Every small business has a process that only one employee knows how to run. Payroll that only the bookkeeper can close. A quoting spreadsheet only the office manager understands. A client onboarding sequence that lives in one rep's head. It works until that person is on vacation, out sick, or gone. Automating it removes the risk, and the way you get there is worth doing even if you never automate a thing.
Why these processes resist automation
The person who runs it can't fully describe it. That's no criticism. They've internalized dozens of small decisions and exceptions, and when you ask "how do you do it?" you get the clean version. The real version has workarounds: the client who always needs a second invoice, the vendor whose file arrives in the wrong format, the step you skip on Fridays.
Automation built from the clean version breaks on the first exception. So the first job is capturing the real version.
Step 1: Record the real process
Ask the employee to record their screen while they run the process, three separate times, narrating as they go. Real work, with real data. The narration catches the decisions ("I check this because sometimes it's blank"). The recording catches the steps they didn't think to mention.
Three runs matter because the exceptions show up in different runs. One recording gives you the ideal. Three give you the process.
Step 2: Write the rulebook in one file
Turn the recordings into a document with three sections.
- Steps. What happens, in order, with the system and screen for each.
- Decisions. Every place the person chooses, written as "if this, then that." This is where the exceptions live, and the exceptions are the rulebook.
- Escalations. The cases where even the expert stops and asks someone. These stay human.
Keep one copy. When we've let two copies of a rulebook exist, they drifted apart within weeks and the drift caused real bugs. One file, and every tool and person points at it.
Step 3: Split lookups, decisions, and writes
Every step in the rulebook is one of three things.
- Lookups read information: pull the invoice, check the balance, find the customer record. Automate these first. They're safe, and they're most of the time spent.
- Decisions apply a rule to what was looked up. Automate the ones with clear rules. Leave the fuzzy ones to a person, with the lookup results already in front of them.
- Writes change something: send an email, move money, update a record a customer sees. Automate these last, and gate them. Draft instead of send. Propose instead of post. A person clicks approve.
This ordering is why a first version can go live in days. Most of the value is in lookups and clear decisions, and neither can hurt anyone.
Step 4: Run it in shadow mode
For the first stretch, the automation runs and produces its output, and the employee keeps doing the job as before. Each day, compare. Where they differ, one of three things is true: the rulebook is missing a rule, the automation misread the rulebook, or the expert made a judgment call the rulebook should capture.
Keep a miss log. Every difference gets a row with a category. A category that repeats means the rule needs to change, and a rule change fixes the whole class of mistake instead of the one instance.
Shadow mode is also your dry run. Nothing it produces reaches a customer or a bank until you've watched it agree with the expert for long enough to trust it.
Step 5: Move the review to someone else
The test that the key-person risk is gone is simple. A different person can approve the automation's output from its report, without the expert in the room. If the report says "three invoices ready, one exception flagged because the PO number is missing," and a second employee can handle that, the process no longer depends on one head.
This is also when the expert gets their time back. They stop running the process and start owning the rulebook, which is a better use of what they know.
Step 6: Keep the human for the exceptions
Some share of runs will always need judgment. Design the exit ramp on purpose. The automation stops, states what it found, and hands it to a named person with everything it looked up attached. A good escalation takes five minutes instead of twenty, because the digging is already done.
What you get even if you stop at step 2
A rulebook. The process is written down, with its exceptions, in a single file. Someone else can run it from the document tomorrow. That alone converts a key-person risk into a training document, and it costs a few hours of recording and a few hours of writing.
Where to start
Pick the process where the risk is highest, which usually means the one that touches money or customers and has the fewest people who understand it. If you're not sure how to rank them, the scoring method in the 5 business processes you should automate first works here too: frequency, time, and cost of error. Add a fourth factor, how many people can run it today. If the answer is one, it goes to the top.
If you'd rather have help with the recording and the rulebook, that's the kind of thing a free automation audit covers. We'll pick the process, map it, and tell you what to automate first.