TL;DR
MIT’s Project NANDA study (July 2025) found that 95% of enterprise generative AI pilots deliver zero measurable P&L impact. In Accounts Payable specifically, the failure pattern is unusually consistent: GenAI handles the easy invoices well, lifts touchless rate from a baseline of 40–50% to about 60–70%, and then plateaus. The remaining 30–40% of invoices look the same on the surface, but they hide seven specific failure modes that probabilistic AI cannot reason through without a deterministic layer underneath it.
The seven places generative AI quietly fails in AP:
- Vendor master ambiguity. Duplicate vendors, name variations, and entity hierarchies that the AI cannot disambiguate.
- Contract escalation drift. Pricing or terms that changed in the underlying contract but were never updated in the PO.
- Goods receipt timing windows. Invoices that arrive before, after, or partially overlapping the GR event.
- Non-PO invoice coding. Chart-of-accounts decisions where GenAI confidently picks the wrong account.
- Tax and FX edge cases. Multi-jurisdiction VAT, withholding tax, and currency conversion timing.
- Exception escalation that creates more work. Human-in-the-loop becoming an unmanaged review queue rather than a control.
- Audit trail invisibility. “Decision: APPROVED. Confidence: 94%” is not an audit trail.
Each of these failure modes has a specific diagnostic you can run during a pilot to catch it before procurement signs the contract. This post walks through all seven, with the questions to ask, the test transactions to run, and the architectural answer that distinguishes deterministic, audit-ready AI Accounts Payable automation from rebranded probabilistic AI.
Why AP pilots plateau at 60–70% touchless
The number that haunts every AP leader is the touchless rate. Three-way match took it from zero to roughly 60–70% over the last decade. Then it stopped. Most AP teams have invested in OCR, workflow upgrades, ERP refreshes, and finally generative AI, expecting each one to crack the next 30%.
Most don’t. The reason is consistent across organizations: the remaining 30–40% of invoices are not “harder invoices.” They are invoices whose context lives outside the invoice itself. A duplicate vendor master entry from 2022. A pricing escalation clause that kicked in last month before anyone updated the PO. A goods receipt logged in the wrong week because a clerk was rushing before quarter close. A French subsidiary’s VAT treatment that the chart-of-accounts mapping doesn’t cover.
Rules-based three-way match has no way to reason about any of these. Probabilistic generative AI can sometimes reason about them, but cannot reliably tell you when its reasoning is wrong. The result is the worst of both worlds: an automation that handles 65% of invoices cleanly, breaks on 35%, and lacks the audit trail to explain which decisions it made and why.
The seven failure modes below are where this pattern shows up most consistently in pilot data. Each one is a place where vendor demos look great, and production reality looks different.
The 7 failure modes
1. Vendor master ambiguity
What happens. Your ERP has 18,000 vendor records. Approximately 800 of them are duplicates, near-duplicates, or stale entries from acquisitions. “Acme Corp” exists three times: as “Acme Corp”, “Acme Corporation Inc”, and “ACME Corp LLC” (the last one created in 2023 after a re-incorporation that nobody told AP about). An invoice from Acme arrives. The vendor name on the invoice matches none of them exactly. GenAI confidently maps it to whichever record’s text is closest to the OCR output. About 40% of the time, that’s the wrong record.
Why GenAI fails. Language models are good at finding semantic similarity. They are not good at understanding that two records that look 80% similar might be:
- The same vendor at different addresses (legitimate to merge)
- A parent entity and a subsidiary (often need to remain separate)
- A vendor and a one-time supplier with a similar name (must not be merged)
- A stale record that should be retired (but the historical PO references still need to resolve)
Without explicit business logic about your vendor hierarchy, the model just picks the closest match. Sometimes it’s right. Sometimes the payment goes to a 2019 banking detail for a vendor that was acquired in 2022.
Pilot diagnostic.
- Pull the 50 vendor records in your ERP with the most near-duplicates. Run 5 invoices per record through the GenAI tool. Track which records it selected and whether they match the AP team’s manual judgment.
- Specifically test acquired or re-incorporated entities, vendors with multiple billing addresses, and vendors with similar names (e.g., “United Healthcare” vs “United Health Group”).
What good looks like. A platform that can express the vendor-matching rules in plain English (“when the vendor name on the invoice resolves to multiple ERP records, route to AP supervisor unless the invoice references a PO whose vendor record is unambiguous”), execute that rule deterministically, and log which record was selected and why.
2. Contract escalation drift
What happens. Your three-year managed services contract with the data center provider includes a 4% annual price escalation clause and a quarterly true-up for power usage. The PO was issued at the original rate. The invoice arrives at the escalated rate plus a power adjustment. Three-way match fails because the invoice doesn’t equal the PO. GenAI tries to “interpret” the variance, often by approving it because “this is the kind of variance that’s usually approved.”
Why GenAI fails. The contract that justifies the variance is not in the AI’s context window. It’s in a contract management system (or, more often, a SharePoint folder). Without explicit retrieval of the underlying contract terms and a deterministic rule for applying them, the AI is guessing at what the variance means.
When it guesses right, nobody notices. When it guesses wrong (approving a variance that wasn’t actually justified by the contract), the error compounds: it sets a precedent the model “learns” from on future invoices.
Pilot diagnostic.
- Identify your top 20 contracts with escalation clauses, volume discounts, or true-up provisions. Pull recent invoices from each.
- Ask the GenAI tool to handle the variances. Then ask it to show you which specific contract clause justified each approval.
- Compare against AP team manual judgment, and especially against what the contract actually says.
What good looks like. An automation that can retrieve the specific contract clause at decision time, apply it to the invoice deterministically, and log the citation alongside the decision. “Approved per Section 4.2 of MSA-2024-127, which permits annual 4% escalation effective January 1” is an audit trail. “Approved with 91% confidence” is not.
3. Goods receipt timing windows
What happens. Your warehouse logged the GR on March 31 to hit a quarter-end target. The actual receipt happened April 2. The invoice arrives April 10. The three-way match works (PO, invoice, GR all align), but the GR is recorded in the wrong period, which means the invoice was incurred in Q2, not Q1. Or: the invoice arrives before the GR. Or: the invoice partially matches a GR for a multi-line PO where only some lines have been received.
GenAI looks at the documents and sees that the totals match. It approves. The expense gets booked to the wrong period.
Why GenAI fails. Period-correct accounting requires reasoning about time, not just amounts. Did the goods receipt happen in the period claimed? Are we close to a cutoff? Does this transaction need an accrual? The AI sees only the documents in front of it, not the period-end policies, the cutoff date, or the materiality threshold for accruals.
Pilot diagnostic.
- Pull 100 invoices from your last quarter-end. Identify the ones where the GR was recorded within 5 business days of period close.
- Run them through the GenAI tool and ask: was this transaction recorded in the correct period? Does this transaction need an accrual?
- Specifically test multi-line POs where partial GRs have been booked.
What good looks like. An automation that knows your fiscal calendar, your cutoff date, and your accrual materiality threshold. It treats period-end transactions differently from mid-period ones, and it can explain why a specific transaction was or was not flagged for accrual review.
4. Non-PO invoice coding
What happens. Roughly 30–40% of invoices in most enterprises are non-PO (utilities, professional services, one-time purchases, employee reimbursements that came in as vendor invoices). These have no PO to match against, which means three-way match doesn’t apply. The AI has to decide the GL coding based on the invoice contents.
GenAI is reasonably good at this for common cases (electricity bill goes to utilities expense). It is dangerously confident on edge cases. A consulting firm’s invoice for “Q1 advisory services” might go to professional services, but if it relates to a capital project, it should be capitalized. The AI doesn’t know about the capital project. It picks the first reasonable answer with high confidence.
Why GenAI fails. Chart-of-accounts decisions are not document-classification problems. They require knowledge of organizational context (which projects are in flight, which budget owners approve which expense types, which transactions get capitalized vs expensed) that lives outside the invoice. GenAI fills the gap with confident-sounding guesses.
Pilot diagnostic.
- Pull 200 non-PO invoices coded by your AP team over the last six months.
- Run them through the GenAI tool and compare coding decisions side-by-side with the team’s actual coding.
- Pay specific attention to: professional services invoices (capex vs opex decisions), facilities-related invoices (capital improvement vs maintenance), and any invoice with a “project” reference.
What good looks like. An automation that knows the rules your AP team applies in their head (capitalized if it relates to a project on the active capex list, expensed otherwise; routes to budget owner X if the amount is over $Y), executes those rules deterministically, and asks for human judgment when the rules are ambiguous rather than guessing confidently.
5. Tax and FX edge cases
What happens. An invoice arrives from a French vendor in Euros. The amount includes French VAT. Your entity is a US LLC, but the goods or services were delivered to a UK subsidiary that’s VAT-registered there. The correct treatment involves: converting Euros to USD at the appropriate date’s FX rate (invoice date? service date? payment date?), determining whether the VAT is recoverable (and by whom), and handling any withholding tax obligations.
GenAI is famously bad at this. The reason is that “correct” depends on jurisdiction-specific rules, your specific entity structure, and timing details that aren’t on the invoice.
Why GenAI fails. Tax and FX are deterministic by their nature. The correct answer is not a probability distribution; it is the answer that satisfies the specific rule applicable to this specific transaction. GenAI’s probabilistic reasoning is fundamentally mismatched to a deterministic problem.
Pilot diagnostic.
- Identify your highest-volume cross-border invoice flows (vendor country to subsidiary country pairs).
- Pull 30 recent invoices from each. Ask the GenAI tool to handle them end-to-end: VAT treatment, FX conversion date, withholding tax assessment, GL coding.
- Compare against your tax team’s actual treatment. Specifically test reverse-charge VAT scenarios, intercompany transactions, and any invoice that requires a permanent establishment analysis.
What good looks like. An automation that encodes your tax position as English-language rules (“French vendor invoicing UK subsidiary: apply reverse-charge VAT; convert at invoice-date ECB rate; route any invoice over EUR 50K to tax team for review”), executes them deterministically, and produces the documentation a tax auditor would expect.
6. Exception escalation that creates more work
What happens. The GenAI tool encounters an invoice it can’t handle confidently. It routes the invoice to a human reviewer. The reviewer opens it, looks at it, and realizes they can’t tell why the AI escalated it. The AI says “low confidence.” It does not say “this invoice matches PO 4521 in total but the line-item description for line 3 says ‘consulting’ while the PO line 3 says ‘software license’; the variance is $4,200; the vendor has 47 prior invoices with this PO.” The reviewer has to recreate that analysis themselves.
Multiply this by 500 escalations a week. The “human in the loop” becomes a triage queue. The team that GenAI was supposed to free up is now bottlenecked on AI-generated work.
Why GenAI fails. Probabilistic systems escalate based on confidence scores. Confidence scores tell you nothing about why the system was uncertain. Without a structured explanation of the exception, every human review starts from scratch.
Pilot diagnostic.
- During the pilot, track three numbers: invoices escalated to humans, time-per-escalation, and rework rate (escalations that come back from humans because the human escalated them again).
- Specifically watch for review-queue burnout: if your AP team is spending more time triaging GenAI escalations than they spent on the original manual process, the AI is creating work, not eliminating it.
What good looks like. An automation whose escalations include a plain-English explanation of what went wrong, the specific fields that triggered the exception, and the most likely resolution paths. “Vendor on invoice resolves to two ERP records (Acme Corp #4521 vs Acme Corp LLC #8830); the invoice references PO 7724 which is associated with #4521; recommend matching to #4521 unless AP supervisor indicates otherwise” is a useful escalation. “Confidence: 71%” is not.
7. Audit trail invisibility
What happens. Your auditor sits down in Q3. They pick a specific invoice processed by your GenAI tool. They ask: walk me through how this decision was made. You open the platform’s audit log. It shows: “Invoice 482919: AI processed. Decision: APPROVED. Confidence: 0.94. Action: posted to GL 6100.”
Your auditor asks the second question. Which specific rule did the AI apply? You don’t have an answer.
In a 2026 audit environment shaped by COSO’s February 2026 generative AI guidance, the SEC’s March 2026 dedicated SOX enforcement group, and the PCAOB’s amended AS 2201 effective December 15, 2026, this is no longer a documentation gap. It is a material weakness. For the deeper auditor playbook, see what your SOX auditor will ask about your AI automation.
Why GenAI fails. Most GenAI tools are built on probabilistic models whose “reasoning” is an emergent property of model weights, not an explicit rule that can be cited. The audit trail captures the inputs and outputs, but cannot reconstruct the decision path in a way an auditor can verify.
Pilot diagnostic.
- Pick 20 invoices that the GenAI tool processed during the pilot, ideally a mix of straightforward and edge cases.
- Ask the vendor to produce, for each one: the timestamp, the inputs received, the specific rule or policy applied, the reasoning expressed in plain language, the action taken, and the user (if any) who reviewed it.
- Then ask: how do we prove this log has not been altered since it was written?
What good looks like. Every decision logged with the 12-field minimum schema we covered in our 2026 AI audit trail checklist: NTP-synced timestamp, decision ID, authenticated user, AI system version, model version, inputs with source attribution, the specific rule or policy invoked, reasoning in plain English, the output produced, the downstream action, human review if applicable, and tamper-evident integrity proof. Anything less than this is going to be a finding in your next audit cycle.
What separates pilots that succeed from pilots that stall
Across the seven failure modes, the same architectural distinction shows up: pilots that succeed have a deterministic layer underneath the AI. Pilots that stall do not.
Deterministic doesn’t mean “no AI.” It means the AI’s reasoning is grounded in explicit, inspectable rules expressed in human language. When the AI handles an invoice, it applies a specific policy you can read. When it escalates, it explains which part of the policy was ambiguous. When it makes a decision, it logs the rule that drove the decision. When your auditor asks why, the answer is the policy, not the confidence score.
This is the difference between agentic AI as a productivity tool and agentic AI as a control. AP is one of the most control-intensive functions in the enterprise. It deserves an AI architecture built for it. For the procurement-side artifact that documents this architecture, see our piece on the AI Bill of Materials (AIBOM).
How Kognitos handles the seven failure modes
Kognitos is a neurosymbolic AI platform built on a deterministic English-as-code foundation. Each of the seven failure modes above maps to a specific capability:
- Vendor master ambiguity. Vendor-matching rules expressed in plain English, executed deterministically, with explicit handling for the disambiguation cases.
- Contract escalation drift. Contract terms retrievable at decision time, with the specific clause cited in the audit log.
- Goods receipt timing. Fiscal-calendar-aware processing with period-end rules expressed explicitly.
- Non-PO invoice coding. Coding rules that encode your organization’s logic (capex vs opex, project mapping, budget owner routing) rather than guessing from the invoice text.
- Tax and FX edge cases. Tax position encoded as English rules per vendor-entity pair, with FX rate sources and dates specified.
- Exception escalation. Plain-English explanations of what triggered the escalation, what the system tried, and what options exist for resolution.
- Audit trail. Every decision logged with the 12-field minimum schema, with tamper-evident integrity proofs and direct mappability to SOX, COSO, and EU AI Act documentation requirements. For our security posture and compliance attestations, see the Kognitos Trust & Security portal.
If you are planning an AP pilot in 2026 and want to see what the deterministic alternative looks like on the seven failure modes above, we’d be glad to walk through a working example on your actual invoice flow.
Book a working session with a Kognitos solutions engineer → Or register for our May 20 webinar: Beyond 3-Way Match →
Last updated: May 2026. This article is intended for informational purposes and does not constitute legal, audit, accounting, or tax advice. Specific requirements vary by jurisdiction, industry, and the structure of your AP program. Engage qualified counsel and your audit, tax, and procurement teams for guidance specific to your situation.