Most owners are not deciding whether to buy these tools. They are already paying for them, bundled into an accounting package, a CRM, an email client or a document editor, with a button that has appeared in the toolbar. The question has moved on. It is now: which job do I point this at first, and how do I know it did anything?
That second half is where most attempts fall over. A task feels faster. Nobody measured it. The subscription renews. Nine months later there is a line item and no evidence.
A triage rule for picking the first task
Before looking at tools, look at tasks. A back-office job is a reasonable first candidate if it passes all four of these:
- High volume. It happens weekly at least, ideally daily. A task done twice a year cannot repay the setup effort.
- Low consequence if it is wrong. Being subtly wrong should cost an awkward sentence, not money, a client or a compliance problem.
- Text-shaped. Words in, words out. Summarising, drafting, rewriting, sorting, tagging. Not calculations that must reconcile, and not judgement calls.
- Reviewable at a glance. A human must be able to tell it is wrong in seconds. If checking the output takes as long as producing it, there is no saving.
That last test is the one vendor listicles never mention, and it is the one that decides the outcome.
Tasks that usually pass
- First drafts of routine emails and enquiry replies, from a short brief
- Turning a meeting recording or transcript into notes and an action list
- Summarising a long document down to the three things a decision depends on
- Rewriting an internal note into client-ready language
- Drafting position descriptions, job ads and induction checklists from an existing template
- Turning an existing process, described out loud, into a first-draft written procedure
- Sorting and tagging inbound enquiries so they land with the right person
- Cleaning up inconsistent text in a spreadsheet — product descriptions, address formats, category labels — where a person still eyeballs the result
Tasks that usually fail the test
- Anything where the numbers must be right: payroll, invoicing, reconciliations, quotes with live pricing
- Anything that answers a tax, employment, safety or contractual question. Those need a qualified professional, and a confident-sounding draft is worse than no draft
- Final client-facing advice, or anything sent without a human reading it
- Anything involving sensitive staff or client information pasted into a tool nobody has checked the terms on
- Decisions. Drafting is a task. Deciding is not.
Start with one. Not three. One task, one person, one fortnight.
The two-week baseline
The measurement is unglamorous and takes about ten minutes a day.
Weeks one and two — before. Pick the task. Have whoever does it record three things every time: the date, minutes spent start to finish, and whether it needed rework afterwards. A phone timer and a shared spreadsheet is enough. Also count how many times the task occurred, because frequency is half the maths.
At the end, work out the average minutes per instance and the number of instances per week.
Weeks three and four — after. Same task, same person, same recording, with three extra columns:
- Minutes spent setting up the request or prompt
- Minutes spent reviewing and correcting the output
- Whether the output was used, heavily edited, or thrown away
The number that matters is not the drafting time. It is the total: setup plus generation plus review plus fixing, compared against the original average. Then multiply by instances per week to get the weekly saving in minutes.
Two honest adjustments. First, discard the first two or three attempts — those are learning, not steady state. Second, if the person is also being asked to learn a new tool while running the trial, expect weeks three and four to look worse than the eventual reality. Run a third fortnight before concluding it failed.
The review tax
This is where paper savings disappear. Output that is 80 per cent right still has to be read closely, because the wrong 20 per cent is not marked. For some tasks that reading is quick — a meeting summary, an action list, an internal note. For others, verifying is nearly the whole job.
A practical rule: if the reviewer has to open the source document to check the output, the saving is probably small. If they can tell it is fine by reading it once, the saving is probably real.
Count the "thrown away" column too. A task where one in four outputs is discarded may still be worth it if the other three are quick, but it needs to show up in the numbers rather than in someone's impression.
Cost, honestly
Pricing varies by vendor, by tier and by whether the feature is bundled into software already being paid for or sold as a separate seat. Check the vendor's own pricing page for current figures rather than relying on anything written about them — these change often and without notice.
Two buying disciplines worth holding to:
- Prove a saving on one task using something already paid for, before adding a new subscription.
- Convert the measured saving into money the same way payroll is costed, then compare it to the annual price of the seats. If the answer is close, the answer is no. Marginal tools create work through the sheer effort of keeping them.
Client and staff data — the objection that actually stops adoption
This is the reason most owners hesitate, and the hesitation is reasonable. Before any real data goes into a tool, get answers in writing to four questions:
- Where is the data stored, and in which country?
- Is the input used to train the vendor's models, and can that be switched off?
- What tier is being used — a consumer account and a business or enterprise account often have different terms for exactly this.
- Who at the business can access what?
Then write a one-page internal rule: what may be pasted in, what must be redacted first, and what never goes near it. Names, tax file numbers, health information, bank details, unpublished financials and anything covered by a client confidentiality clause are the usual "never" list.
Privacy obligations differ depending on the size and nature of the business and the kind of information it holds. That is not a question to settle from an article. Check the Office of the Australian Information Commissioner's current guidance, and if client contracts or sensitive personal information are involved, get legal advice specific to the business.
What to do this week
- List the back-office tasks that happen at least weekly.
- Score each against the four tests: volume, consequence, text-shaped, reviewable at a glance.
- Pick the single highest scorer and nominate one person to own it.
- Start the before-measurement on Monday. Two weeks, three columns, no tool yet.
- While that runs, get the four data questions answered by whichever vendor is already in the stack.
The baseline is the whole exercise. Without it, the only available evidence is whether the work felt faster, and that is not evidence.
Sources
No figures, rates or thresholds are cited in this article, as none were verified for it. For the matters that depend on outside authority:
- Privacy obligations and current guidance — Office of the Australian Information Commissioner
- Pricing, data storage location and model-training terms — the vendor's own current pricing and terms pages