.svg)
Invoice fraud does not succeed because AP teams are careless. It succeeds because the AP workflow was never designed to catch it.
Phacet Labs describes the problem precisely: "The standard invoice approval process assumes that invoices arriving from known suppliers, referencing real transactions, for plausible amounts, should be paid. Fraudsters design their attacks to satisfy exactly those assumptions." The result is a category of financial loss where the attack succeeds not because the controls failed, but because the relevant controls were never triggered.
Human review catches the obvious. A duplicate invoice with an identical reference number. A new supplier with an implausible name. An invoice for an amount wildly out of proportion to anything the business has ever purchased. These things get caught.
What manual review consistently misses is the subtle. A duplicate invoice with a reference number that differs by one digit. A supplier whose invoiced amounts have been creeping upward by 3% per month for six months. A bank detail change that arrives on a Friday afternoon, two days before the monthly payment run.
At the volume most mid-market finance teams process, catching these patterns through manual review is not a question of diligence. It is a question of what is structurally possible.
A finance team processing 500 invoices a month cannot give every invoice meaningful scrutiny. The operational reality is that invoices are checked against the purchase order and delivery note, routed for approval, and processed. The review is transactional, not analytical. It confirms that the invoice matches what was ordered and received. It does not compare this invoice against the supplier's full history, check whether the amount is consistent with seasonal patterns, or flag that this is the third time this supplier has invoiced for an amount just below the approval threshold.
That analytical layer requires time that manual processing does not have. And it requires a view of data across time that no individual reviewer holds in their head.
According to AllAboutAI's research, enterprise-grade AI fraud detection detects anomalies in milliseconds and processes every transaction without the attentional limits that fraud exploits. The comparison is not between AI and a diligent human. It is between AI and the realistic level of attention any team can give to each invoice in a high-volume processing environment.
Modern invoice fraud is designed around the gaps in manual review. Trustpair's 2026 Fraud Trends Report confirms that 71% of companies faced an increase in AI-powered fraud attempts in the past year. Fraudsters are using AI to generate more convincing fake documents, identify the optimal timing for fraud attempts, and target the specific weaknesses in each organisation's approval process.
The timing exploitation is particularly significant. Month-end is when fraud most often succeeds in businesses with manual review, because that is when approval pressure is highest and scrutiny per invoice is lowest. A fraudulent invoice submitted during a period when the AP team is under maximum pressure to process quickly has a statistically higher chance of clearing without detailed review than the same invoice submitted mid-month.
An AI system does not have a month-end. It applies the same analysis to every invoice regardless of when it arrives.
We covered in our invoice fraud guide that according to the ICAEW, only 14% of invoice fraud incidents are ever reported to authorities. That figure reflects not just embarrassment but the nature of the losses: many instances of invoice fraud go undetected for months or years, during which the losses accumulate. By the time the scheme is discovered, the full scale is often much larger than any single incident would suggest.
The ACFE estimates that companies lose an average of 5% of annual revenues to fraud, with fraud schemes that go undetected for 12 months or more causing the largest average losses. The link between detection speed and loss magnitude is direct.
The most valuable fraud signal is not what is on an individual invoice. It is how a supplier's invoicing behaviour has changed over time relative to their historical baseline.
A supplier who invoiced consistently for £2,000 to £3,000 per month for two years and then begins submitting invoices for £5,000 to £7,000 is exhibiting a deviation that the AP team processing this month's invoice will not notice, because they are not comparing this invoice against two years of history. They are comparing it against the purchase order and the delivery note, which may well be correctly matched.
AI builds a behavioural baseline for every supplier from the full transaction history. When an invoice deviates from that baseline, the anomaly is flagged regardless of whether the invoice passes every other check. The flag is not evidence of fraud. It is a signal that warrants investigation.
Lucid.now's AI anomaly detection research confirms that "AI establishes behavioural norms for departments, vendors, and time periods, flagging deviations without relying on fixed thresholds that often trigger false alarms." The behavioural approach produces fewer false positives than threshold rules while catching the slow-moving, incremental fraud that rules miss entirely.
AI-powered intelligent document processing reads financial documents at a level of detail that visual review cannot match at scale. Font inconsistencies within a single document. Metadata that does not match the claimed creation date. Subtle differences in the formatting of VAT numbers or bank account details that indicate the document was edited rather than generated by the supplier's system.
A human reviewer scanning an invoice for the key fields, amount, purchase order reference, supplier details, will not notice that the font on line 7 is 0.5 points larger than the rest of the document, or that the embedded PDF metadata records a creation date three days after the invoice date. An AI system that has been trained on document forensics catches both of these, consistently, on every invoice processed.
Some fraud schemes are invisible when viewed at the invoice level but visible when patterns are examined across the full history. The AP manager who has been approving invoices from a supplier they control will not reveal that relationship in any single invoice. It becomes visible when the pattern of approvals is examined: invoices from this supplier cluster around certain approvers, approval times are unusually fast, and the invoiced services correlate suspiciously with periods when that approver was responsible for procurement decisions.
This kind of correlation analysis is straightforward for AI and impossible for a human team to run routinely across their full invoice history.
The foundation of AI fraud detection in AP is the behavioural baseline: a model of what normal invoicing looks like for each supplier, built from the full transaction history. The baseline captures the typical invoice frequency, the usual amount range, the normal timing within the month, the payment terms typically applied, and any seasonal patterns.
When a new invoice is received, it is evaluated against this baseline before any other check. An invoice that is statistically consistent with the supplier's history proceeds normally. An invoice that deviates significantly from the baseline is flagged for additional review, regardless of whether it passes the standard matching checks.
Research from the International Journal of Advanced Research (2025) found that modern AI fraud tools achieve 90 to 97% accuracy in detecting anomalies, compared to legacy rule-based systems at 60 to 75%. The gap is primarily attributable to the behavioural approach: AI learns what normal looks like, rather than relying on predefined rules about what fraud looks like.
The distinction between real-time and batch detection matters for fraud prevention. A fraud detection system that reviews invoices in a nightly batch finds fraud after the fact, potentially after an invoice has already been approved and queued for payment. A real-time system flags anomalies at the point of processing, before approval, while the fraud can still be stopped.
Emburse's 2026 AI fraud detection analysis describes modern AI systems as "proactive, agentic defence networks, continuously analysing transactions, detecting anomalies in real time, and autonomously escalating suspicious activity before losses occur." That continuous, real-time coverage is the structural difference between AI detection and any version of periodic manual review.
A fraud detection system that generates too many false positives is one that the team learns to ignore. The alert fatigue problem, where genuine warnings are dismissed because the signal-to-noise ratio has degraded, is one of the most common failure modes of rule-based fraud detection systems.
AI-powered detection addresses this through confidence scoring. Every anomaly is assigned a score based on how significantly the invoice deviates from the baseline, how many independent signals are present, and how consistent those signals are with known fraud patterns. High-confidence anomalies, where multiple independent signals converge, generate immediate alerts and payment holds. Low-confidence anomalies, where a single marginal deviation from the baseline is present, are logged for monitoring without disrupting the workflow.
AllAboutAI reports that modern AI fraud systems achieve false positive rates below 2%, compared to historic levels of 10 to 20% for rule-based systems. That reduction in false positives means the alerts that are generated carry genuine signal, and the team can act on them with confidence rather than dismissing them as noise.
The classic duplicate invoice has an identical reference number and identical amount. Those are caught easily. The sophisticated duplicate has a reference number that differs by one character and an amount that differs by a small percentage, just enough to make automated exact-match detection fail.
AI duplicate detection uses fuzzy matching across multiple dimensions simultaneously: supplier, approximate amount, date range, and reference number pattern. A near-duplicate that passes exact-match rules triggers an anomaly flag based on the probabilistic assessment that two invoices from the same supplier for similar amounts in the same period are more likely to be a duplicate than a coincidence.
The U.S. Department of Treasury's AI fraud detection deployment reported identifying 234 fraudulent invoices out of 12,547 processed, boosting detection accuracy from 77% to 99.7% while reducing false positives from 8% to 0.2%.
We covered the mechanism of payment diversion fraud in the vendor verification guide: a fraudster changes a supplier's bank details, and the next payment run sends the funds to an account they control. The timing of the change relative to the payment run is the most reliable signal.
AI payment fraud detection flags any bank detail change and applies automatic scrutiny to the next invoice from that supplier. It also looks for the timing pattern: bank detail changes that occur within a defined window before a payment run are treated with higher suspicion than changes that occur immediately after a payment, which are more consistent with a legitimate account update.
Deepfake invoices, generated using AI to produce convincing forgeries, are designed to pass visual inspection. They use the correct logo, the correct formatting, the correct VAT number. The document looks authentic.
What AI document forensics examines is not what the document looks like but what it is. Document metadata, font embedding patterns, pixel-level consistency analysis, and the correlation between the document's claimed generation date and its actual file properties all provide signals that visual review cannot access. An AI system that has been trained on document forensics catches structural inconsistencies that no human reviewer would identify by looking at the invoice.
Ghost vendor fraud, where a fictitious supplier is added to the approved vendor list for the purpose of submitting fraudulent invoices, leaves patterns that are visible at the vendor master level rather than the invoice level. New vendors added by specific users, approved unusually quickly, with invoices approved by the same user who added them, with bank accounts that share characteristics with other vendors in the system: these correlations are the signature of ghost vendor fraud.
AI analyses the full vendor master and the approval patterns across it, not just individual invoice-supplier relationships. The patterns that indicate a ghost vendor scheme are detectable at the system level even when each individual transaction appears legitimate in isolation.
AI fraud detection changes what humans need to do, not whether humans are needed. The cases that AI flags are the ones that require human judgement: is this a genuine anomaly or a legitimate business change? Is this timing coincidence or coordinated fraud? Is this a supplier error or deliberate manipulation?
Those decisions require context, relationship knowledge, and ethical judgement that AI does not have. The finance team's role in a well-designed fraud detection system is to investigate alerts, not to generate them.
The quality of an AI fraud alert is not just whether the anomaly was detected. It is whether the relevant context is surfaced alongside the flag. An alert that says "this invoice deviates from the supplier's historical baseline" is useful. An alert that says "this invoice is 340% above the supplier's 12-month average, arrives 2 days before the monthly payment run, and was submitted 1 week after the supplier's bank details were changed" is actionable.
Dost's AP automation platform surfaces anomalies with the context already assembled: the supplier's historical pattern, the specific deviation, the relevant recent events, and the risk score. The approver can make a decision without conducting their own investigation from scratch.
The practical value of AI fraud detection is only realised if the organisation has a clear response process for the alerts it generates. Who receives a high-confidence fraud alert? What authority do they have to hold a payment? What is the escalation path? What documentation is required before a payment that triggered an alert can proceed?
These are process design questions, not technology questions. The most sophisticated AI detection system generates limited value if the alerts it produces disappear into an inbox with no defined response protocol.
Dost's AP automation platform includes fraud detection logic built into the processing workflow. Every invoice is evaluated against the supplier's behavioural baseline as part of standard processing. Bank detail changes trigger an automatic hold and a verification step before any invoice from that supplier can proceed. Duplicate submissions are identified through fuzzy matching across reference, amount, supplier, and date.
The approval workflow enforces segregation of duties by design, preventing any single user from both adding a vendor and approving that vendor's invoices. The complete audit trail records every action on every supplier record, including every anomaly flag and every resolution decision.
This is the harder case for AI fraud detection, because the behavioural baseline requires historical data to be meaningful. For genuinely new suppliers, AI relies on cross-referencing against external data sources, checking document forensics, and applying the vendor master controls described in our vendor verification guide: bank account verification, company registration checks, and duplicate supplier detection. The behavioural baseline builds over time, which means a new supplier detected through document forensics today will be protected by behavioural monitoring from the second invoice onward.
The false positive rate for modern AI fraud detection is below 2%, which means the vast majority of flagged invoices are genuinely anomalous. When an investigation confirms that a flagged invoice is legitimate, the resolution is recorded, the context is documented, and the AI model updates its baseline for that supplier. If the anomaly reflects a genuine business change, such as a new contract at a higher value, the updated baseline will prevent the same scenario from triggering a flag in future months. The investigation time for a false positive is typically minutes rather than hours, because the context is already assembled by the system.
Unlike rule-based systems that perform identically regardless of how long they have been running, AI fraud detection models improve as the volume of processed transactions grows. The behavioural baselines become more accurate as more data is available to define what normal looks like. The model learns from resolution decisions: when a flagged invoice is confirmed as fraudulent, that pattern is reinforced. When a flagged invoice is confirmed as legitimate, the model adjusts its threshold for that type of deviation. After 12 months of operation, an AI fraud detection system is meaningfully more capable than it was on day one, without anyone having to update rules or retrain the model manually.
AI fraud detection does not compete with diligent manual review. It does something that diligent manual review cannot do at scale: maintain consistent, analytical scrutiny of every invoice, comparing it against the full supplier history, checking for document anomalies at a level of detail beyond visual inspection, and correlating patterns across the entire vendor master simultaneously.
The fraud that human teams miss is not the result of inattention. It is the result of what is structurally possible when a finance team is processing hundreds of invoices under time pressure, comparing each one against the purchase order and delivery note rather than against two years of behavioural data and a cross-system correlation analysis.
Closing that structural gap is what AI brings to AP fraud detection. And in a market where 71% of companies faced an increase in AI-powered fraud attempts in the last year, closing it is no longer optional.
