AI in Property Accounting: What the Evidence Shows
Median cost to process an invoice is $5.83 and median cycle time is 15 days in APQC data. Here is what neutral research says AI changes, and what is unproven.

Most numbers quoted about AI and accounts payable come from companies that sell AP software. This report leaves those out. It uses a nonprofit benchmarking body, a university working paper, analyst survey results reported by trade press, and a professional association survey to answer three questions: what invoice processing costs today, what AI has been shown to change, and what remains unproven.
No neutral body publishes these benchmarks for property management. The figures below are cross-industry. The last sections explain how to apply them to a property portfolio and what to measure yourself.
Key findings
- The median organization spent $5.83 to process one invoice. The best quarter spent $2.07 or less and the worst quarter spent $10 or more. Source: APQC benchmarking data from 1,485 organizations, reported in CFO, February 2018.
- The median organization takes 15 days from receiving an invoice to transmitting payment. The fastest quarter averages 12 days and the slowest averages 24. Source: APQC data reported in CFO, December 2025.
- 58 percent of finance functions were using AI in 2024, up from 37 percent in 2023. In 2025 the figure was 59 percent. Source: Gartner surveys of 121 and 183 finance leaders, reported by the Journal of Accountancy and CFO Dive.
- In field data from 79 small and mid-sized firms, accountants using generative AI closed the month 7.5 days faster and produced general ledgers with 12 percent more detail. Source: Choi and Xie, Stanford Graduate School of Business working paper, May 2025.
- 91 percent of finance leaders in the 2025 Gartner survey described the impact of AI so far as low or moderate. Source: CFO Dive, November 2025.
What does invoice processing cost today?
APQC is a nonprofit benchmarking organization. Its accounts payable measure is the total cost of the AP process divided by invoices processed. The cost includes labor, systems, overhead and outsourcing.
| APQC measure | Best quarter | Median | Worst quarter | Reported |
|---|---|---|---|---|
| Total cost to process one invoice | $2.07 or less | $5.83 | $10 or more | February 2018, 1,485 organizations |
| Days from invoice receipt to payment transmitted | 12 days (average at or below the 25th percentile) | 15 days | 24 days (average in the upper quartile) | December 2025, sample size not stated |
Sources: CFO, "Metric of the Month: Accounts Payable Cost" and CFO, "Hidden delays that slow payments and strain cash flow". Both columns are by Perry D. Wiggins, identified in the 2018 column as APQC's chief financial officer, and both draw on APQC's database.
The cost figures are from 2018 and have not been adjusted for inflation. APQC's full tables are available to members only, and we could not find a more recent public release of the cost measure. Treat $5.83 as a floor for a current median.
The gap between best and worst is the main lesson. A separate APQC analysis from May 2018 put it in ratios: top performers' AP cost per invoice was one-fifth that of the bottom quartile, they processed five times as many invoices per full-time employee, and they took half as many days from receipt to payment. Source: APQC, "4 KPIs That Set Good Accounts Payable Organizations Apart".
Why do invoices take two weeks?
The December 2025 column names three causes. Invoices are misrouted to the wrong inbox or department and sit. Invoices arrive with errors in tax, price or quantity and bounce back to the supplier. And some companies hold payment on purpose to manage cash, which is a choice and should not be counted as waste.
The first two are where automation is aimed. The third is a reminder that a shorter cycle is only better when it is the result you wanted.
How widely is AI used in finance?
Gartner surveys finance leaders each year. Its press releases were not readable by us, so the figures below are taken from coverage in the Journal of Accountancy, published by the AICPA, and in CFO Dive.
| Survey year | Finance functions using AI | Respondents | Most common uses reported |
|---|---|---|---|
| 2023 | 37% | Not stated in coverage | Not stated in coverage |
| 2024 | 58% | 121 finance leaders | Intelligent process automation 44%, anomaly and error detection 39%, analytics 28%, operational assistance 27% |
| 2025 | 59% | 183 CFOs and senior finance leaders | Knowledge management 49%, accounts payable process automation 37%, error and anomaly detection 34% |
Sources: Journal of Accountancy, September 16, 2024 and CFO Dive, November 19, 2025.
Adoption jumped in 2024 and then stopped rising. The 2025 coverage reports that data literacy, technical skills and data quality were the largest obstacles, that 16 percent of finance departments had no AI implementation planned for the coming year, and that 91 percent of respondents saw low or moderate impact so far. Teams further along in adoption were more than twice as likely to report moderate impact.
These samples are small and the respondents are finance leaders at companies that answer analyst surveys. The numbers show direction. They are not a census.
What has AI been shown to change in accounting work?
The strongest public evidence is a 2025 working paper by Jung Ho Choi of Stanford Graduate School of Business and Chloe Xie of MIT Sloan. It combines a survey of 277 accountants with field data from 79 small and mid-sized firms that use an AI-based accounting platform, covering hundreds of thousands of transactions.
| Outcome for AI adopters | Reported effect |
|---|---|
| Monthly close | 7.5 days faster |
| General ledger granularity | 12 percent higher |
| Weekly client support | 55 percent higher |
| Accountant time | About 8.5 percent moved from routine data entry to work such as communication and quality assurance |
Source: Choi and Xie, "Human + AI in Accounting: Early Evidence from the Field", Stanford GSB Working Paper No. 4261, May 2025.
Three cautions apply.
- It is a working paper and has not completed peer review.
- The field data comes from firms working with one AI accounting software provider. Firms that chose that software may differ from firms that did not.
- The firms are small and mid-sized. Nothing in the paper is specific to property management or to utility invoices.
The paper's finding on behaviour is as useful as its numbers. Experienced accountants did not accept everything. They used automation selectively and stepped in more often when the software reported low confidence. The authors read this as AI adding to professional judgment. A pilot experiment in the same paper looks at what happens when the AI is wrong, and its results are described as preliminary.
Where does AI not yet have neutral evidence?
- Extraction accuracy on real invoices. We found no neutral, published benchmark for how accurately AI reads utility or vendor invoices in production. Published accuracy figures come from vendors, including ours.
- Cost per invoice after AI. APQC's public figures do not separate organizations by automation level.
- Property management. No neutral body publishes AP cost or cycle-time benchmarks for property operators.
- Fraud. The Association for Financial Professionals reported that 76 percent of US organizations experienced attempted or actual payments fraud in 2025, that business email compromise affected 74 percent, and that 17 percent used AI to combat payments fraud. The survey covered 465 treasury practitioners and was underwritten by a bank. It measures exposure. It does not show that AI reduces losses. Source: AFP, 2026 Payments Fraud and Control Survey press release.
How do these benchmarks apply to a property portfolio?
Property accounting has a feature most industries lack: a large share of its invoices are utility bills. They arrive monthly for every account, from many providers, in many formats, and each one has to be coded to a property, a unit or meter, and a general ledger account. Late payment carries a fee and, eventually, a shutoff.
That makes the APQC measures easy to translate. The arithmetic below is an illustration using an assumed portfolio. It is not a benchmark.
| Assumption or result | Value |
|---|---|
| Utility accounts (assumed) | 1,000 |
| Invoices per year (12 per account) | 12,000 |
| Annual processing cost at APQC worst quarter ($10) | $120,000 |
| Annual processing cost at APQC median ($5.83) | $69,960 |
| Annual processing cost at APQC best quarter ($2.07) | $24,840 |
The spread between worst and best quarter is $95,160 a year for this assumed portfolio. Whether a given operator sits near $10 or near $2 is something only its own data can show.
Cycle time matters more for utilities than for most payables. If a provider allows three weeks from bill date to due date, a 15-day median process leaves little room and a 24-day one leaves none. We did not find a public source for typical utility due-date windows, so check the terms on your own bills.
What should a property accounting team measure?
Measure these before and after any change, using APQC's definitions so the numbers are comparable.
- Cost per invoice. Total AP cost, including staff, systems and outsourcing, divided by invoices processed.
- Cycle time. Calendar days from invoice receipt to payment transmitted.
- Invoices per full-time employee.
- First-pass rate. The share of invoices coded and approved with no human correction. There is no neutral benchmark for this, so track your own trend.
- Late fees paid. In dollars and as a count of invoices.
- Days to close. The working paper above suggests this is where AI shows up most clearly.
When a vendor quotes an accuracy figure, ask what it measures: fields or whole invoices, which fields, on whose documents, and whether a person reviewed the output before it was counted.
How should AI be controlled?
The research supports a simple operating rule. Let the software handle the routine invoices and send the uncertain ones to a person. The Stanford and MIT paper found that experienced accountants already work this way.
In practice that means a confidence threshold on every extracted field, a review queue for anything below it, a record of who approved what, and segregation of duties that software cannot override. Payment fraud statistics are a reason to keep human approval on changes to vendor bank details, whatever the invoice workflow looks like.
Method and sources
All figures were read from the linked pages in November 2026. APQC figures are as reported by APQC staff in CFO and on APQC's site. Gartner figures are as reported by the Journal of Accountancy and CFO Dive, because Gartner's own releases could not be retrieved. The portfolio table is arithmetic on stated assumptions. Nothing in the research sections comes from AirBills customer data.
- CFO, "Metric of the Month: Accounts Payable Cost", February 5, 2018. APQC cost per invoice.
- CFO, "Hidden delays that slow payments and strain cash flow", December 3, 2025. APQC cycle time.
- APQC, "4 KPIs That Set Good Accounts Payable Organizations Apart", May 22, 2018.
- Journal of Accountancy, "Organizations turning to AI in finance function", September 16, 2024.
- CFO Dive, "CFOs' AI adoption slows as challenges mount", November 19, 2025.
- Choi and Xie, "Human + AI in Accounting: Early Evidence from the Field", Stanford GSB Working Paper No. 4261, May 7, 2025.
- Association for Financial Professionals, 2026 Payments Fraud and Control Survey press release, April 14, 2026.
How AirBills helps
These are AirBills' own figures, not neutral benchmarks. AirBills AI-powered accounting captures utility invoices with 99%+ accuracy, cuts bill processing time by 90% and cuts close time by 70%. It scores each extraction for confidence, codes bills to the general ledger with reviewer override, and flags duplicates and rate spikes. Customers go live in 2 to 4 weeks.