How to write a process down so an AI can run it

An AI cannot run a process you keep in your head.

You already know the steps. You have done this task fifty times. The sequence lives in your fingers, not on a page. That is the problem. No tool, no model, no automation can begin until the process leaves your brain and lands somewhere stable.

Most process documentation fails because it aims at the wrong reader. It explains the big picture. It includes org charts. It uses verbs like "manage" and "coordinate." An AI does not need any of that. An AI needs a trigger, a list of inputs, a sequence of small decisions, and a clear definition of done.

This article gives you a single template. You can write it in twenty minutes. It works for the person who does the job now, because you are the only one who knows where the real friction lives.

The cost of the missing page

When a process is not written down, three things happen.

First, the work stays on one person's desk. That person becomes a bottleneck. Every request, every variant, every edge case flows through them. The company cannot scale because the company is a person.

Second, the work drifts. Two people handling the same task will produce different outputs. One uses the old template. One forgot a compliance step. The drift is invisible until a client complains or a reconciliation fails.

Third, the work cannot be improved. You cannot measure a process you cannot see. You cannot hand it to a new hire without sitting beside them for a week. And you cannot give it to an AI, because an AI has no intuition. It has no "you know what I mean." It has only the words you give it.

Sarah runs accounts payable at a 40-person logistics firm. She processes 120 invoices a month. Each one requires her to check the PO number, match the amount, flag discrepancies over 5%, route to the right department head for approval, and post to the ERP.

Sarah knows all of this. She has done it for three years. When she is on holiday, the invoices pile up. When the company hired a junior to help, training took six weeks. Not because the work is hard, but because the process was invisible.

Sarah is not the problem. The missing page is the problem.

Why most process docs fail an AI

The typical process document reads like a meeting agenda. It lists objectives. It describes roles. It says things like "ensure timely resolution" and "maintain stakeholder alignment."

An AI reads that and produces nothing useful. It cannot act on "ensure." It needs an instruction at the level of a click, a keystroke, a field name.

Here is what an AI needs instead:

A trigger. Something that says "start now." An email arriving in a specific inbox. A row appearing in a spreadsheet. A time on the clock.

Inputs. Named, located, accessible. "The PDF in the shared drive, folder /invoices/unprocessed." Not "the invoice."

Steps. One action per line. "Open the PDF. Read the total amount. Find the PO number on page one. Open the ERP. Search for the PO number."

Decisions. Every fork in the road spelled out. "If the amount matches exactly, go to step 8. If the difference is under 5%, flag for review. If over 5%, stop and email the department head."

Exceptions. The weird cases Sarah handles without thinking. The vendor who sends two invoices for one PO. The currency mismatch. The duplicate that looks like a new invoice but isn't.

A done state. Something concrete an AI can check. "Row added to the ERP with status 'posted' and a confirmation number." Not "invoice processed."

A checker. Who verifies the output, how often, and what they look for.

That is the whole template. Seven things. Most companies never write them down because the person doing the work does not need them. Until the work needs to leave that person.

The seven-part template

Here is the template in full. You can copy it into a blank document and fill it out for any repetitive task in your business.

1. Trigger

What exactly starts this process. Be specific about the channel and the format.

Bad: "When an invoice comes in." Good: "When an email with a PDF attachment lands in accounts@company.com and the subject line contains the word 'invoice'."

Bad: "Every Monday." Good: "Monday at 09:00 GMT, triggered by a scheduled check of the shared inbox."

If the trigger is a person saying "do the thing," the process is not ready for AI. Find the upstream signal that person is responding to.

2. Inputs

List every piece of data the process needs to start. For each one, name where it lives and how to access it.

Format: "[Name of input]: [location], [format], [access method]."

Example:

• Invoice PDF: /shared-drive/invoices/unprocessed, PDF, read access via service account
• PO database: ERP module Purchasing, table POs, read/write access via API key
• Approval matrix: /shared-drive/admin/approval-thresholds.xlsx, Excel, read access

If an input is "Sarah knows it," stop. That is a dependency. Write it down or accept that the process cannot be automated.

3. Steps

Write one action per line. Use verbs that a machine can execute: open, read, search, compare, copy, paste, send, update, log.

Number each step. Keep each step small. If a step contains the word "and," split it.

Example:

1. Open the invoice PDF.
2. Read the invoice number from the top right corner.
3. Read the total amount from the bottom right corner.
4. Read the PO number from the reference field.
5. Open the ERP.
6. Search the PO table for the PO number.
7. Read the expected amount from the PO record.

Notice how granular this is. "Process the invoice" became seven steps. That granularity is what lets an AI execute without guessing.

4. Decisions

At every point where the next action depends on a condition, write the condition and the two paths.

Format: "If [condition], then [action]. If not, then [alternative action]."

Example:

• If the invoice amount matches the PO amount exactly, go to step 12 (post).
• If the difference is less than 5% of the PO amount, go to step 10 (flag for review).
• If the difference is 5% or more, stop and send an email to the department head listed in the approval matrix.

Do not skip the "if not" branch. An AI with only one path will walk off a cliff when the condition fails.

5. Exceptions

List every weird case you have seen in the last six months. The vendor who sends a credit note as a negative invoice. The invoice that references two POs. The PDF that is scanned sideways.

For each exception, write what the AI should do. Often the answer is "stop and alert a human." That is fine. It is better to stop cleanly than to guess.

Example:

• Duplicate invoice number: check if already posted. If yes, stop and log "duplicate detected." If no, proceed.
• Missing PO number: stop and reply to the sender asking for the PO number. Log the exception.
• Scanned PDF with no readable text: stop and move to /invoices/manual-review. Alert the accounts team.

Exceptions are where most automations break. The happy path is easy. The edge cases are the work.

6. Done state

Define what "finished" looks like in a way that can be checked programmatically.

Bad: "Invoice is processed." Good: "A new row exists in the ERP table 'Posted Invoices' with the invoice number, amount, PO number, and a status of 'posted'. The confirmation number is logged."

The done state is your test. If you cannot write it in one sentence that references specific data in a specific system, the process is still fuzzy.

7. Who checks

Name the person who reviews the output. Say how often and what they look for.

Example: "Sarah reviews a daily summary email listing all invoices processed in the last 24 hours. She checks the count against expected volume and spot-checks two invoices per day for accuracy. If the count is off by more than 10%, she investigates."

This step matters because AI output is probabilistic. It needs a human in the loop, not to micromanage, but to catch drift early. The checker is not a failure of automation. It is the control that makes automation safe.

How to fill this out in twenty minutes

Sit down with the person who does the job now. Open a blank document. Talk through the last three times they did the task. Do not ask "what is the process." Ask "what did you do on Tuesday at 10am when the thing arrived."

Write down exactly what they say. Do not clean it up. Do not remove the bits that sound too obvious. The obvious bits are the ones an AI needs most.

Then go through the seven sections and fill each one. The trigger and the inputs come first. The steps will be the longest section. The decisions and exceptions will be the hardest, because the person doing the work has internalized them. Prompt them: "What made you stop and think? What made you call someone?"

When you have a draft, hand it to someone who has never done the task. Ask them to follow it literally. Where they hesitate, add detail. Where they take a wrong turn, add a decision point.

You now have a process an AI can run.

What happens after the page exists

Once the process is written to this template, three doors open.

You can hand it to a new hire and cut training time from weeks to days. The document is the trainer.

You can measure the process. How many invoices per hour, how many exceptions per week, how long the average cycle takes. You cannot improve what you cannot see.

And you can hand it to an AI. Not a generic chatbot. A system built to read this exact structure and execute it against your tools. The trigger fires. The inputs are fetched. The steps run. The decisions branch. The exceptions are caught. The done state is verified. The checker gets a summary.

The AI does not need to understand your business. It needs to understand the page.

Some companies try to skip the page. They hire an AI vendor and say "automate accounts payable." The vendor spends six weeks extracting the process from Sarah's head. The cost is higher, the result is fragile, and the knowledge still leaves when Sarah does.

Writing the page yourself costs twenty minutes. It gives you a reusable asset. It gives Sarah a weekend without a backlog. It gives the business a process that can scale beyond one person's memory.

Nexibeo does this work for companies every month. We take the page you write, or we help you write it, and we build the automation that runs it. One fixed fee, no platform for you to manage, no model to choose. You hand us the process and we hand back the done state. If that fits how you want to work, you can find us at nexibeo.com.

But the page comes first. The page is free. You can write it today. Pick one task, the one that haunts you when someone is on leave, and fill out the seven sections. You will see the process clearly for the first time. And you will have something real to build on.