The process everyone knows and nobody can describe
Someone asks how an order gets approved. The answer comes quickly: it arrives, someone checks it, a manager signs off, it goes out. Four steps. Easy to automate.
Then you sit next to the person who actually does it. The order arrives by email, except the large ones, which come by phone. The check is done against a spreadsheet, except for one customer who has a special arrangement. The manager signs off, unless she is away, in which case a colleague approves it and tells her later.
The four steps were true. They were just not the whole process.
Why automation fails on unwritten processes
Software does exactly what it is told. People do what the situation needs. When a process has never been written down, the difference between the two is held in a handful of habits, exceptions and quiet fixes that nobody thinks to mention, because to them it is simply the job.
Automate the four-step version and the first unusual case stops the line. Add AI on top, and it learns from the same incomplete picture. The tool gets blamed, but the real gap was the map.
Map the work as it happens, not as it is described
The most useful step comes before any code: watch the work and write it down.
Follow real cases from start to finish
Pick a few recent items, an order, a request, an invoice, and trace each one through every hand it passed. Note where it waited, who touched it and what they looked at to decide.
Ask about the last time it went wrong
"What usually happens?" gets the tidy version. "Tell me about the last one that was unusual" gets the exceptions. Those stories are where most of the rules live.
Write down the workarounds
The extra spreadsheet, the message sent to a colleague, the note on someone's desk. Each workaround is a step the process really has. Some should be automated, some should be removed, but none should be ignored.
Decide what to automate and what to leave to people
Once the map is on paper, sort the steps. Repeated steps with clear inputs and a clear rule are good candidates for automation. Steps that need reading messy text or sorting requests into types can be a good fit for AI, as long as someone can check the result. Steps that depend on judgement, a relationship or an exception should stay with a person, with the software making that hand-off easy rather than hiding it.
This is also where you choose one problem to start with. Automating one well-mapped step and watching it work teaches more than automating the whole flow at once.
A process map checklist you can copy
1. Trigger: what starts the process, and every channel it can arrive by.
2. Steps: each action in order, with who does it.
3. Decisions: what is checked at each point, and against what source.
4. Exceptions: the cases that do not follow the usual path, and what happens to them.
5. Workarounds: spreadsheets, messages and notes that fill gaps today.
6. Hand-offs: where work waits for another person, and how they find out.
7. Owner: the person who can confirm the map is right and approve changes.
8. First candidate: one step to automate first, and how you will know it works.
What a written process gives you
Even before any automation, the map pays off. New team members learn the job faster. Disagreements about how things should work surface early, when they are cheap to settle. And when you do build, the people who know the work can review the plan and say, with confidence, "yes, that is how it really happens."
We build AI features and automation aimed at a measurable problem, and every build starts with discovery: goals, users and requirements before any code. If you have a process worth automating that nobody has written down yet, [tell us about it](/contact).
Part of the team building and running the products behind these posts at Hedaya Global Solutions.