According to the Ardent Partners AP Metrics that Matter 2025 report, the average invoice processing time across all organisations stands at 9.2 days. Automation leaders process the same documents in under 3 days. The cost gap is even starker: $9.40 per invoice at the median versus $2.36 for best-in-class — four times cheaper, on the same document, in the same working day.
That gap is not about software. It is about automation maturity. And that maturity level has dramatically shifted with the arrival of LLMs (Large Language Models) in accounts payable solutions.
OCR promised to “kill manual data entry” for a decade. It kept that promise on standardised invoices from large, recurring suppliers. It failed on everything else — poorly formatted PDFs, unstructured confirmation emails, foreign vendor invoices, partial credit notes. The reason is simple: OCR reads pixels. LLMs read meaning.
This article is aimed at CFOs, finance managers, and controllers at mid-market companies (100–2,000 employees) running an ERP in production who want to move beyond OCR. It covers inbound accounts payable (AP) only — expense reports are out of scope.
The 4 AP Automation Maturity Levels
Level 1 — Manual Entry and Email Workflows
This is the starting point for most mid-market organisations. Invoices arrive by email or post. An accountant enters them into the ERP. An email goes out for approval. The manager responds — or not — in an ever-growing thread.
Processing cost exceeds $15 per invoice. Average cycle time: 15 to 20 days. Data entry error rate: 1–3%. This is not just a cost issue — it is a direct operational risk: one accountant on sick leave can leave 200 invoices in limbo.
Level 2 — Basic OCR
OCR extracts predefined fields from a PDF: invoice number, net/gross amount, VAT, IBAN, due date. It works reliably on standardised invoices from large, recurring suppliers (telecoms, utilities, maintenance). It fails on:
- image-only PDFs (a greyscale scan with no text layer)
- invoices with non-standard layouts
- confirmation emails with no structured table
- multi-line invoices with complex cost-centre allocations
In practice, basic OCR enables 30–50% straight-through processing but requires systematic human review for the rest. Manual work does not disappear — it migrates to the hard cases.
Level 3 — Adaptive Machine Learning
Adaptive ML goes beyond recognising predefined fields: it learns recurring patterns by supplier. After 20–30 invoices from the same vendor, the engine recognises its layout and standard cost allocations. Touchless processing rates reach 60–70% for a stable vendor panel.
The limitation: ML remains fragile on unseen cases. A new supplier, a project invoice with unusual cost allocation, a partial credit note — the system requests human validation. And since unseen cases often represent 30% of volume, ML’s ceiling becomes apparent quickly.
Level 4 — Generative AI (LLMs)
LLMs shift the equation on difficult cases. Where OCR and ML see characters and patterns, LLMs understand context:
- “Does this invoice cover a service already ordered in the ERP?”
- “Does the description ‘Q3 preventive maintenance’ map to which cost centre under our internal rules?”
- “Is the 12% variance between the purchase order and the invoice consistent with this supplier’s framework contract?”
In production, LLM-native solutions (SAP Document AI, Dynamics 365 Document Intelligence) and specialised tools embedding generative models achieve 80–85% touchless rates across the entire invoice flow — including irregular vendors and unstructured formats. Best-in-class organisations reach 85.7% touchless versus 36.2% for the average (Ardent Partners 2025).
Technical Architecture: Native ERP Module or Third-Party Solution?
Native ERP Module
The major ERP vendors have embedded their document AI capabilities directly into the platform:
- SAP S/4HANA — SAP Document AI: built on SAP BTP, it extracts invoice data and posts directly into the MM/FI workflow. The generative version (2025) handles PO-free invoices and image-based PDF formats.
- Microsoft Dynamics 365 Finance — Document Intelligence: built on Azure AI, it covers extraction and automatic matching against purchase orders within the same AP module interface.
- Sage Intacct — Native AP Automation: cloud-native for SME and mid-market, with an approval workflow configurable by amount and vendor category.
- Oracle Fusion Cloud — AI-Powered Invoice Processing: Oracle AI Applications module, with automatic extraction and PO matching within the P2P flow.
The advantage of going native: no critical interface to maintain, unified governance, all data in a single repository. The limitation: functional depth can be shallower than specialised solutions, particularly for complex cost-allocation rules and international vendor management.
Specialised Third-Party Solutions (Middleware)
Specialised solutions sit between invoice receipt and the ERP, enriching data before injection:
- Esker (international vendor, certified French PDP for e-invoicing): strong on the ETI/enterprise segment, deep P2P functional coverage, multi-ERP integration (SAP, Oracle, Dynamics, and others). Custom pricing, no public rate card.
- Tipalti: widely used in North America and Europe for high-volume AP automation, strong on multi-currency and multi-entity scenarios, pre-built connectors to major ERPs.
- Basware (ScanSource group): international solution for multi-site mid-market companies, strong on framework agreement management and 3-way matching (PO/goods receipt/invoice).
- Celonis: more a process orchestrator than an AP automation tool. Useful for identifying bottlenecks in your existing AP flows before deploying a capture solution.
Comparison: Native ERP vs. Third-Party Solution
| Criterion | Native ERP Module | Specialised Third Party |
|---|---|---|
| Critical interfaces | None | 1 (ERP connector) |
| AP functional depth | Moderate to good | Very good |
| Cost-allocation rules | Standard ERP config | Highly flexible |
| Implementation cost | Included in ERP licence | €20–50K depending on volume |
| Time-to-value | 6–12 weeks | 8–16 weeks |
| E-invoicing compliance | Depends on vendor roadmap | Varies by solution |
Simple decision rule: if your ERP is on a recent cloud version (SAP S/4HANA 2024+, Dynamics 365) and your volume is under 300 invoices per month with a stable vendor panel — the native module is sufficient. Beyond that, or if you have many international vendors and complex allocation rules, a third-party solution justifies the extra cost.
Business Rules to Configure for Automatic Coding
Rules by Vendor and Purchase Category
70–80% of your automation target rate is determined by business rule configuration, not technology. An LLM engine without well-structured rules will produce codings that accountants must correct systematically.
Priority rules to model:
- By key supplier: every recurring vendor (more than 12 invoices per year) should have documented coding rules — expense account, default cost centre, project analytics if applicable.
- By purchase category: IT services to account 611xxx, equipment maintenance to 615xxx, vehicle leasing to 6135xxx. A reference set of 30–50 rules typically covers 80% of recurring invoices.
- By amount threshold: invoices under $500 with no PO can be auto-approved if the supplier is known and active. Above that, N+1 approval is required.
Exception Handling: 3 Cases That Always Require Manual Review
Even at Level 4, three types of invoices remain in the human queue:
- Partial credit notes: a 30% credit on a multi-line invoice with cost allocation requires a human accounting decision, not an algorithmic rule.
- Invoices spanning multiple periods: an annual maintenance contract invoiced in November must be split between year N and N+1. The LLM can suggest the split, but an accountant validates.
- New suppliers: the first three invoices from any new vendor always go through human review, including IBAN verification — the classic fraud vector exploits precisely this onboarding moment.
Multi-Level Approval Workflow
A clear delegation schedule is the second automation lever. Without formalised delegations, every invoice escalates to the CFO. With a three-tier scheme:
- Under $1,000: auto-approved if supplier is known and coding rule exists
- $1,000–$10,000: department manager approval (target: 24h turnaround)
- Over $10,000: CFO and director co-approval (target: 48h turnaround)
Configuring these thresholds in your workflow tool — and enforcing actual compliance by approvers — drives overall cycle time more than extraction technology does.
Mandatory E-Invoicing and AP Automation
The wave of mandatory B2B e-invoicing across Europe is creating a concrete accelerator for AP automation. France’s mandate (effective September 2026), Germany’s (2025–2027 rollout), Belgium’s (since January 2026), and similar rules across the EU are pushing suppliers to issue invoices in structured formats: UBL, CII, or Factur-X/ZUGFeRD hybrid PDFs.
What Structured E-Invoicing Changes for AP
Under these mandates, invoices from registered suppliers arrive via certified exchange networks in a machine-readable structured format — not a free-form PDF. The data is clean at source, with normalised fields (number, amount, tax, supplier registration ID). No OCR extraction required.
For your AP automation solution, this is a significant advantage: error rates on these invoices drop to near zero at capture. If your solution connects directly to the relevant national exchange network or accredited service provider, structured invoices flow in without human intervention.
For a detailed look at France’s specific implementation, see our France mandatory e-invoicing guide.
What Structured E-Invoicing Does Not Cover
The mandates do not cover:
- invoices from foreign suppliers not subject to the local e-invoicing obligation
- small sole-trader invoices below registration thresholds
- legacy paper or free-form PDF still in circulation among some small vendors
These residual flows represent 20–35% of a typical mid-market company’s volume and still require OCR/LLM processing.
What to Do Now
If your AP solution already connects to the relevant certified network (France’s PDP ecosystem, Germany’s PEPPOL network, Belgium’s mandatory PEPPOL channel), you receive structured invoice flows automatically. If not, check your vendor’s compliance roadmap before year-end: structured invoicing is rapidly becoming the norm across Europe, and a solution that is not network-connected will create an increasing bottleneck as more of your suppliers switch formats.
ROI and Measurable KPIs
Data from the Ardent Partners AP Metrics that Matter 2025 study enables clear benchmarking:
| KPI | Before AP Automation | After (Level 3–4) |
|---|---|---|
| Average cost per invoice | $9.40 (market median) | $2.36 (best-in-class) |
| Average processing time | 9.2 days (average) | 2.9–3.1 days |
| Touchless processing rate | 36% (average) | 85% (best-in-class) |
| Data entry error rate | 1–3% | 0.1–0.3% |
| 3-year ROI | — | 3.2x on average |
A target of 60–80% touchless in 6 months is realistic for a mid-market company processing 200–500 invoices per month with a stable vendor panel. Reaching 80–85% typically takes 12–18 months of engine learning and business rule refinement.
Security: Preventing AP Automation from Becoming a Fraud Vector
Automation increases payment fraud risk if specific controls are not in place:
- Automatic IBAN verification: any IBAN change must trigger an alert and require dual human sign-off, even for suppliers you have worked with for five years.
- Dual validation for new suppliers: the first three invoices from any new vendor always require human review regardless of amount.
- Mandatory exception log: any rule override — manual validation of an invoice normally processed automatically — must be logged, timestamped, and available as an audit trail.
For a detailed analysis of vendor fraud risks in your ERP, read our AML, UBO and KYC compliance guide.
6-Month Implementation Roadmap
Months 1–2: Flow Audit and Solution Selection
Before selecting a solution, document:
- Monthly invoice volume by intake channel (email, EDI, supplier portal, post)
- Your top 50 suppliers by volume — they typically represent 70–80% of total flow
- Amount distribution: what proportion falls under $500? Under $1,000?
- Existing coding rules: are they documented or only in the senior accountant’s head?
This mapping determines native ERP vs. third-party and the complexity of initial configuration.
Months 3–4: Configuration and Pilot on 20 Priority Suppliers
Start with the 20 suppliers representing the highest volume of recurring invoices. Configure coding rules for each. Run the engine in “suggestion” mode (the tool proposes, the accountant validates) for four weeks. Adjust rules based on observed errors.
This run-in phase generates the learning data that ML/LLM engines need and allows rule correction before full rollout. It cannot be shortened.
Months 5–6: Phased Rollout and KPI Measurement
Expand to the full scope in waves of 30 additional suppliers per week. Measure weekly:
- Touchless processing rate (target: over 60% by end of month 6)
- Invoices awaiting validation for more than 48 hours (target: fewer than 20 at any time)
- Average processing time (target: under 5 days by month 6)
AP Automation Pre-Go-Live Checklist (8 Points)
- Invoice volume analysed by channel and by supplier
- Coding rules documented and validated by the CFO for the top 50 suppliers
- Tests run on the 100 most frequent suppliers in suggestion mode
- Exception workflow configured: credit notes, multi-period invoices, new suppliers
- IBAN control activated: any change triggers an alert and human validation
- Finance team trained on the new workflow and exception cases
- Tracking KPIs defined with a real-time dashboard accessible to the CFO
- Rollback plan documented: if error rate exceeds 5%, what is the manual fallback circuit?
To complement this topic, read our treasury management and cash forecasting ERP guide — AP automation accelerates visibility into short-term cash outflows, a key lever for cash forecasting. For organisations operating in France that want to understand the full regulatory picture on supplier invoicing, our article on France mandatory e-invoicing 2026 covers the obligations now in force.
