BizBlocz
Finance Three-Way Matching AP03
Operations reference

Finance: Three-Way Matching

You own this process. What the work is and where its difficulty sits — then how much better it could run, who can run it, where AI fits, and how to choose.

The short answer

8-step finance work whose binding step is probabilistic matching — the part you can’t fully automate away. Best-fit AI is ML / Predictive (~35%); best-in-class teams reach 65–85% improvement potential.

Tasks
8
The bottleneck
probabilistic matching
Improvement potential
65–85% · Improvement potential
Best-fit AI
ML / Predictive · 35%
01
Section 01 / 05
Overview · understand the work

What the work actually is

Three-way matching reconciles the invoice against the purchase order and the goods receipt — at header and line level, within tolerance — releasing the clean matches to payment and routing the mismatches to a buyer or AP analyst.

Inputs · documents in
Purchase orderGoods receipt / service entry sheetSupplier invoiceTolerance rules
Outputs · documents out
Matched invoice released to payGR/IR clearing entryVariance exception ticket
Volume
high
Risk / control
moderate
Shape of the work
Mostly rule based · gated by predicting

The 8 tasks — the nature of each, and the oversight it needs

Tag each task in plain terms — what kind of work it is and how hands-off it can run — before any mention of AI. The kind of work is what later decides which tool, if any, fits.

Naturerule-basedreadingpredictingjudgingpeople / hands-on
01
Retrieve purchase order data and confirm PO validity and open balances
retrieving PO dataunattended
02
Retrieve goods receipt / service confirmation (GRN/SES)
retrieving goods receiptunattended
03
Match invoice to PO and GRN at header and line level (quantity, price, delivery)the bottleneck
probabilistic matchingexceptions
04
Apply configurable tolerance rules (price variance, quantity variance)
applying tolerancesunattended
05
Flag and route exceptions for buyer or AP resolution
routing exceptionsexceptions
06
Handle partial receipts, over-deliveries, and blanket PO scenarios
resolving complex matchesperson decides
07
Update PO and GRN balance trackers after successful match
updating balancesunattended
08
Release matched invoices to payment workflow
releasing to paymentunattended

Most steps (retrieve, apply tolerances, update, release) are rule-based. But the binding step is the match itself: comparing quantities, prices, and delivery across three sources within tolerance is probabilistic — a classification-and-anomaly problem, not a fixed rule. Get the match right and the rest releases automatically; the edge cases — partial receipts, over-deliveries, blanket POs — are where human judgment stays.

Matching errors flow downstream into payment; tolerance and exception controls keep accuracy and a human on the edge cases.

02
Section 02 / 05
Improvement potential · how much better it could run

How much better this process can run

The question isn’t only “is there savings” — it’s can I run this better: cheaper, faster, higher quality, better service? Here’s what best-in-class looks like, and how teams get there. (How much of it AI specifically drives — and how proven that is — is Section 04.)

Best-in-class · what “better” looks like
65–85%
Improvement potential
Gartner
85–95%
Straight-through match rate
Everest
80%
Invoice/payment handling cost ↓
Basware
97%
RPA ROI (3-yr)
UiPath / Forrester
How best-in-class teams get there

Process discipline first, then automation — AI is one slice of the second column, not the whole answer.

Process & standardization
  • PO-compliance discipline
  • Tolerance-rule design
  • Goods-receipt (GRN) discipline
  • Blanket-PO policy
Automation & AI
  • ML matching engine
  • Exception prediction & routing
  • Auto-release of clean matches
  • RPA PO/GRN retrieval
Best-in-class teams reach 65-85% improvement (Gartner), with ML matching engines hitting 85-95% straight-through rates (Everest) — among the most automatable AP steps.
03
Section 03 / 05
Executor · who can run it

Your levers — five ways to run this work

“Who runs the work” is its own question, separate from AI. AI shows up across these options — sometimes heavily, sometimes not at all. Vendor-neutral; the real options mapped to AP03.

Lever 01
Internal staff
Your own team runs it — the status quo.
AI: optional copilotdata: in-house
Your people, on your ERP, optionally AI-assisted.
Best when volume is low, formats vary wildly, or you need full control and a person accountable on every step.
Lever 02
ERP / platform
Your system of record runs it natively.
AI: some nativedata: platform-resident
Oracle
Best when you're already on SAP/Oracle and want least integration — data never leaves the ERP.
Lever 03
Specialized SaaS
Buy a best-of-breed product; run it in-house.
AI: usually coredata: vendor-cloud / customer-cloud
HighRadius · Esker · Basware · Stampli · Medius · UiPath · Tipalti · BILL · AvidXchange
Best when you want capability your ERP lacks and will run another system; data processed in the vendor cloud.
Lever 04
AI agents
Autonomous AI runs the pipeline; you handle exceptions.
AI: it IS the executorcross-cuts the delivery models
HighRadius · Basware · UiPath · Tipalti · BILL · AvidXchange · Oracle · WNS
Best when volume is high and formats are stable — you want touchless and only manage exceptions.
Lever 05
BPO / managed service
Hand the whole process to a partner.
AI: people + toolingdata: service-mediated
EXL Service · WNS
Best when you want an outcome and an SLA, not a tool to operate — partner works on your ERP, data stays with you.
Note on AI agents: they aren’t bought separately — you get them through a delivery model (your ERP, a SaaS product, or the BPO). Listed on their own because “should an agent run this autonomously?” is a distinct decision (Section 05), not because it’s a separate kind of vendor.
04
Section 04 / 05
AI · where it fits this work

Match a solution to each kind of work

Recall the tasks and their nature from Section 01. AI is one lever, not the whole story — the mix below is simply the result of matching the right kind of solution to each kind of work, weighted by where the work concentrates.

Nature of the work → the solution that fits
Read a document you didn’t designDocument AI
Deterministic routing, validation, postingAgentic / RPA
Anomaly detection & predictionML / Predictive
Draft, summarize, correspondGenerative AI
Answer questions in natural languageNLP / Conversational
See / digitize images & scansComputer Vision
The AI mix · weighted by where the work concentrates
35%
ML / Predictive leads the mix — matched to where this work concentrates and to its binding step.
ML 35%
Agentic 30%
Document 25%
ML / Predictive35%
Agentic AI / RPA30%
Document AI25%
Generative AI5%
NLP / Conversational5%

ML/Predictive leads (~35%) because matching PO vs. receipt vs. invoice within tolerance is probabilistic fuzzy matching — classification and anomaly detection. Agentic/RPA (~30%) runs retrieval and release; Document AI (~25%) reads the source documents. (Everest Group IDP PEAK Matrix 2024, Gartner Finance 2024.)

AI target value
65–85% — AI the dominant lever toward Section 02’s targets
AI’s contribution toward the best-in-class targets · personalized in the assessment
Strong
evidence
The grade is for the AI value/results, not the mix (which is directional). AI target value: ~65-85% (Gartner) — among the most automatable AP steps, with ML the dominant lever. Strong evidence. Sources: Gartner Finance 2024, Everest Group IDP PEAK Matrix 2024, UiPath / Forrester TEI, Basware.See your number →
05
Section 05 / 05
How to choose · which lever fits you

Matching the approach to your situation

The right lever fits your volume, variability, control needs, and appetite to operate a system. Start here.

If your situation is…
Lean toward
High, stable volume; you want touchless
AI agentvia your ERP or a SaaS platform — runs itself, you handle exceptions
Formats vary widely, exceptions frequent, or a person must stay accountable
Copilotyour team, AI-assisted — the human still presses enter
Already standardized on SAP/Oracle; data must stay in the ERP
ERP-embeddedleast integration, platform-resident data
Need capability your ERP lacks; willing to run another system
Specialized SaaSbest-of-breed; data processed in vendor cloud
You want an outcome & SLA, not a tool to operate
BPO / managed serviceoffload the function; partner works on your ERP

The autonomy question: agent or copilot?

Whichever delivery model you pick, one choice cuts across them — who presses enter.

It acts

AI agent

Runs the steps end-to-end, completes the clean cases on its own, and routes only the exceptions to a person.

Best: high volume, stable inputs, a clear accountability surface.
vs
It assists

AI copilot

Sits beside the person and speeds up each step; the human acts on every decision.

Best: high variability, frequent exceptions, or a need for a person in the loop.

What to evaluate — whichever you choose

  • Accuracy on your own inputsvendor benchmarks are on clean data; test your messiest cases.
  • Straight-through / touchless ratethe real efficiency number, not “AI-powered.”
  • Exception-handling experiencemost of your team's time goes here, not the happy path.
  • ERP write-back & integration depthdoes it post cleanly to your system of record?
  • Data residencydoes data leave your environment, and is that acceptable to compliance?
  • The accountability surfacewhat happens, and who owns it, when the model is confidently wrong?
Related blocks

See every lever across your processes

Run your portfolio through the assessment — work profile, improvement potential, confidence, and executor options across all your blocks, scored against 127 enterprise subprocesses.

Open the AI Value Assessment →