Expense reports: what one costs to process

What this page records

Every article on this topic quotes the same $58 figure with the confidence of a fact that has been checked exactly once, by someone else, a long time ago. Processing an expense report is one of the few operations tasks with a published price on it, and the number everybody quotes is $58. That figure deserves less weight than it gets: it was published in 2015, it covers a single night’s hotel stay rather than an average claim, and the study behind it was produced by a travel trade association in partnership with a commercial supplier. The figures that hold up better are less flattering to the usual business case. One in five reports carries an error or a gap, at $52 and 18 minutes each to fix. Expense reimbursement is the third most common occupational fraud scheme at 13% of cases, and it is also the second slowest, bleeding a median $1,900 a month for a median 18 months before anyone notices. And the machine that is supposed to read the receipt scores 70% on a blurred image and 92% on a sharp one, which makes photograph quality a larger variable than model choice. Each record below carries the figure, who was counted, over what period, and the publication it was read from, so any single row can be checked without trusting this page. What a build in this area has to hand over at the end is what running an expense system costs after go-live.

What the process profile says

Six structural dimensions decide whether a process is worth automating, and expense reports sit at an unusual corner of that space: the individual amounts are trivial, the individual errors are cheap, and the aggregate is neither. This is a high-frequency, low-stakes-per-event process wrapped in a tax rule that makes a procedural failure expensive at the level of the whole arrangement rather than the single claim.

Three of those rows point the same way, and they are not the rows the usual pitch is built on. The exception-variance row says the hard input is a photograph, not a transaction. The cost-of-an-error row says no single claim is worth much attention. The reversibility row says the correction window is short and the detection window is long. Put together, they describe a process where the return on automation comes from doing a cheap check on every claim continuously, and not from doing a careful check on any particular one.

The volume row is what separates expense reports from the reporting and reconciliation processes alongside them. Those arrive on a calendar at a size known in advance. Claims arrive whenever somebody travels, in a stream that grows with headcount, which means throughput reasoning genuinely applies and a per-claim saving genuinely multiplies. That is the honest half of the standard business case, and it is the half worth keeping.

The approval process, and what it is actually for

An expense approval chain looks like a control and mostly is not one. A manager approving a claim sees a total and a category, rarely the receipt, and almost never the comparison between the receipt and what the claimant typed. The paper that supplied this page’s extraction figures found 89 invoices in one multinational’s corpus where the amount read off the receipt was lower than the amount the employee had entered, in a system where reimbursement follows the typed figure. No approval step catches that, because the approver is not reading the image either.

This is what makes the regulatory rows load-bearing rather than decorative. The tax rule does not ask whether a manager approved the claim. It asks whether the expense was substantiated with documentary evidence, whether it had a business connection, and whether excess amounts came back within a reasonable period. An approval workflow that records a decision without recording the evidence behind it satisfies the organisation’s own process and not the rule the organisation is actually exposed to.

The practical consequence is that approval and verification are different jobs, and only one of them is worth automating first. Verification is a comparison between a document and a claim, performed on every claim, and it is the part machines do at a cost per document low enough to be uninteresting. Approval is a judgement about whether the spend should have happened, which needs context no receipt contains. Automating the first does not remove the second; it gives the second something to look at.

What most claims about this get wrong

The genre’s central error is quoting the $58 figure as a current, universal cost per expense report. It is none of those things. It is eleven years old, it is scoped in its own source to a report covering a single night’s hotel stay, and it comes from a trade association study run with a commercial partner whose interest was in the answer. The figure is cited on this page because it is the only widely circulated per-report cost with a traceable publication and date behind it, and because a reader deserves to see what it actually is. A number that cannot be traced to a live primary publication with a date is not a statistic however often it is repeated, and most of the cost-per-report figures in circulation fail that test entirely.

The second error is selling expense automation on fraud prevention while quoting the wrong fraud numbers. Expense reimbursement fraud is common, at 13% of cases in a study of 2,402. It is also, by the same study, one of the two slowest schemes measured, at a median $1,900 a month against $9,400 across all schemes. The organisations that lose spectacular sums to fraud do not lose them here. What is genuinely striking in that data is the duration: a median 18 months undetected, second only to financial statement fraud. The problem expense fraud presents is not magnitude, it is latency, and those call for different systems. A control that reduces the size of a loss is not the same as a control that shortens the time to finding it.

The third error is reading benchmark extraction scores as a forecast for a particular company’s receipts. The same labelling method that scored 94% against exact labels on a public benchmark of 1,000 receipts scored 50%, 70% and 76% on merchant name, amount and date against a real multinational’s own expense documents. Some of that gap is the reference labels being imperfect rather than the extraction being wrong, and the authors say so. Most of the rest is image sharpness: about 90% of that organisation’s receipts were sharp enough to score above 90%, against about 60% of the public benchmark’s. The variable that decides whether receipt reading works is how the photographs are taken, which is a policy and interface question rather than a modelling one.

What this page cannot show you

It cannot tell any particular organisation what its own expense processing costs. The published per-report figure is old and narrowly scoped, the regulator’s burden estimate measures the traveller’s time rather than the employer’s, and neither is a substitute for counting the hours a specific finance team spends on a specific claim volume. Nothing here is tax advice, and no position is taken on whether any particular expense is deductible, substantiated or reimbursable under any arrangement. The regulations are cited for what they require a system to keep, not for how any transaction should be treated.

It also cannot say whether the association between proactive data monitoring and shorter fraud duration is causal. The study compares organisations that had the control against organisations that did not, across cases that were all eventually detected. Organisations that monitor their data proactively differ from those that do not in many ways beyond the monitoring, and the study does not attempt to separate them. The direction is worth knowing and the size of the effect is not a prediction.

How these figures were compiled

Every record was read from its primary source on 12 August 2026. The trade association’s article was read on its own site, and its publication date taken from the page’s own timestamp. The fraud report was downloaded as a PDF and read in full text, with the scheme table, the duration figure, the velocity figure and the two anti-fraud control tables each located separately rather than taken from a summary. Both regulation sections were read through the Electronic Code of Federal Regulations renderer rather than a secondary restatement, and the paragraph references in the records point at the specific provisions read. The Federal Register notice was read in the government publishing office’s text. The extraction figures were read in the HTML version of the paper, which carries the results tables the abstract does not.

Two sources were dropped rather than approximated. The most frequently cited cost-per-report benchmark from a process benchmarking body could not be retrieved, as its host refuses server-side requests, so no figure from it appears here in paraphrase and none was taken from a search result quoting it. The fraud report’s detection-method data is not broken out at the level of an individual scheme, so no claim is made here about how expense fraud in particular tends to be found, despite that being the most useful thing the report could have said for this page’s purposes.

One methodological point is stated rather than buried. The fraud study is a sample of cases that were detected and investigated by professionals who chose to report them. It cannot describe fraud that was never found, which matters more here than on most subjects, given that the same study measures this scheme’s median time to detection at 18 months.

The verdict

The published evidence supports automating the capture and checking of expense claims, and does not support justifying that work on a per-report cost saving alone. The check is worth automating because it is cheap enough to run on every claim and there is no other point in the process where anyone compares the document to the claim. The overpayment risk measured on one real corpus was about $2 a document, which is uninteresting per claim and is not uninteresting across 51,000 of them. The cost saving is real too, but the figure normally used to size it is eleven years old and scoped to one trip shape, so a business case resting on it is resting on the weakest number in this page.

The stronger argument is about time rather than money. Expense fraud is frequent, small per month and long-lived, running a median 18 months before detection. The only control in the data associated with a materially shorter detection window is continuous monitoring of the transactions themselves, at 9 months against 16, and that association is not proof of cause. Even discounted for that, it points at the same design as the cost argument does: a system whose job is to look at everything cheaply and continuously, rather than to look at anything carefully once. The audited artefact that build has to produce is the comparison it made on each claim and what it did about the difference, because that is the record the accountable-plan rules ask for and the record an investigation 18 months later has to reconstruct.

The brake is the receipt. Extraction accuracy on real corporate expense documents tracks image sharpness far more than it tracks anything a buyer is being sold, moving from 70% to above 90% across that one variable. A finance team weighing this should establish three things before choosing anything: how many claims it actually processes and what the last quarter’s corrections cost, whether its own receipt images are sharp enough for a machine to read, and whether its current approval step ever compares the document against the claimed amount. A team that can answer those is buying on evidence. A team that cannot is being asked to automate a check nobody has yet performed manually, and performing it once on a sample is cheaper than any of the alternatives. If you have run that sample and want a second read on it, send it to us.

Sources

1. GBTA Foundation with HRS, 'How Much Do Expense Reports Really Cost a Company?', published 20 October 2015 - read on gbta.org. A trade association study produced in partnership with a commercial travel supplier Retrieved 12 August 2026
2. Association of Certified Fraud Examiners, 'Occupational Fraud 2026: A Report to the Nations', 14th edition, published May 2026 - scheme table, duration and velocity figures read at report pages 16 to 19, anti-fraud control tables read later in the same report Retrieved 12 August 2026
3. 26 CFR 1.62-2, 'Reimbursements and other expense allowance arrangements' - current text read through the Electronic Code of Federal Regulations renderer, paragraphs (c), (e), (f) and (g) Retrieved 12 August 2026
4. 26 CFR 1.274-5, 'Substantiation requirements' - current text read through the Electronic Code of Federal Regulations renderer, paragraph (c)(2)(iii) on documentary evidence Retrieved 12 August 2026
5. US Nuclear Regulatory Commission, 'Information Collection: NRC Form 64, Travel Voucher', Federal Register document 2024-04078, published 28 February 2024 - read on govinfo.gov Retrieved 12 August 2026
6. US General Services Administration, 'GSA selects IBM to provide technology-managed service for governmentwide travel and expense management', news release dated 27 November 2024 Retrieved 12 August 2026
7. Bhattacharyya and Tripathi, 'Information Extraction from Heterogeneous Documents without Ground Truth Labels using Synthetic Label Generation and Knowledge Distillation', arXiv:2411.14957, submitted 22 November 2024 and last revised 25 November 2024 - results read in the HTML version of v2 Retrieved 12 August 2026

Foxnut Studios works on briefs like this one from Bengaluru and Paris. If you want the shape of that before you talk to anyone, here is what an AI engagement costs before you ask.