Documenting processes so they can be automated with AI
By Keith Hamilton · · 4 min read
Documenting processes for automation comes before any technology, and it is where many automation projects stall. Someone asks how a task is done and gets three different answers, or a list of steps that leaves out the awkward cases that take most of the time. This article explains how to document a process so it can be automated, what to capture, and how to tell which parts suit AI and which need a person.
Why is undocumented work hard to automate?
Automation cannot fill gaps it does not know about. A process that lives in people's heads works because those people fill the gaps without thinking. They know which customer always needs a phone call, which supplier sends invoices in a strange format and which field in the system is never filled in properly.
If the rules are not written down, the automation follows the simple version, and the exceptions land back on your team, often at the worst moment.
Documenting a process pays off in its own right too. It makes training easier, reduces the risk when someone leaves and shows up steps that can simply be removed.
Which process should you document first?
Document one process that happens often, takes noticeable time and follows broadly the same pattern each time. Starting small is the core of choosing tasks to automate, and our guide on where to start with AI will help you pick.
Then name one owner: the person responsible for the process working well. They do not need to write everything themselves, but they decide what the correct version is when people disagree.
How do you document a business process?
Sit with the people who do the work, ideally while they do it, and have them talk you through each step. The manager's description of a process and the way it is actually done are not the same thing.
Ask questions such as:
- What starts this task?
- Where does the information come from?
- What do you check before moving on?
- What do you do when something is missing or wrong?
- Who else do you need, and how do you contact them?
- How do you know it is finished?
Record the walkthrough, with the person's agreement, and use AI to turn the transcript into a first draft of the steps. Check that draft carefully with the people who do the work, because an AI draft smooths over awkward detail unless someone checks it.
What should a process document capture?
A process document captures eight things:
- Trigger. What starts it, such as an email, a form or a date.
- Inputs. The information and documents needed.
- Steps. What happens, in order, in plain language.
- Decisions. Each point where someone chooses between options, and the rule they use.
- Systems. Every system the information goes into or comes out of.
- Outputs. What exists when the process is finished.
- Exceptions. What goes wrong, how often and what people do about it.
- Volume and time. How often it happens and how long it takes.
Write decisions as rules
Write each decision as a simple rule wherever you can: if this, then that. For example, if an order is over a certain value, a manager approves it.
If a rule needs judgement that is hard to put into words, say so. That step stays with a person.
Give exceptions their own section
Exceptions get their own section because they are where the time goes. The normal path is quick. List the common exceptions, how often they happen and how they are handled. This is the most valuable part of the document.
How do you decide what to automate?
Go through each step of the documented process and ask four questions. This is process mapping for AI in practice:
- Is it rules-based? Clear rules suit automation.
- Is the information digital? Paper and phone calls need an extra step to capture them.
- Does it need judgement? Reading and summarising suit AI, but decisions about money, complaints or people's circumstances stay with a person.
- What happens if it goes wrong? The higher the risk, the more a person checks.
The answer is usually a mix: automation handles the routine path and passes exceptions to a person with the information they need.
Why measure before you change anything?
The volume and time you captured are your baseline, and without one you cannot show what automation achieved. The arithmetic is simple: a task that happens 40 times a week and takes 15 minutes each time takes 600 minutes, or 10 hours a week.
Our guide to the real cost of manual work shows how to turn hours into pounds, so you can judge whether automation is worth it.
How do you keep the document current?
Give it a review date and update it whenever the process or a system changes. A process document that is out of date is worse than none, because people trust it.
Once you automate the process, the document also becomes the reference for how the automation should behave.
Where Forwardcycle fits
Our AI Opportunity Diagnostic includes a process inventory and scored opportunities, so you know which tasks to automate first. Our efficiency and capacity service then helps redesign the workflow so work reaches the right person first time. To get started, take the free AI Readiness Assessment or arrange a free call.