Which jobs a 5-to-20 person business should hand to AI first — and how to check it's saving money

2026-08-28

Most small businesses are past the question of whether to try these tools. The problem now is a handful of subscriptions on the company card, one or two people who use them daily, everyone else who signed up once and stopped, and a renewal date approaching with no clear answer on whether any of it saved money.

The Australian Bureau of Statistics reported on 25 June 2026 that business adoption of artificial intelligence accelerated in 2024-25. The National AI Centre publishes periodic adoption insights, the most recent covering December 2025 to February 2026 and published on 7 May 2026. Third-party reporting also tracks readiness and barriers among Australian small and medium businesses, though that work is commercially produced and should be read as indicative rather than official. What none of it can tell you is whether your particular spend is worth it. That is an arithmetic job, and it is one you have to do yourself.

A four-part test for choosing the first task

Don't start with the most exciting task. Start with the one that passes all four of these:

High volume. It happens weekly at minimum, ideally daily. A task done twice a year will never repay the effort of setting it up.
Low judgement. The work is assembly and rephrasing, not deciding. If the value is in someone's read of a client or a market, leave it alone.
Tolerant of error. A wrong output is annoying, not dangerous or expensive. Nobody gets underquoted, misdiagnosed or misadvised.
Already checked by a human. Someone reviews the output before it leaves the business anyway. You are replacing the blank page, not the reviewer.

If a task fails any one of these, it is not your first task. It may be your fifth.

The jobs that usually qualify

The wins in a 5-to-20 person business are dull. That is the point. Dull tasks are high volume, low judgement and already checked.

Quote and proposal first drafts. Not the pricing. The scope description, the inclusions and exclusions, the covering note. Someone with authority sets the numbers and signs off.
Email triage and drafting. Sorting an inbox into categories, drafting standard replies to enquiries that arrive in the same three or four shapes every week.
Meeting notes and action lists. Recording, transcribing and summarising internal meetings and site visits. Check your obligations and get consent before recording anyone.
Position descriptions, ads and induction material. First drafts only, then a human pass for tone and accuracy.
Tidying and reformatting data. Cleaning up supplier lists, standardising addresses, turning messy notes into a consistent format.
Turning one piece of content into several. A case study into a client email, a LinkedIn post and a line for the capability statement.

What usually does not qualify first: customer-facing chatbots. They are the most visible option and the most likely to embarrass you. They sit in front of customers, they answer without a human check, and errors are public. Also leave alone anything touching tax positions, legal drafting, safety documentation or clinical decisions — those need a qualified professional, and a plausible-sounding draft is worse than a blank page because it is harder to spot as wrong.

Expect the first six to eight weeks to cost you

This is the part vendor material skips. When you hand over a task, you spend real hours writing prompts, finding the failure modes, fixing the output, and rewriting the process so the rest of the team can use it. During that period you are net negative on time. Someone senior is usually doing the work.

Plan for it. Give one task to one person for six to eight weeks. Do not roll out four tools to twelve people in a month — you will get twelve half-learned habits and no way to attribute any change to anything.

Working out whether it actually saved money

You need three numbers, and you need the first one before you start.

1. The baseline. For two weeks, have the person doing the task record how long it takes and how often. Rough is fine. Twenty quotes a month at forty minutes each is a usable baseline. Guessing after the fact is not.

2. The fully loaded cost. Not the headline seat price. Add every seat you are actually paying for, including the ones nobody logs into. Add any usage charges above the base plan. Add the setup hours at the internal cost of the person who spent them, spread over the subscription term. Add whatever your bookkeeper or IT support charged to sort out access.

3. The hours actually saved, valued honestly. Re-measure the same task after the learning period. The saving is the difference, minus review time — the checking is part of the new process, not a rounding error.

Then value the hours at what they are genuinely worth to you. If the saved time belongs to an owner or a billable specialist who now does more chargeable or revenue-generating work, value it at that rate. If it belongs to an admin role where nothing else changes, the honest value is close to zero. That is uncomfortable and it is the single most common reason claimed savings never turn up in the accounts.

As an illustration of the method only: if a task took eighteen hours a month and now takes seven including review, that is eleven hours. If those eleven hours go into work that generates revenue, compare their value to the total monthly cost from step two. If the eleven hours go into nothing in particular, you have bought idle time, not savings.

What happens to the saved hours

Decide this before you start, in writing. There are only three honest destinations: more revenue-producing work, a role you no longer need to hire for, or fewer hours worked. Anything else and the productivity gain quietly disappears into a slightly longer lunch.

Say the decision out loud to the team too. People who suspect the aim is to cut their job will not help you find the savings.

Before the renewal date

Run a short review on each subscription:

Who logged in during the last month? Not who has a seat — who used it.
Which specific task is it attached to, and what is the measured before-and-after?
Does another tool you already pay for do the same job adequately?
If it disappeared tomorrow, what breaks?

Cancel anything that fails the first two questions. Consolidate where two tools overlap. Keep and expand only what has a measured task behind it.

This week

List every subscription and what it costs. Pick one task that passes the four-part test. Baseline it for two weeks before touching anything. Name one person to own it and give them a review date eight weeks out. Write down where the saved hours will go.

One task, measured properly, will tell you more than ten tools bought on the strength of a demo. If the task involves personal information, financial records or anything covered by professional obligations, get advice from someone qualified in that field before the data goes anywhere near a third-party tool.

Sources

Written by ARC Editorial, drafted and reviewed with claude-opus-5. ARC publishes these to help business owners find answers; if something here matters to your situation, a Facilitator can point you at someone who has solved it.