Ask ten business owners how much time their AI tools save and you'll get ten different answers, most of them guesses dressed up as facts. "Hours a week" is not a number you can put in a board pack or a budget submission. If you're going to make a case for expanding AI use, or defend the spend you've already committed, you need a measurement method that survives someone asking "how do you know?"
Why "it saves us hours" isn't a number
Most time-savings claims come from a single enthusiastic staff member describing their best day. That's a real experience, but it isn't data. It skips the setup time, the corrections, the days the tool wasn't used at all, and the fact that saved minutes on a task don't automatically turn into saved dollars unless that time gets redeployed to something billable or valuable.
A proper time-and-motion study borrows from lean manufacturing: you measure the task as it actually happens, before and after the change, across enough occurrences to smooth out the noise. It is slower to set up than asking someone how they feel about the new tool, but the number that comes out the other end is defensible in front of an accountant, a bank, or a business partner who wasn't in the room when the tool was adopted.
This matters more than it sounds. Once an AI tool is embedded in daily work, the temptation is to justify the licence cost with a vague sense that things feel faster. Feelings don't survive a budget review. A measured number does, and it also tells you honestly where the tool is and isn't earning its keep.
A four-step method that holds up under scrutiny
Baseline first. Time the task the old way, at least 10 to 15 occurrences, across different staff and different days of the week. Write down the actual minutes, not an estimate made from memory.
Isolate the task, not the role. Measure the specific unit of work, drafting a client email, reconciling a ledger line, summarising a call, rather than someone's whole day, which is contaminated by interruptions and unrelated work.
Run the same measurement post-adoption. Same task, same sample size, same observer where possible. Include the review and correction time. If a person has to check Claude's output before it goes out, that checking time counts against the saving, not in favour of it.
Convert minutes to dollars using a loaded rate. Use full employment cost per hour, including superannuation and on-costs, not just the award wage. A saving that looks generous at the base rate often looks modest once it is loaded properly.
None of this requires special software. A stopwatch, a shared spreadsheet, and a willingness to sit with someone for an afternoon will get you a number that a finance manager in Sydney, Melbourne, or anywhere else will accept without an argument. The method is the credibility. Anyone can produce an estimate; fewer businesses bother to test it.
Where the measurement usually goes wrong
Measuring the best case instead of the typical case, then presenting it as the average across the whole team.
Forgetting to count the time spent writing prompts, fixing formatting, or re-running a request that came back wrong the first time.
Treating time saved as automatically equal to money saved, when the freed-up hours are actually just absorbed into other unstructured work.
Sampling for a single week right after rollout, when everyone is paying close attention and performing above their normal baseline.
Comparing a rushed manual process against a carefully prompted AI one, rather than holding quality constant on both sides.
A worked example: a Sydney bookkeeping practice
A ten-person bookkeeping firm in Sydney wanted to know whether Claude was worth the seat licences before renewing. They picked one task: drafting the plain-English summary that accompanies each client's monthly management report. Before the trial, five bookkeepers timed themselves drafting 40 summaries between them. Average time: 34 minutes each, at a fully loaded staff cost of $58 an hour, giving a baseline cost of roughly $32.90 per summary.
After four weeks using Claude to produce a first draft that a bookkeeper then reviewed and edited, the same task took an average of 11 minutes, including the review step. That works out to $10.63 per summary at the same loaded rate, a saving of about $22 per summary. Across roughly 400 summaries a year, that is a little over $8,800 in freed staff time on this one task alone, before counting the two other report types the firm later applied the same method to.
Scaled across the practice's full reporting workload, the firm's honest, sampled estimate landed at roughly $45,000 a year in redeployable staff time, well short of the six-figure claims some vendors like to wave around, but a number the practice's director could put in front of the partners without hedging or rounding up.
Turning hours into a defensible dollar figure
The last step is the one people skip: deciding what the freed time is actually worth. If a bookkeeper spends the saved 23 minutes on more billable client work, that time has real dollar value and belongs in the calculation. If it just makes their afternoon less rushed with no change to output or billable hours, the honest answer is that you have improved working conditions, not cut costs. Both outcomes are worth having, but only one belongs in an ROI figure.
If your firm handles client financial or personal information as part of the task you are measuring, keep the methodology note explicit about what data went into any AI tool during the trial. This matters for your obligations under the Privacy Act 1988, and it makes the eventual business case easier to sign off internally, because nobody has to go back and ask what was actually shared with the tool during testing.
Run the same four-step method again in six months. Tools improve, staff get faster at prompting, and the honest number often moves, in either direction. Treat the first measurement as a baseline for the practice, not a permanent verdict.
If you want help designing a measurement study for your own team rather than guessing at the number, book a short call and we will set up a baseline that holds up to scrutiny.
