Three recent papers on what actually happens when accountants work with AI, and why the most interesting finding isn’t the speed.
Here’s the short version, and then I’ll earn it: AI is making accounting work faster, and we can now measure it instead of guessing. But the most useful finding isn’t the speed. It’s what happens to the human in the loop. They don’t disappear. They move.
I’ve been reading three papers that approach this from different directions: a stopwatch experiment from Portugal, a field study from Stanford built on more than 200,000 real transactions, and a conceptual paper from Oxford arguing that finance is moving from AI that predicts to AI that acts.
Read together, they tell a surprisingly coherent story.
The stopwatch study
Miguel Resende, at the University of Aveiro, took one routine task, the vertical analysis of an income statement, and ran it three ways: manually, with a custom GPT, and with an AI agent.
The task involved expressing every line of an income statement as a percentage of revenue and writing an executive summary for the client.
The AI agent went further than generating a summary. Built using the no-code platform MAKE, it monitored a Google Drive folder, ran the analysis through OpenAI and Gemini, stored the results in Google Sheets for review and, after human approval, emailed the reports to clients.
Each method was tested 50 times across 50 anonymized Portuguese small businesses over three months.
Here’s what came out:
| Method | Average time per report |
|---|---|
| Manual | 622.4 seconds (10 min 22 sec) |
| Custom GPT | 281.2 seconds (4 min 41 sec) |
| AI agent | 153.5 seconds (2 min 33 sec) |
Going from manual work to GPT cut the time by 55%. Moving from GPT to the agent reduced it by another 45%. Overall, the agent was about 75% faster than doing the task by hand.
But the number I keep coming back to isn’t the average. It’s the drift.
When Resende plotted the manual runs over time, execution times crept upward. He attributes this to repetitive work and human fatigue. If you’ve ever done the fortieth report of the day, you know exactly what that feels like.
The GPT version also showed some drift because a person still had to upload, copy and send each report individually. The agent was the only one whose execution times remained flat.
That’s the real difference between a chatbot and an agent. The GPT made the thinking part faster. The agent eliminated the shuffling: uploading, copy-pasting and emailing.
And that shuffling turned out to be a significant part of those ten minutes.
One important detail: even with the agent, a person reviewed every summary before it went out. The accountant wasn’t replaced. The work changed.
The field study
The Portuguese experiment is useful, but it covers one task in a controlled setting. What happens in actual accounting practice?
That’s what Jung Ho Choi and Chloe Xie investigate in Human + AI in Accounting: Early Evidence from the Field, published in the Journal of Accounting Research.
Their research draws on a survey of 277 professional accountants, data from an AI-enabled accounting platform serving 79 small and mid-sized businesses, more than 200,000 transaction-level records, and a field experiment.
The headline findings from the working-paper version:
- AI adopters saw a 55% increase in weekly client support.
- Around 8.5% of accountant time shifted away from routine data entry towards business communication and quality assurance.
- General ledgers became 12% more granular.
- Monthly closes finished 7.5 days faster.
That last finding deserves a pause. We’re not talking about shaving minutes off a report. That’s more than a week off the close.
Look at where the time went, too. Not simply away from accounting, but towards talking to clients and checking work.
That aligns with Resende’s findings. The machine takes over repetitive work, while the person moves towards review and communication. Two very different studies, pointing in the same direction.
The survey also found considerable variation in how accountants adopt AI, perceive its benefits and respond to its risks. They’re not all embracing it at the same pace, and they don’t necessarily agree on what it’s good for.
The most interesting finding
For me, the most interesting part of the Stanford paper isn’t a productivity number. It’s a behaviour.
The accounting platform assigns confidence scores to AI-generated transaction classifications. Researchers found that accountants selectively intervened when those scores were low. Experienced professionals, in particular, used automation selectively, stepping in when the AI signalled uncertainty.
That’s what good collaboration looks like. Nobody checks every line, and nobody blindly accepts every recommendation. The experienced accountant recognises uncertainty and directs their attention where it matters.
But there’s a warning here.
AI assistance improved classification accuracy on average. However, relying on AI recommendations that experts disagreed with could increase the risk of error.
In other words, the average result can improve while individual decisions get worse. If you’re looking only at overall accuracy, you might miss the mistakes that matter.
The conclusion is worth remembering: AI works best when it augments professional judgment rather than replaces it.
The bigger shift: From predicting to doing
The third paper takes a broader view. In The Agentic Turn in Finance: AI Agents in Financial Decision-Making, Marc Schmitt of the University of Oxford examines how AI is changing finance.
For years, AI in finance largely meant prediction. A model generated a forecast or score, and a human decided what to do with it.
What’s emerging now is delegated decision-making. AI systems can interpret financial situations, pursue assigned goals, plan multiple steps, use tools, retain information and adapt based on feedback.
Schmitt calls this the Agentic Finance Framework. It distinguishes genuine AI agents from predictive models and decision-support tools, examining applications across investment research, trading, portfolio management, risk and compliance, and customer-facing services.
His central argument is that once AI can act, accuracy isn’t the only concern. You also need to know what the system is authorised to do and who remains accountable when something goes wrong.
That’s a shift from managing model risk to managing agent risk.
Schmitt writes about finance broadly, not bookkeeping. But look at Resende’s experiment through this lens. The custom GPT drafted a summary, leaving a person to handle everything else. The agent monitored files, ran the analysis, recorded results and sent approved reports to clients.
It’s a small, relatively low-stakes example of the delegation Schmitt describes, already being implemented in small-business accounting using no-code tools.
So what does this mean if you run a practice?
Pulling the three papers together, here’s what I take away.
The speed is real, but the shuffle is where the hours hide. Resende’s findings show that automating uploads, copying and sending can matter as much as speeding up the analysis. If all you’ve given your team is a chatbot, you may have captured only part of the opportunity.
The job moves; it doesn’t vanish. Across the empirical studies, human effort shifts towards validation, client communication and quality assurance. Accountants increasingly need analytical reasoning, data interpretation and AI oversight skills.
Expertise means knowing when to doubt the machine. The confidence-score finding is particularly practical. The skill isn’t simply knowing how to use AI or distrusting it. It’s recognising which outputs deserve a second look, especially when the system produces an unusual recommendation.
Scale changes the economics. Batch processing allows firms to handle larger client portfolios without a proportional increase in manual effort. Combine that with the reported 55% increase in weekly client support, and the potential to change how accounting practices operate becomes clearer.
When AI acts, governance must catch up. The more responsibility you delegate to an AI system, the more you need clear boundaries, oversight and accountability, not just accuracy.
The caveats
None of these papers is the last word.
Resende’s study covers one repetitive task across 50 Portuguese SMEs in a controlled setting. It doesn’t establish that the same time savings will apply to every accounting process. His agent also still required someone to place statements in the Google Drive folder.
Choi and Xie’s findings come from AI adopters using a single platform, and the research itself is presented as early evidence. Schmitt’s paper proposes a framework rather than measuring productivity gains.
Treat the percentages as signposts, not promises. What’s striking is that all three papers point in a similar direction, despite their different approaches.
Where this leaves us
The honest picture of what AI can do for accounting is it can take over the repetitive work that wears people down, leaving more room for reviewing, exercising judgment and talking to clients.
Those were always the parts that needed a professional.
The machines got the boring part. The judgment is still ours.
AI is changing how accounting gets done, but figuring out where it fits into your practice is a different challenge.
If you’re exploring ways to automate routine finance and accounting work while keeping professional judgment at the centre, let’s talk.
