What to Automate First: A Practical Method for Picking the Right Process

The most expensive automation mistakes are not technical. They are picking the wrong process — spending forty thousand dollars automating something that ran eleven times last year, or automating a process that turns out to need judgement on every instance.

This is the method we use in discovery, written out so you can run it yourself before speaking to anyone.

The four tests

Run every candidate process through these in order. A process needs to pass all four.

Test 1: Frequency — does it run more than fifty times a month?

Below roughly fifty instances a month, engineering cost rarely clears the saving. There are exceptions where a single instance is very expensive or very risky, but they are exceptions.

Count actual instances rather than estimating. People consistently overestimate the frequency of annoying tasks and underestimate the frequency of routine ones.

Test 2: Rules — could a competent new employee learn this in a week?

This is the single best proxy for automatability.

If the answer is yes, the process has explicit rules and those rules can be encoded. If the answer is "well, it depends, you sort of get a feel for it" — that is judgement, and judgement automates badly. You will get a system that is confidently wrong on the cases that matter.

A useful sub-test: could you write the rules down in two pages? If you cannot, neither can we.

Test 3: Cost — what does one instance actually cost today?

Time multiplied by loaded hourly rate. Loaded means salary plus superannuation plus overhead, which is typically 1.3 to 1.5 times base salary.

Measure rather than estimate. Ask someone to time it for a week. The number is almost always different from what management assumes — sometimes far lower, occasionally much higher once the interruption cost is counted.

Test 4: Failure — what happens when it goes wrong?

The most-skipped test and the one that determines your budget.

If a mistake is cheap and obvious, the automation can be simple. If a mistake is expensive or invisible — a wrong price on a quote that goes out, a missed compliance step — the system needs validation, logging and escalation logic. That checking layer is frequently more than half the build cost.

Be honest here. Underestimating this is why projects run over.

Scoring it

Score each candidate 1 to 5 on four axes, and multiply rather than add — multiplication punishes a weak score properly, which is what you want.

Anything scoring above about 200 is a strong candidate. Below 60, leave it alone for now.

Run this across eight or ten processes. The ranking is usually surprising — the process everybody complains about is often not the one worth automating first, because complaining correlates with annoyance rather than cost.

The three that usually win

Across Australian service businesses, three processes come out on top repeatedly.

After-hours call handling

High frequency, clear rules for the first conversation, high cost per instance because a missed call is often a lost job, and contained failure because escalation to a human is straightforward.

Typical numbers: five to fifteen missed calls a week outside hours, average job value of several hundred dollars, and roughly half of missed calls representing real work.

Quote generation from a fixed price list

The qualifier matters. If your quoting pulls from a defined price list with consistent margin rules, it automates well and typically takes an evening task down to minutes.

If you eyeball each job and adjust on instinct, it fails Test 2 and you should not automate it. You will get quotes that are precisely wrong.

Document handling and data entry between systems

Invoices arriving as PDFs that need to become records. Job sheets becoming schedule entries. Two systems that were never designed to talk to each other.

High frequency, explicit rules, moderate cost per instance, and failures are usually detectable if you build the validation. This is unglamorous and reliably worth doing.

What usually fails the tests

Anything involving negotiation or persuasion. Fails Test 2.

Genuinely bespoke creative work. Fails Test 2, and the output quality bar is subjective.

Compliance-critical processes without a checking layer. Passes Tests 1 to 3 and fails Test 4 badly. These are automatable but the budget has to include the validation, and that is often two-thirds of the cost.

Anything you are about to change. If the process is under review, wait. Automating a moving target means building it twice.

The two-week test before you spend anything

Before commissioning work, run the process manually with an off-the-shelf assistant for two weeks and log every instance: how long it took, where it broke, what needed a human decision.

You will end up in one of three places.

1. Manual plus assistant is fine. Stop. You have saved yourself a project.
2. It works but the volume hurts. Automate the specific bottleneck rather than the whole process. This is usually much cheaper than a full build.
3. It keeps needing information the assistant cannot reach. That is your build, and now you can specify it precisely rather than describing it vaguely.

That fortnight is the cheapest specification document you will ever produce. It also gives you a genuine before-measurement, which is the only way to know later whether the thing worked.

One thing worth saying

We are a consultancy that builds automation. It is worth noticing that this method regularly produces the answer "do not automate this yet", and we would rather you reach that conclusion in an afternoon than after an invoice.

The businesses that get real value from automation are the ones who picked two processes and did them properly. The ones who get nothing are usually the ones who bought a platform and hoped.