AI Saved Your Team Time. It Also Handed Them a Checking Job Nobody Wrote Down. Measure the Net Hours.
AI Saved Your Team Time. It Also Handed Them a Checking Job Nobody Wrote Down. Measure the Net Hours.
A two-column tracker for AI tools you already use, and a simple rule for deciding when a task isn't worth automating.
Ask a finance team whether AI has saved them time and most will say yes. Ask how much time they now spend checking what it produces, and you'll usually get a pause. Nobody has counted.
A survey released last Thursday puts some numbers on that pause. Accounting Seed, a US accounting software vendor, asked nearly 200 business and finance professionals at small and midsize companies about AI at work. 81% said AI had saved them meaningful time. 70% said it had handed them new responsibilities without taking any existing ones away. 54% said they often have to correct AI outputs before they can use them.
Those three numbers sit together without contradiction. AI saves time on the doing, and creates time on the checking. The question for a finance lead is the net, and who carries it.
|
70%
Say AI handed them new responsibilities without removing any existing ones (Accounting Seed, November 2026)
|
54%
Often have to correct AI-generated outputs before they can use them
|
|
46% vs 28%
Finance team members vs executives who say AI hasn't delivered the productivity gains they expected
|
96%
Of 270 large-company finance leaders in a separate Datarails survey spend at least 10% of their workday verifying or correcting finance AI outputs
|
A fair caution before going further: both surveys come from software vendors, neither is Australian (Accounting Seed's release frames its sample around US small and midsize companies, and Datarails surveyed 270 finance leaders at companies with at least 1,000 employees), and the Accounting Seed sample is small. Treat them as a signal, not as Australian evidence. But the pattern they describe will be familiar to anyone running a finance team that has started using AI in earnest.
Why the checking job stays invisible
The most telling number in the Accounting Seed release is the gap between levels. 77% of executives said AI had freed up time for higher-value work. Only 61% of finance team members agreed. And nearly half of team members (46%) said AI hadn't delivered the gains they expected, against 28% of executives.
That gap has a simple explanation. The saving is visible from above: the report arrived faster, the coding was done overnight. The checking happens below: someone reads every line of AI-drafted commentary, re-ties the figures, fixes the three invoices coded to the wrong account. It doesn't appear in any system. It's just a longer day.
Left alone, this creates two problems. Leaders overestimate the saving and plan as if it's real, sometimes in headcount. And the people doing the checking, under pressure, start checking less. That's the point where AI errors start reaching the management reporting pack.
The two-column tracker
We've written before about building the measurement before a pilot starts. This is the version for tools you're already using, where nobody captured a baseline. It's deliberately simple. For two ordinary working weeks, everyone who uses AI on a recurring finance task logs two numbers per task, per week:
- Time saved: their honest estimate of how long the task used to take, less how long the AI-assisted version takes to produce.
- Time checking: every minute spent reviewing, correcting, re-running or cleaning data so the tool could work. Data clean-up counts. The same survey found 45% of respondents spend time cleaning up data just to make AI tools usable.
Add one more column: who did the checking. That column usually tells you more than the hours do.
Here's what a fortnight of results might look like for a small team. The numbers are illustrative, not from a real organisation:
| Task (hours per week) | Saved | Checking | Net | Checked by |
| Supplier invoice coding | 6.0 | 2.0 | +4.0 | AP officer |
| Bank rec matching | 3.0 | 0.5 | +2.5 | Accountant |
| Variance commentary drafts | 3.0 | 2.75 | +0.25 | Finance manager |
| Grant report narrative | 2.0 | 2.5 | −0.5 | Finance manager |
| Total | 14.0 | 7.75 | +6.25 |
The headline saving is 14 hours a week. The real one is just over six. And the most senior person on the team, who was meant to be freed up, is now spending an extra five and a quarter hours a week checking AI drafts, and on their own two tasks the net is slightly negative. In this example, the whole gain sits with the routine tasks lower down.
When a task isn't worth automating
The tracker gives you the evidence. You still need a rule for acting on it. Mine is a rule of thumb from practice, not a research finding, and it's simple:
- If checking consistently takes more than about half the time saved, look hard at the task. Either the inputs are poor (fix the data, not the prompt) or the task needs judgement the tool doesn't have.
- If checking needs the same expertise as doing, the saving is often an illusion. Variance commentary and grant narratives are the usual examples. The person who checks them has to understand the numbers as well as if they'd written it themselves.
- If the net is still negative after a month, and you've already tried fixing the inputs and the instructions, stop using AI on that task and say so. That's a good result, not a failure. It frees the time for tasks where the net is positive.
This connects to a point we've made about cost: the real unit is the cost per accepted output, not the price per use. Review time is the biggest part of that cost, and it's paid in your team's hours rather than on the invoice.
Decide who keeps the net hours
Once you know the real saving, decide on purpose where it goes. If nobody decides, the hours get absorbed by whatever is loudest that week, and in six months you won't be able to point to anything.
Options worth naming in writing: a faster close, a proper look at the aged debtors, a monthly cash forecast instead of quarterly, or simply less overtime for an overstretched team. And when executives start talking about AI and headcount, put the tracker in front of them. The survey's exec-versus-team gap is exactly the conversation it helps settle.
Can AI track its own checking time?
Partly. Many tools keep a record of what they produced and when. Pairing that with a simple "accepted / corrected / rejected" tag from the reviewer gives you a correction rate per task without asking anyone to fill in a timesheet. A correction rate that stays high is a strong sign the checking time is real, even before anyone counts minutes. But the minutes themselves are best estimated by the people doing the work. Two weeks of honest estimates beat any automated guess.
Starting next Monday
You don't need a project to do this. A shared spreadsheet with five columns (task, person, minutes saved, minutes checking, checked by) and a two-minute entry at the end of each day is enough. Pick two weeks that include a month-end, because that's when checking time peaks and when the saving matters most.
At the end of the fortnight, sort by net hours and look at the bottom three rows first. Then take a one-page summary to your next management meeting: the headline saving, the real saving, where the checking sits, and which tasks you propose to keep, fix or stop. It's the kind of evidence that turns an argument about whether AI "works" into a decision about where it works for you.
Repeat it every six months, or whenever a tool changes. Models and features move quickly, and a task that wasn't worth automating in November may be in May.
Is your AI saving the hours you think it is?
PFL provides senior-level outsourced finance, management reporting, and AI automation for Australian NFP, NDIS, and SME organisations. We can help you measure the real net saving, decide which tasks stay automated, and make sure the review work sits with the right person.
Book your free health check →
Comments
Post a Comment