Local AI Guy Book a 15 minute call
Playbook

Playbook · which processes to automate first small business

Which job to automate first in a business of twelve people

Three tests decide which job to automate first in a small Ontario business: measured hours, whether the process is written down, and whose data it touches.

Lasse Pettersen

In a business of about twelve people there are usually four obvious candidates for automation and the budget for one. Choosing correctly is worth more than building well, because a build aimed at the third best job returns third best forever.

Three tests decide it, in this order. Measured hours. Whether the process is written down. Whose personal information it touches. The second test eliminates more candidates than the other two combined, and almost nobody applies it first.

Test one: measure the hours, do not estimate them

Owners underestimate administrative work by roughly a factor of two and overestimate how often the awkward exceptions occur that they believe prevent automation. Both errors point the same way: the boring job is bigger than it looks and less variable than it feels.

The measurement is not complicated. For one week, whoever does the task writes down the start and stop time each occurrence. Not a study, not a stopwatch, just a line in a notebook. One week is enough to be directionally right and short enough that people actually do it.

Then the arithmetic, out loud:

  • Minutes per occurrence, times occurrences per week, equals hours per week.
  • Hours per week, times 50, times loaded hourly cost, equals dollars per year.
  • Compare against a build estimate. On this site those are published: $2,500 to $12,000 for workflow automation.

A working threshold in a small Ontario business paying $30 to $40 loaded per hour is around five hours a week. That is roughly $8,000 to $10,000 a year, which pays back a mid-range build inside a year. Below three hours a week the arithmetic rarely works on labour alone.

The exception is risk. A task that takes two hours a week but where a missed step costs a contract is not a labour saving, it is an insurance policy, and it should be valued as one. Certificate and qualification tracking is the usual example: an expired document discovered by an auditor rather than by you is not a two hour problem.

That exception is not rare in Ontario. A contractor working inside the Sarnia petrochemical complex, a supplier to the mines around Greater Sudbury, or an engineering firm in the Pickering energy supply chain all carry qualification obligations set by a customer several orders of magnitude larger than themselves. Two hours a week of tracking dated records is trivial against one crew standing down at a gate, and that comparison, not the hours, is the one to put in front of whoever signs.

Test two: is the process written down anywhere

This is the test that kills candidates, and it is the one worth applying before you get attached to an idea.

Ask the person who does the job to write down every step, in order, including what they do when something is missing. Not a diagram. Sentences. If they cannot, the process does not exist in a form anything can act on. It exists as judgement, accumulated, in one person.

Two things happen when they do write it down, and both are useful:

A third of the process disappears. Steps that exist because of a constraint that was removed years ago become obvious the moment they are read in sequence. This is free and it happens before any software is bought.

The exceptions get counted. People believe their process is 40% exceptions. Written down and tallied for a week, it is usually 5% to 10%, and the exceptions cluster into two or three named cases rather than being infinitely various. That changes a project from impossible to straightforward.

If nobody will write it down, choose a different candidate. A process that only exists in one person’s head cannot be automated, and attempting it produces a system that encodes what somebody guessed that person does.

Test three: whose personal information does it touch

Not a formality, and not last because it matters least. It is third because it is a filter rather than a ranking.

A process touching your own internal data can use whatever tool is best. A process touching your customers’ or your staff’s personal information cannot, and the reasons are set out in what PIPEDA means for your AI tools. The practical effects are a shorter eligible tool list, a written data boundary in the scope, and usually a higher price.

That does not make it a bad candidate. It makes it a differently priced one, and knowing that before you rank the shortlist prevents the common outcome where the winner is reranked after the quote arrives.

The four that come up most, and how they usually score

CandidateTypical hours a weekWritten down?Personal informationUsual verdict
Double entry between two systems10 to 25Yes, implicitlySometimesAlmost always first
Quoting, where one person writes them all8 to 15RarelyNoFirst, once written down
The monthly report rebuilt by hand1.5YesNoReal, and rarely worth it yet
After-hours enquiries going unansweredNot a time costNoYes, by designDepends on missed call volume

The third row is the one businesses most often pick and most often should not. Six hours a month is 72 hours a year, roughly $2,300, against a $3,000 to $5,000 build. It is a genuine automation with a two year payback, which means it is a next year project.

The fourth row is different in kind, because the loss is revenue rather than hours. You never learn about the job that called the next company on the list. Sizing it means asking your phone provider how many calls arrive outside working hours and how many go unanswered, which takes one email. If the answer is meaningful, an intake agent usually beats every internal candidate on return, and that case is made in the four automations that pay in a trades business.

Two rules about scope

Build one thing completely, including the failures. A half-automated process needs somebody watching it, which means you have replaced a manual job with a monitoring job. Completeness means the automation knows when it did not work, tells somebody, and leaves the record in a state a person can pick up.

Do not automate a process you are about to replace. If a new system is being implemented in six months, wait. This comes up constantly in fast-growing businesses and the honest answer costs a supplier the work, which is why you rarely hear it.

Why this is worth a fixed fee to get right

The three tests take about ten business days to run properly in a business of this size, most of it spent watching how work is actually done rather than how it is described. That is what the $999 assessment is, and the fee comes off any build that follows.

You can run it yourself. The tests above are the whole method and there is nothing proprietary in them. What is hard to do from inside a business is noticing the person whose job title says one thing while their actual day is spent being the interface between two systems. That role is created by growth, is almost never designed, and is invisible to everyone including the person doing it.

Statistics Canada found that 40.0% of Canadian businesses not adopting AI said it was not relevant to what they do. In most cases what that means is that nobody has counted the hours yet. Counting them is the project.

Questions on this

How many hours a week does a task need before automating it is worth it?

As a working threshold, around five hours a week of genuinely repeated work in a business paying $30 to $40 loaded per hour. That is roughly $8,000 to $10,000 a year, which pays back a mid-range build inside about a year. Below three hours a week the arithmetic rarely works unless the task also carries risk, and risk is a separate and often larger number.

Should I automate the task that annoys people most?

Not automatically, but pay attention to it. Annoyance is a decent signal for repetition and a poor signal for volume. The task everyone complains about is sometimes a two-hour monthly job that is simply unpleasant. Measure it before you fund it, then decide.

What if the process is not written down anywhere?

Then writing it down is the first project, and it is not a delay. Describing a process end to end typically removes a third of it before any software is involved, because the steps that exist only because of an old constraint become visible. A process that lives in one person's head cannot be automated until it leaves it.

Is it better to automate one thing well or several things partly?

One thing, completely, including the failure cases. A half-automated process needs a human watching it, which means you have added a monitoring job rather than removed a manual one. Partial automation is the most common way these projects fail to produce the saving they were sold on.

Local AI Guy

Before you spend anything

Tell me how many people work there, what the busiest hour of the week looks like, and which task everybody complains about. That is usually enough to say on a first call whether an assessment is worth your $999 or whether you have one obvious problem that needs one obvious fix.

Reply within one business day · Mon to Fri, 9am to 5pm Eastern

Book a 15 minute call Call (705) 555-0182