The Report Took AI a Few Minutes. The Invoice Said Four Hours.
The Report Took AI a Few Minutes. The Invoice Said Four Hours.
Keeping the time AI saves you as margin is fair. Billing more than the job would take without AI is not. Here is the benchmark I think AI-assisted work should be billed against, and the questions to ask when an invoice doesn't add up.
A client asked its outsourced payroll provider for a simple report: total payroll for the quarter, per employee. The provider had made a point, more than once, of telling the client how much it now relies on Copilot across its service.
The report arrived. It had been built by exporting a summary from the client's accounting system and another from the provider's payroll system, then matching the two by employee with a lookup. Tidy enough. Then the invoice arrived: four hours.
That number is what this post is about. Not because four hours of payroll time is a fortune, and not because the provider kept the time AI saved it. Providers keeping some of that saving is fine; I do too. The problem is that, on everything the client could see, four hours is more than the job would take without AI, done the slow way.
Three ways to do the same job
Matching two system reports by employee is about as routine as finance work gets. Here is roughly how long it takes depending on how you do it, assuming both exports are clean and the employee identifiers line up:
|
Minutes
Both exports into a business-grade AI tool, asked to match by employee and flag differences. Plus a short human check of the result.
|
~30 min
A spreadsheet with an XLOOKUP across the two exports, including checking totals and investigating any mismatches.
|
|
~2 hours
Printing both reports and comparing them line by line, typing each employee's figures in by hand. The slowest honest way to do it.
|
4 hours
What was billed, for a report built with a lookup by a provider promoting its AI use.
|
These are my own working estimates for a small-to-medium headcount, not a published benchmark. Your numbers will move with the size of the workforce and the state of the data. But the shape doesn't change: the invoice was double the time it would take to do the job completely by hand, with pen and paper.
And if the provider genuinely did spend four hours on a clean, routine match, that tells you something too: it's what it costs when a team hasn't yet learned to use the tool it's promoting. That's a real cost. It just isn't the client's to pay.
AI is becoming the baseline, and clients have noticed
Using AI for routine reconciliation and reporting is no longer a differentiator. It's fast becoming the expected way to work, and clients are starting to ask what that means for their bills.
A General Assembly survey of 258 senior leaders at large consulting, accounting and legal firms in the US and UK, reported by Accounting Today in March this year, found 42% said clients were questioning their pricing model because of AI. Among accounting firms it was 35%, and 73% said they were changing how they talk about pricing. That's an overseas sample of big firms, so treat it as a direction rather than an Australian measurement. But the direction is clear, and smaller Australian clients are asking the same question in their own way: if the tool did most of it, why does the bill look the same, or bigger?
There's nothing wrong with a provider using AI, or with it earning a better margin because of it. I use AI every day and would encourage any finance team to do the same. The problem is promoting AI as a selling point and then billing a premium on top of what the work would cost without it.
To be fair: what can legitimately take longer
Not every long invoice is a padded one. Some things genuinely add time to a job like this, and a good provider should be able to point to them:
Messy data. Employees missing from one system, duplicate records, terminated staff still active, different identifiers in each system. Cleaning that up is real work.
Investigating differences. If the two systems disagree, working out why (a late adjustment, a back-pay, a misposted journal) takes judgement and time, and is often the most valuable part of the job.
Review and sign-off. A second person checking the output before it goes to the client should be billed. Checking AI output isn't optional; it's what makes the output trustworthy.
First-time set-up. If it's the first time the report has been built, designing the format and the matching logic takes longer than running it next quarter.
If any of those happened, the invoice should say so. "4.0 hours — quarterly payroll report" doesn't. And if the next quarter's version is billed at the same four hours, the set-up argument has run out. A simpler habit avoids most of this: ask for a quote before the work starts, then compare it with what the job would take if you gave it to the Excel guru on your own team, and decide for yourself whether it's worth sending out.
Five questions to ask about an invoice for AI-assisted work
None of these are hostile. Any provider confident in its billing will answer them easily.
1. What was produced? One report, or the report plus an exceptions list and an explanation of each difference? The second is worth more.
2. How was it produced? Which systems were the data pulled from, and which tools were used to match it? If AI was used, that's fine. You just want to know.
3. Where did the time go? A split between extraction, matching, investigating differences and review tells you far more than a single line.
4. What did the review find? If the answer is "nothing, it all matched", then most of the time was mechanical, and mechanical work is exactly what AI shrinks.
5. Can this be a fixed fee next time? Repeatable reports are ideal for fixed pricing. A provider that resists pricing a routine, recurring output is usually telling you something.
The benchmark: what the job takes without AI
Since this is a criticism of a provider, it's only fair to say what I think the right standard is. It applies to us as much as anyone.
The baseline is a competent person doing the work without AI. For this report, that's the spreadsheet version: two exports, a lookup and a check, around half an hour. That's what the job is worth in time, whatever tool actually does it.
If AI was used to save time, not to add quality, bill no more than the baseline. If a person then verified the output, the provider has two honest choices. Bill the baseline and keep the saving as margin, which is a fair reward for investing in the tools and the skills. Or pass some of the saving on to the client. Both are reasonable. Many good firms will do a bit of each.
If AI added genuine quality, say what it added. An exceptions list, a reconciliation to the general ledger, an explanation of every difference. That's more than the baseline report, and it can be priced as more, as long as the client can see what the extra is.
Learning the tool is the provider's cost. If the work took longer than the baseline because the team is still working out how to use the AI it's promoting, that time belongs to the provider's training budget, not the client's invoice.
In my view, billing above the manual baseline is the one option that should be off the table. Adding a premium for work that is now faster, while marketing the tool that made it faster, is an approach from before the tools existed. It won't survive clients who use the same tools themselves.
Where I've landed
I'll be honest: I don't pass every minute AI saves me back to clients either. The saving is part of what makes investing in the tools worthwhile. What I don't do is inflate the hours. I'd rather bill the minimum the work genuinely justifies, which in practice usually sits at the lower end of what comparable advisory work costs, so that both sides come out ahead. The provider earns a sustainable margin, and the client pays less than it would have before AI.
That, to me, is what using AI well in professional services should look like: a win for both sides, not a new way to charge more for less. Isn't that the direction we should all be heading?
Note: The scenarios in this post are based on real experiences — mine and those shared by colleagues across the sector. Details might have been changed and modified slightly to protect confidentiality, and mostly used 1st person perspective for convenience.
The time estimates in this post are the author's working estimates, not a published benchmark.
Not sure what your finance and payroll work should cost?
Our free Finance Function Health Check looks at what you're currently paying for, which recurring work could sit on a fixed fee, and where AI can genuinely save you time. PFL provides senior-level outsourced finance, management reporting, and AI automation for Australian NFP, NDIS, and SME organisations.
Book your free health check →
Comments
Post a Comment