How to choose what to automate first

What should we automate first?

Automate the boring, frequent, rule-based process that currently costs you the most and annoys the most people. In most South African small and mid-sized businesses that is data entry between two systems, reminders that depend on someone remembering, and reports that are rebuilt from scratch each month. It is rarely the customer-facing process everyone wants to talk about.

The reason this is the right first choice is arithmetic. A process you do forty times a month and save ten minutes on returns 6 600 minutes a year, which is over a hundred hours. A process you do twice a year, no matter how elegant, will not change anything. Volume is the whole argument.

Is automation the right answer at all?

Sometimes it is not, and finding that out early is worth a week of work. Three cases where the answer is no:

• The process happens too rarely to be worth building anything. Doing it by hand twice a year costs you almost nothing.
• The process changes every month. You will be rebuilding the automation as fast as you build it.
• The problem is not the process, it is a missing decision. An unclear rule cannot be automated, only guessed at, and a guessed rule is a new class of error.

Before automating, try the cheap version first. Delete the step. Combine two forms into one. Get the customer to send the information in the format you need instead of reformatting it. Fix the process first, automate the improved version second. Automating a broken process just runs the breakage faster.

How do you list the processes that are worth automating?

Spend a week writing down every recurring task anyone does, without judging any of them. Ask the people who do the work, not the owner, because the owner remembers the big things and the team remembers the ones that eat the day.

A useful prompt list:

• Anything copied from one screen or message into another.
• Anything sent out on a schedule, even informally.
• Anything assembled by pulling figures from two or more places.
• Anything checked by a person because getting it wrong has consequences.
• Anything a customer waits for, where you would like them to wait less.
• Anything you re-do because the first attempt was done the wrong way.
• Anything done at month end that takes longer than it should.

Then cut the list down. Remove anything that happens fewer than a few times a month, anything nobody owns, and anything where you cannot name a person who would notice if it stopped happening. A process with no owner has no definition of done, and no definition of done means no safe automation.

How do you score candidates and pick one?

Score each survivor from one to five on five factors, then add them up. The score is only a way to force a conversation, so argue about the scores rather than about the answer.

The winner is normally a high scorer on frequency, time and rule clarity, because those three are what make automation reliable. A process scoring five on error cost but two on rule clarity is a candidate for a checklist and a second check, not for automation.

How much does a process need to cost before it is worth automating?

Work out the cost before you start, so you have a number to return to. Do it in this order:

annual time cost = (runs per month) x (minutes per run) x (loaded hourly cost) x 12

plus: the cost of errors caused today
plus: the value of the delay, if the output being late costs you something
minus: what a manual fix would cost, such as a checklist or a template

Use your own loaded hourly cost. For a person paid R28 an hour who costs the business R40 an hour to employ, use R40. If the process also needs a manager to check it, include that share too, because the checking is often the more expensive half.

Once you have the figure, a useful threshold: if a process costs less than a few thousand rand a year, the build will struggle to pay for itself, and you should do the manual fix instead. A fuller worked example is in the companion templates.

Which processes should you leave alone?

Some things are not worth automating, and recognising them saves you from a project that fails.

• Anything requiring genuine judgement about a person. Who gets the discount, which supplier you trust, whether a customer complaint is serious.
• Anything that happens very occasionally, even if it is painful.
• One-off work, migrations and clean-ups. There is no second run to amortise the build against.
• Anything a customer is actively waiting for and you want a human on.
• Processes where the source data is not reliable. Automate that later, once the data is trustworthy.

There is a fifth case worth naming: don't automate a process nobody has agreed should exist. Automation applied to an unnecessary step makes the unnecessary step happen faster and more reliably, which is a very expensive way to waste money.

What does a first automation actually involve?

Keep the first one small. A first automation that works in three weeks and saves ten hours a month builds confidence and funds the next one. A first automation that takes four months burns the goodwill you need for months two and six.

A workable shape:

1. Write down the current process, including every exception anyone has to handle.
2. Define the trigger precisely. What event starts it, and how does the system know?
3. Set the rules, and write down what happens when a rule cannot be satisfied. The exception path is the part people forget, and it is where automation projects die.
4. Build it to run alongside the manual process, not instead of it, for the first month.
5. Compare outputs side by side. If they disagree, fix the rules before you switch over.
6. Switch over, then delete the manual step so nobody quietly keeps doing both.

How do you know it worked?

Measure the same thing you measured before, using the same method, six weeks later. Specifically: the number of runs per month, the minutes per run, the number of errors, and the time between the trigger and the output being right.

If the numbers have not moved, the automation did not work, and the useful outcome is knowing which rule was wrong. Say that openly on your next audit, because the alternative — quietly keeping a half-working automation alongside the manual process — is how a small team ends up doing double work and blaming the tool.

The companion checklist walks you through a first automation from process list to measurement, and the templates include a candidate process inventory, a prioritisation scoring sheet, a cost-of-delay calculator, and a thirty-day plan you can put in front of a team on a Monday morning.

Still weighing two candidates against each other? We can score them with you. Book a free strategy session with Agentcy.