AI expense audit is the automated review of every expense claim before reimbursement. A platform tests each claim in seven layers, from document extraction and image provenance through merchant resolution, arithmetic validation, cross-report matching, anomaly scoring, and disposition. Each layer catches something the others structurally cannot see, and each one fails differently.
Almost every expense platform now advertises audit driven by artificial intelligence. A good share of what ships under that label is a rules engine with a classifier bolted to the front. The term has a technical meaning, and it comes down to which layers actually run. A platform tests each claim in the seven layers below, and each layer has the failure mode the others exist to cover. Three questions at the end separate a real implementation from a relabeled one.
Vendors adopted the label faster than the technology, because the problem underneath got loud in a hurry.
PYMNTS reported our platform data in 2026. Documents produced by generative AI rose from zero percent of flagged fraudulent receipts in March 2025 to 70.8 percent by mid-May 2026, according to that coverage of AI-generated fake receipts. The analysis covered 1,471 fake receipts from 745 employees across 174 companies, worth $148,143 in fabricated reimbursements. The average AI-generated fake came in around $100, against $182 for older template-based fakes. Cheaper forgeries mean higher volume, filed below the amounts that draw a reviewer's eye.
Duration is what turns that into money. The Association of Certified Fraud Examiners studied 2,402 cases across 143 countries and territories for Occupational Fraud 2026. The median loss was $104,000 per case, with a median of 12 months before detection. The median loss was $40,000 in schemes caught inside six months.
An expense audit is an evidence problem before it is a technology problem. Every line item makes an assertion, meaning the employee claims that this merchant charged this amount on this date for this business purpose. The employee offers a document as proof. The audit tests whether the document supports the assertion, and whether the assertion is permitted.
That splits into two jobs vendors tend to blur. Extraction turns a picture into structured fields. Verification tests those fields against something outside the document. A system with excellent extraction and no verification reads a forgery perfectly and passes it, because nothing it does leaves the four corners of the file.
No single model covers verification either. An image forensics model says nothing about whether the same dinner was claimed six weeks ago. A duplicate matcher says nothing about whether the restaurant exists. The failure modes are independent, so detection has to be layered.
This layer turns a photo, a PDF, or a multi-page folio into labeled fields. Merchant, date, currency, line items, tax, and total get roles rather than positions. Optical character recognition returns text, and document understanding returns structure.
It catches nothing on its own, and it decides whether anything else works. A misread tax line quietly breaks validation downstream. It fails on unfamiliar layouts, low-contrast photographs, handwriting, and folios where a description runs across two columns. A mature implementation scores confidence per field rather than per document.
This layer examines the file instead of reading it. Metadata, generator signatures, compression history, resampling traces, and font rendering all belong here. Our research on detecting AI-generated fake receipts covers how the individual checks work.
Provenance catches the forgery that is internally flawless, where the arithmetic adds up and the merchant is real. Nothing else in the stack sees that document as suspicious. It fails on a laundered file, meaning a photograph of a screen, a print and rescan, or an app that strips metadata on send. It also misfires when a scanner rewrites a legitimate receipt.
This layer resolves a printed merchant name to a real business. It then tests the address, the category, and the naming variants against the rest of the organization's data. Entity screening belongs here too, including politically exposed person checks on attendees.
It catches invented merchants and real names attached to impossible locations. A generated receipt reads fine alone and looks wrong against external reference data. Resolution fails on businesses with a thin external footprint, such as market stalls, new restaurants, and franchises trading under several names. Those need to return low confidence rather than a flag.
This layer tests whether the numbers agree with each other and with the world. Line items should sum to the subtotal. Tax should match a rate that exists in that jurisdiction. Tip percentages, currency conversions, and rounding conventions all get checked here.
Generators produce plausible figures rather than consistent ones, so a large share of fabrications break at this layer. Tax computing to 8.8 percent in a 6.5 percent jurisdiction is the strongest single arithmetic signal. Validation misfires on legitimate documents with service charges, split tenders, or loyalty redemptions applied after tax.
This layer compares the claim against the submitter's history, colleagues traveling on the same dates, the corporate card feed, and prior periods. Matching on amount, date, and merchant catches the easy cases. Matching on the receipt image catches submissions that arrive months apart or through different systems.
It is the only layer that sees a duplicate at all, because a single report contains no evidence that the employee filed the same claim before. It fails early in a deployment, when there is no history to match against. In our own data, teams that test systematically find the number of duplicate expenses they identify grows 700 percent between month one and month twelve.
This layer learns what normal looks like for an employee, a cost center, a merchant category, and a city, then scores departures from it. It is the only part of the stack that surfaces something nobody wrote a rule for, and new patterns appear faster than rule libraries get updated.
Anomaly output is a prioritization signal rather than a finding. A relocation or a first trip to an expensive city produces a legitimate outlier. Models drift as travel patterns shift, so a team that treats scores as verdicts trains its reviewers to dismiss them.
This layer decides what happens next. It sorts each exception by risk and dollar value, auto-approves clean reports, queries the employee on a missing document, and escalates the rest. It writes the finding with the line item, the policy clause, and the observed evidence attached.
It detects nothing, and it routes what the other six layers find to the right reviewer. An audit that routes every exception into one queue creates a backlog. Teams then tune thresholds down until the queue is manageable, which quietly restores the coverage gap the AI was bought to close.
Three limits are worth stating plainly, because a vendor who will not state them is selling something else.
It does not replace judgment on genuinely ambiguous policy calls. Whether a Saturday dinner with a client served a business purpose is a decision a person makes from context outside the system.
It does not establish intent. A regenerated receipt for a real purchase produces the same signals as a fabrication, so a finding has to describe what the evidence shows.
It does not see spend that never enters the system. Card charges that bypass expense reporting stay invisible until the card feed itself gets audited.
Three questions do most of the work on an evaluation call. Ask for each answer in the product rather than the deck.
What happens on a receipt from a merchant the system has never seen? A rules engine passes it, because no rule fires. A real implementation resolves the entity against external data and returns low confidence, which is a different outcome from silence.
Does the system explain its finding? Ask to see raw output on one flagged report. Look for the field, the observed value, the expected value, and the source of that expectation. A risk score with no reasoning underneath cannot be sent to an employee.
Is coverage complete or sampled? Ask for the denominator in writing, since most teams still review 10 to 20 percent of expense transactions, a gap examined in our analysis of 100 percent audit coverage. Ask whether card transactions get the same treatment as reimbursements.
Extraction is where the market has genuinely converged. Receipt capture works, card feeds arrive clean, and mileage tracking is solved. Most products stop there and hand the extracted fields to a rules engine, which catches only what somebody already anticipated and encoded as a rule.
The symptom is timing. New fraud patterns stay invisible until a person notices a trend in a report. By that point the duration clock in the ACFE data has been running for months.
We built Expense Audit around the whole stack rather than a single model. Our AI reads every line of every receipt on every report before reimbursement. We cover 100 percent of expense reports and 100 percent of card transactions as purchases post.
The layers map onto the architecture above. Image provenance, pattern recognition, merchant authentication, mathematical validation, and completeness verification run in one pass rather than in sequence. A document that clears one check still has to clear the rest. More than 40 pre-built travel and expense (T&E) audit models apply those tests across 42 languages and 97 countries, including fapiao and value-added tax validation that manual programs apply unevenly.
The platform sorts every exception it raises by risk and dollar value. AI Agents resolve the routine exceptions on their own, and AI Analytics turns the pattern data into something a controller acts on. Clean reports clear without human attention, which leaves auditors on the judgment calls the models are not there to make.
Ask a current vendor which of these seven layers their product runs, then ask to see the output of each one on a real report from last quarter. The gap between the layers you have and the layers you assumed you had is usually where the losses occur. Our AI expense audit page describes full pre-payment coverage at enterprise volume.
AI expense audit is the automated review of every expense claim before reimbursement. Models extract the receipt data, test the document's authenticity and arithmetic, verify the merchant against external data, compare the claim against history, and route each exception for disposition.
A rules engine flags only what someone anticipated and encoded as a threshold. AI expense audit adds verification against evidence outside the document, matching across reports and periods, and anomaly models that surface patterns nobody wrote a rule for.
It misses spend that never enters the expense system, and it does not settle ambiguous business-purpose judgments. It also does not establish intent, since a regenerated receipt for a real purchase produces the same signals as a fabricated one.