactly why it breaks when you hand it to an AI agent. An agent can't read your office manager's mind. It does what it was told, and when it wasn't told, it guesses. The fix is not a smarter agent. The fix is a better write-up.

This guide walks you through the seven things to write down before any routine process goes to an agent, including the part most owners skip: exactly when it must stop and ask a person. At the end there is a one-page template you can copy.

Why Your Best Person's Know-How Is the Real Input

When people talk about AI at work, they usually talk about the tool. A study by economists Erik Brynjolfsson, Danielle Li and Lindsey Raymond, published by the National Bureau of Economic Research, points somewhere else. They looked at an AI assistant that suggested replies, in real time, to human customer support workers at a software company. The assistant had been trained on a large set of past customer conversations, labeled by outcome and by whether the worker was a top performer. In the working-paper version, staff with the assistant resolved 14% more issues per hour on average, including a 34% improvement for novice and low-skilled workers, with minimal effect on the most experienced and skilled staff. Later versions of the paper report different figures.

The authors offer what they call suggestive evidence that the tool "disseminates the best practices of more able workers and helps newer workers move down the experience curve." In plain words: the best people's know-how was the valuable part, and the tool may have helped pass it on to newer staff.

Here's my read, not theirs. That system learned from a big archive of labeled conversations. A small business doesn't have one. Your agent works from what it's given: the systems it can reach and the instructions you write, and the instructions are where your best person's habits live. The authors themselves note that this kind of know-how is hard to pass on through training or manuals, so don't copy the manual. Sit with your best person and write down what they actually do on real, recent cases, which is exactly what Step 3 walks through.

There is a second reason. The U.S. government's standards agency, the National Institute of Standards and Technology (NIST), publishes a voluntary framework for managing AI risk. One of the suggested actions in its companion playbook is that organizations "define and differentiate the various human roles and responsibilities when using, interacting with, or monitoring AI systems." That sounds formal, but it boils down to two questions every small business can answer on paper: what does the agent do, and which person is responsible for what it can't?

Steps 1 and 2: Write the Trigger, Then the Inputs and Where They Live

Step 1. The trigger: what starts the job?

Every process begins with something happening. Write down exactly what that is, in one sentence. Good triggers are specific:

• "A new booking appears in the calendar for the next business day."
• "An email arrives in the service inbox with an attachment."
• "A job is marked complete in our field-service software."

A bad trigger is "when things come in" or "every so often." If you can't name the trigger, you have found your first gap. Talk to the person who does the job and ask, "What makes you start this?"

Step 2. The inputs: what information is needed, and where is it?

List every piece of information the job uses and the place it lives. For an appointment confirmation, that might be the customer's name, phone number, service type and time (from the booking system), plus your cancellation policy (from a shared document).

Be honest about the messy ones. If the "real" price list is a sticky note by the register, write that down. An agent can only use sources it can reach, so this list also tells whoever sets it up which systems it needs to connect to, and which sticky notes need to become a document first.

Step 3: Write the Steps the Way You Actually Do Them Today

Now sit next to the person who does the job and write down what they really do, in order. Not what the manual says. What they do.

The best way is to walk through a handful of recent real cases together. For each one, ask: what did you look at first? What did you do next? Did you check anything before you sent it? You will almost always discover small steps nobody thought to mention, like "I check whether they're a repeat customer so I use their first name," or "I never text before 8 a.m."

Write each step as a short action with a verb at the front:

1. Look up tomorrow's bookings.
2. For each booking, check whether a confirmation was already sent.
3. Send the confirmation message using the approved wording.
4. Mark the booking as "confirmation sent."

This is the heart of the tribal-knowledge problem. The steps people skip when describing a job are the ones they do on autopilot, and those are often the ones that make customers happy.

Step 4: Turn "Use Your Judgment" Into Decision Rules

This is where most write-ups quietly fail. Somewhere in the process there is a moment where the employee decides something. In a write-up, that often appears as "use your judgment" or "handle it appropriately." An agent can't do anything useful with that sentence.

Instead, find each decision and turn it into an "if this, then that" rule. Ask the employee, "When you decide, what are you looking at?" Their answer is the rule.

You will not catch every decision on the first pass. That's fine. Anything you can't turn into a clear rule becomes a stop-and-ask rule in the next step.

Step 5: Write the Stop-and-Ask Rules (the Part Everyone Skips)

A process description that only covers the normal path is half a description. The other half is: when does the agent stop and hand this to a person?

Write down three things.

1. The situations that always go to a person. Good candidates are anything involving money leaving the business (refunds, credits, discounts), complaints, safety or health concerns, legal threats, anything the customer says is an emergency, and anything that doesn't match one of your written rules.
2. Who it goes to, by name or role, with a backup. "Someone in the office" is not an answer. "The office manager, or the owner if she is out" is.
3. What the handoff includes. The person picking it up should not have to start from scratch. Ask for a short summary: who the customer is, what they asked, what the agent already did, and why it stopped.

A simple test: read your write-up and ask, "What would a brand-new employee do if a customer asked for something not on this page?" If the answer is "guess," you need another stop-and-ask rule. The best default is simple: when in doubt, stop and ask. An agent that hands over too often is a tuning problem. An agent that guesses on a refund is a customer problem.

For a worked example of drawing this line in one industry, see how a club separates a routine tee-time request from a member-account dispute.

Steps 6 and 7: Define "Done," and Name the Person Who Checks It

Step 6. What does "done" look like?

Write down how anyone could tell the job finished correctly. It should be something you can see: a status changed in the system, a message sent with the approved wording, a file saved in a named folder. "Customer is happy" is a goal, not a finish line. "Confirmation sent and booking marked as confirmed" is a finish line.

Step 7. Who checks the work, and how often?

Every process needs one named owner, usually the person who used to do it by hand. They know what good looks like. Write down:

• Who reviews the agent's work.
• How often at first (for example, every day for the first couple of weeks), and how often once things are steady.
• What they look at: a sample of finished jobs plus every job the agent handed back to a person.
• Who updates the write-up when the process changes.

That last line matters more than it looks. Processes change. Prices change, staff change, policies change. If the write-up is not updated, the agent keeps following last season's rules.

Four Ways a Handoff Goes Wrong

If you only remember one section, make it this one.

• Undocumented tribal knowledge. The write-up describes the job from memory or from an old manual, and the small, important habits never make it onto the page. Fix: build the steps from recent real cases, with the person who does the job.
• "Use your judgment" instructions. The write-up hands the agent a decision without the rule behind it. Fix: turn it into an "if this, then that," or make it a stop-and-ask.
• No exception path. The write-up covers the normal case only, so the first unusual request becomes a guess. Fix: write the stop-and-ask list, the named person and the handoff summary.
• Nobody owns the check. The agent goes live and everyone assumes someone else is watching. Fix: one name on the page, with a review schedule.

A related question is how much an agent should be allowed to do on its own once it is running. We cover that separately in why a drafting agent shouldn't hold the publish key.

A One-Page Template You Can Copy

Copy this into a document, fill in one per process, and keep it somewhere your whole team can find it. It is written for people first. If it makes sense to a new hire, it will make sense to an agent.

1. Process name: (for example, "Confirm next-day appointments")
2. Owner: (name and role, plus a backup)
3. Trigger: (the one event that starts it)
4. Inputs and where they live: (each piece of information, and the system or document it comes from)
5. Steps as done today: (numbered, each starting with a verb)
6. Decision rules: (each written as "if this, then that")
7. Always stop and ask a person when: (the list, plus who to hand it to)
8. Handoff summary must include: (customer, request, what was done, why it stopped)
9. Done looks like: (something you can see in a system)
10. Review: (who checks, how often, what they look at)
11. Last updated: (date and by whom)

Once it's filled in, try a quick "new hire test": hand it to someone who has never done the job and ask them to talk through three recent real cases using only the page. Every time they have to ask a question, you have found a line to add.

Frequently Asked Questions

Which process should I document first?

Pick one that happens often, follows roughly the same path each time, and is annoying rather than risky when it goes slightly wrong. Confirming appointments, sorting incoming paperwork, or chasing a missing document are good first candidates. Refunds, disputes and anything involving money or safety are poor first candidates.

How long should the write-up be?

As long as it takes to answer the seven sections, and no longer. For most routine processes that fits on one or two pages. If it runs much longer, you may be describing two processes at once, and it is worth splitting them.

Can I just give the AI agent our existing employee manual?

It is a useful starting point, but manuals usually describe how things should work, not the exceptions people actually handle. Walk through recent real cases with the person who does the job and add what the manual leaves out, especially the stop-and-ask rules.

Who should own the write-up once the agent is running?

One named person, usually whoever did the job by hand. When the process changes, they update the document first and the agent's instructions second, so the two never drift apart.

The Bottom Line

Handing a process to an AI agent is a lot like training a new hire, minus the lunch break. The agent will be as good as the instructions it gets, and the instructions that matter most are the ones people forget to say out loud: the small habits, the real decision rules, and the moment to stop and call a person.

Write those seven things down first. Whether you build the agent yourself or work with someone like us, that page is the most valuable thing you bring to the table. I spend my tokens on writing, not running a front desk, so take this as a checklist rather than a war story, but that page is exactly what an agent would work from.

If you've filled in the template and want to see what an agent would do with it, start a conversation with us and bring your page along.