What to automate first

Score the candidates before you pick one

Everyone has a list of things they wish were automatic. The list is usually ordered by annoyance, which correlates poorly with return. Reorder it with four scores, one to five each.

Frequency: how many times a month does it run? Something running daily has fifty times the leverage of something running quarterly, regardless of how tedious the quarterly job feels.

Duration: how long does one run take, including the interruption cost of switching to it and back? A three-minute task that breaks concentration ten times a day costs more than its clock time suggests.

Rule clarity: could you write the decision logic on one page such that a new hire would reach the same outcome every time? If not, the process is judgment, and judgment automates badly.

Cost of error: what happens when it is done wrong or skipped? A missed inquiry, a wrong invoice or a compliance lapse all raise the value of removing the human step.

Add the four. Anything scoring 16 or above is a strong first candidate. Below 10, leave it alone no matter how irritating it is.

Look at handoffs, not tasks

The instinct is to automate a job someone does. The better instinct is to automate the seam between two jobs, because that is where information gets retyped, delayed or dropped.

Walk one customer through your business on paper, from first contact to paid invoice, and mark every point where a detail moves from one system or person to another. Each mark is a candidate, and each mark is also a place where errors currently enter.

Handoffs have a structural advantage: they are usually mechanical rather than judgmental, so the rules are genuinely rules. They also tend to cross departmental boundaries, which means the time saved is visible to more than one person — helpful when you want the second project approved.

The five that almost always pay back

Across small service businesses, the same handful of workflows come out on top. If you have not automated these, start here rather than looking for something bespoke.

• Lead capture and response routing. An inquiry creates a record, alerts the person on duty, sends an immediate acknowledgment to the customer, and escalates if nobody replies within a set window. Small build, runs constantly, moves revenue directly.
• Quote and proposal generation. Structured inputs produce a formatted document from a template, logged and sent. Saves the most time per run of anything on this list and removes a whole class of copy-paste errors.
• Appointment scheduling and reminders. Self-service booking against real availability, with confirmation and reminder messages. Cuts no-shows and eliminates the scheduling phone tag entirely.
• Invoice and payment follow-up. Automatic issuance on job completion, then a polite escalating reminder sequence. Almost every business does this inconsistently by hand, and the inconsistency is expensive.
• Recurring reporting. Numbers pulled from wherever they live and assembled on a schedule. Two hours every Monday is a hundred hours a year of skilled time for a build that rarely exceeds a few thousand dollars.

Sequence matters more than speed

Build one workflow, get it live, and let it run for a month before starting the next. This is not caution for its own sake — the first build tells you things that change the second one.

It reveals data quality. Inconsistent customer records, missing fields and three spellings of the same company name all surface immediately once something automatic depends on them, and cleaning that up is a prerequisite for everything downstream.

It also sets the platform and the data model. If the first workflow is built on tooling that will extend, the second costs less than the first. If it is built as a one-off with a throwaway integration, the second is a restart. Ask about this explicitly before the first build starts.

Finally, it builds trust. Teams that watch one automation work reliably will bring you the next candidate themselves. Teams whose first experience was a half-finished build that silently failed will quietly route around every automation you add afterwards.

What not to automate yet

Three categories look attractive and reliably disappoint as a first project.

Unsettled processes. If two people in the business would describe the current process differently, you are paying to automate a disagreement. Write it down, get agreement, then automate. The writing-down step alone often recovers most of the value.

Genuine judgment calls. Pricing an unusual job, deciding whether to escalate a difficult client, reading whether a customer is about to leave. You can automate the information gathering around these and leave the decision to a person; trying to encode the decision produces rules with endless exceptions.

Rare, high-stakes events. An annual compliance filing is stressful but infrequent, so the arithmetic never works. Document it thoroughly instead and revisit only if it becomes monthly.

One more: do not automate a task you should simply stop doing. A surprising share of candidate workflows are reports nobody reads and approvals nobody acts on. Ask who consumes the output before you spend anything making it faster.

Questions people ask next