Planning
How to describe a process so it can be automated
A practical method for writing down a process, from trigger to final check, so that someone can automate it without guessing.
November 11, 2026
6 min read
The hardest part of automating a process is rarely the automation itself. It is getting a clear description of what the process is. People who do a task every day carry much of it in their heads, and when they explain it, they skip the parts that feel obvious. Those skipped parts are often exactly where an automation breaks.
A good description is specific and written down. Whoever builds the automation, whether that is a colleague, a freelancer or an agency like ours, can then work from the page instead of from guesses. Writing it also tends to expose problems in the process before anyone builds anything.
Why the description comes before the tool
A process description does not depend on any tool. The same written process can be built in n8n, Make or Zapier, or handed to an AI agent with access to your inbox and spreadsheets. If you choose the tool first, you start bending the process to fit it.
Writing it down also forces decisions that are easy to postpone in conversation. "We usually reply within a day" has to become a rule: reply by when, using which template, and who signs off. If nobody can agree on the rule, that is worth knowing early.
The seven parts of a process description
Write each part in plain language, ideally on one or two pages in total.
Trigger
Write down what starts the process and where it happens, and be precise about the event. "When we get an order" is too vague. "When a new email from the web shop arrives in the shared orders inbox, with a subject starting 'New order'" is something a builder can work with.
Some processes start on a schedule (every Monday at 8am) or by hand (someone marks a row as ready in a sheet). If there are several triggers, list each one.
Inputs
List the information the process needs and where each piece comes from: client name, email, order number, amount, attachments. Note the format each one arrives in (a form field, free text in an email, a PDF) and whether it is always present. A field that is "usually there" needs a rule for when it is missing.
Attach two or three real examples with personal details removed. Real examples show the mess that a tidy description smooths over.
Steps
Write the steps in order as short, numbered actions, each starting with a verb: read, check, copy, create, send. Keep one action per step. If a step says "process the request", break it down until each line describes something you could watch a person do.
Decisions
Mark every point where the path splits. For each one, write the question, the possible answers and what happens after each answer. "If the amount is over 1,000, send it to the manager for approval. Otherwise, approve it automatically." Words like "urgent", "large" or "important" need a definition, even a rough one.
Note which decisions call for judgment. Those are the places where an AI model can help by classifying or summarizing, or where a person should stay involved.
Exceptions
List what goes wrong and what should happen when it does: missing information, duplicates, messages in another language, attachments that cannot be opened, the client who replies to an old thread about something new. You will not catch everything. Write down the cases that have actually happened, then add a default rule that anything else goes to a person.
Outputs
Describe what exists when the process is finished: a row in a spreadsheet, a contact in the CRM, an email sent, a file saved in a particular folder under a particular name. Be precise enough that you could check each output: which fields, what format and where it lives.
Who checks
Name the person or role who reviews the work, what they look at and how often. In the first weeks, that may be every output. Later, it may be a daily look at a "needs review" list. Also name who gets told when something breaks.
A worked example
Take a hypothetical property management company with eight staff that handles tenant maintenance requests. Its description might look like this.
Trigger: A tenant submits the maintenance form on the website, or sends an email to the shared maintenance inbox.
Inputs: Tenant name, property address, unit number, a description of the problem and optional photos. Form submissions always include the address. Emails often do not, so the sender has to be matched by email address against the tenant list in Google Sheets.
Steps:
- Read the request and identify the tenant and property.
- Classify the issue as plumbing, electrical, heating, appliance or other.
- Rate the urgency using the decision rules below.
- Create a row in the maintenance tracker with all fields and a link to the original message.
- Send the tenant a confirmation with a reference number.
- Notify the property manager responsible for that building.
Decisions: A request is an emergency if it mentions an active water leak, no heating during the cold months, a smell of gas or sparking electrics. Everything else is standard. Emergencies also go by text message to the on-call phone. For any mention of gas, the confirmation tells the tenant to call the gas emergency line first.
Exceptions: If the tenant cannot be matched, the request gets a "needs review" label and no confirmation is sent. If the same tenant reports the same issue twice within 48 hours, the second message is added to the existing row instead of creating a new one. Requests written in another language are translated for the tracker, and the confirmation goes out in the tenant's language.
Outputs: One tracker row per request, one confirmation email to the tenant, one notification to the manager, and a text message to the on-call phone for emergencies.
Who checks: The office coordinator reviews every new row each morning for the first month, checking the category and urgency. After that, a weekly spot check of ten rows, plus everything labeled "needs review".
The AI model's job here is narrow: read free text, pull out the fields, classify the issue and flag emergencies. Everything around it is ordinary rules.
The more of a process you can write as plain rules, the smaller and more dependable the AI part becomes.
Gaps we often see in first drafts
When you reread your description, look for these:
- Undefined words. "Urgent", "VIP client" and "large order" each need a definition the builder can apply.
- Knowledge in one person's head. "I know that supplier always sends invoices late" is a rule nobody wrote down.
- No rule for silence. If the client never replies or the manager never approves, something should still happen. Add a time limit and a next step.
- Several versions of the process. If two people do the task differently, settle on one before anything is built.
- No volume figure. Note roughly how many times the process runs each week. It affects both the running cost and the right tool.
Where to start
Pick one process that repeats at least weekly and write it out using the seven parts: trigger, inputs, steps, decisions, exceptions, outputs and who checks. Attach a few real examples with personal details removed. Then hand it to a colleague who has never done that task and ask them to walk through one example using only the page. If they can, an automation can probably be built from it too. Every question they ask points to something missing. Fix those gaps and you have a description someone can build from.
If you have several processes written up and want help deciding which to automate first and in what order, that is the job of our AI Automation Roadmap.


