How to test AI invoice matching for landscape material purchases

How we work: This guide uses current public documentation and does not claim hands-on testing unless stated. Some links may later become affiliate links, at no added cost to you. Rankings are independent of commissions.

Test AI invoice matching with synthetic purchase orders, delivery records, invoices, and duplicates before it touches a live vendor bill. The system should extract the document, compare each material line with what was approved and received, stop risky exceptions, preserve corrections, and export one reviewed transaction to accounting. A convincing OCR preview is not proof that those controls work.

Download the landscape invoice OCR test-case manifest. It contains fictional vendors, invoice numbers, amounts, and dates. Use those values to create documents only in a vendor sandbox or other approved test environment.

This is a desk-researched test protocol, not a report that we tested QuickBooks, Ramp, Dext, Microsoft Dynamics 365, or another account. The official product, NIST, and National Association of Landscape Professionals sources were opened and checked on September 16, 2026.

What does invoice matching need to prove?

Separate four jobs that software marketing often compresses into one phrase:

1. Ingest: accept a supported file or email without losing its sender, attachment, or entity.
2. Extract: turn the visible vendor, invoice, and line-item text into structured fields.
3. Match: compare those fields with approved and received records.
4. Post: create one controlled bill or expense in the accounting system after required review.

Microsoft’s current accounts-payable documentation gives the useful formal distinction. Invoice-total matching compares overall totals. Two-way matching compares invoice prices with a purchase order. Three-way matching also compares invoice quantities with product-receipt quantities. Configured tolerances decide which differences become discrepancies.

For landscape materials, three-way matching is the strongest starting point when the business orders physical goods. The purchase order says what was approved. The delivery or goods-receipt record says what arrived. The invoice says what the supplier wants paid.

Do not label all four methods “AI matching.” Record the evidence the tool actually compares.

Which fields should a landscape material test include?

Use stable identifiers first, then descriptions:

Units deserve their own check. Ten bags, ten cubic yards, and ten pallets are not interchangeable because the number happens to match. The same applies to plant container sizes, tons versus yards, and each versus case.

The landscaping job-costing guide shows why job and cost-code allocation belongs in this test. An invoice can balance in accounts payable while sending irrigation parts to a maintenance contract or mulch to the wrong property.

What can current OCR products actually extract?

Capabilities differ, so use each vendor’s current documentation rather than a generic “AI-powered” label.

QuickBooks currently documents an email intake that extracts dates, amounts, vendor, and the last four card digits from a receipt or bill, then places the document in a review flow where a user can edit, add, or match it. Its email guide requires one receipt per image and identifies supported file types and permissions. That is an extraction-and-review workflow; the page does not establish line-level three-way matching.

Dext’s line-item documentation, updated August 4, 2026, says it extracts each line’s description, total amount, tax amount, and quantity. It explicitly says tax codes and rates are not extracted, and blurry fields may be left blank. Those limits make tax-code preservation and unreadable-field handling required test cases.

Ramp documents another shape. Its purchase-order workflow can use OCR to read a PO number and suggest an imported PO, then match bill lines to PO lines. Its AP email workflow says OCR may prefill invoice number, due date, total, line items, vendor details, and payment details. Exact duplicate files are filtered by file hash; a changed file may instead be flagged when vendor name and invoice number match an existing bill.

These are documented capabilities, not test results. Confirm the plan, accounting integration, entity behavior, and approval configuration in the account being evaluated.

How should you build the synthetic test pack?

Start from the downloadable manifest. Create no real vendor, employee, customer, property, bank, or card data.

Use one clean baseline set:

approved PO → full receipt → matching invoice

Then change one condition at a time:

• invoice only part of a fully received order;
• receive only part of an order, then invoice that received amount;
• invoice more units than were received;
• change one unit rate;
• change tax without changing the pre-tax lines;
• add freight that was not on the PO;
• resend the identical file;
• rescan the same logical invoice into a different file;
• omit the receipt;
• omit the PO;
• blur a critical field; and
• send a credit memo through the same inbox.

Keep the clean source values in the manifest. The reviewer should not rely on the OCR output to decide what the OCR output was supposed to say.

For a strict pilot, configure zero variance where the product allows it. That does not mean the company’s permanent business tolerance must be zero. It makes every deliberate difference visible while the team is learning which rules and fields the tool evaluates.

How do you test exact and partial matches?

The baseline must preserve every line, not merely the invoice total. Check:

• vendor and entity;
• invoice and PO numbers;
• item description or approved mapping;
• quantity and unit;
• unit rate and line extension;
• subtotal, tax, freight, credit, and total;
• job, property, and cost code;
• receipt quantity; and
• currency and dates.

For the partial case, receive fewer units than ordered and invoice exactly the received amount. Pass only if the system links the correct lines, leaves the unreceived balance open when that is the intended configuration, and does not mark the whole PO complete.

Then split one order across two approved synthetic invoices. The second invoice must see what the first already consumed. A tool that compares each invoice with the original PO independently can allow cumulative overbilling.

How do you test duplicates and unreadable files?

Run two duplicate cases:

1. submit the same bytes under a different filename; and
2. print or rescan the same logical invoice so the file bytes change while vendor, invoice number, and amount remain the same.

Both must stop or receive a prominent review flag before accounting export or payment. Dext’s current duplicate guide shows why the second case matters: its invoice rule depends on supplier, total, and document reference, and missing key fields can prevent recognition. The guide also distinguishes automatic deletion from review mode and restricts suspected-duplicate review to authorized users.

For the unreadable case, obscure one critical field at a time. A blank invoice number must not become a made-up value. A low-confidence quantity must not default to the ordered amount. A missing tax value must not silently inherit another vendor’s rule without a visible decision.

The pass condition is an owned exception with the original file attached. “No draft created” is not a pass unless the sender or AP owner receives a traceable rejection and can recover the document.

Which discrepancies must stop automatic posting?

Use the company’s approved materiality policy for ordinary variances, but treat these test conditions as hard stops:

• probable duplicate;
• inactive or unapproved vendor;
• missing or ambiguous invoice identity;
• no approved PO where a PO is required;
• no receipt for physical goods;
• invoiced quantity above the received quantity;
• unexplained rate, tax, freight, or total difference;
• wrong legal entity, job, property, or cost code;
• unreadable critical field; and
• a credit memo classified as a payable invoice.

An exception needs a reason, owner, and permitted action. “Approve anyway” without a retained explanation defeats the test.

Landscape trade reporting supports testing the operating process, not just the model. In June 2026, NALP described one company standardizing material, tool, and consumable requests through a mobile process with photos and descriptions, centralized purchasing, tracking, and distribution. If the ordering and receiving records are inconsistent, OCR will automate inconsistent inputs.

How should permissions and corrections be tested?

Use at least four synthetic roles:

Attempt every restricted action, including by direct link and after a role change. Record denied actions as evidence.

For each correction, capture the original extracted value, corrected value, person, time, reason, and resulting match state. Reprocess the same vendor format later and confirm that a learned rule does not “correct” a different invoice incorrectly.

The landscaping tech-stack audit template can assign the source of truth and owner for purchasing, receiving, payables, and job-cost fields.

How do you verify accounting export and payment handoff?

Use a sandbox or non-posting test company when the accounting product supports one. Export only the accepted baseline and approved partial case.

Confirm:

1. one source invoice creates one accounting record;
2. vendor, date, number, amount, tax, terms, and entity agree;
3. every line reaches the intended account, job, property, and cost code;
4. the original image and PO or receipt references remain retrievable;
5. a corrected draft does not leave an orphan first version;
6. a retry does not create a second bill; and
7. the AP system and accounting system expose stable IDs for reconciliation.

Do not test payment with real bank details. The payment payout reconciliation worksheet starts after a valid payable or customer payment exists; it does not repair a bad source invoice.

What are defensible acceptance gates?

NIST’s AI RMF says test sets, metrics, tools, and methods should be documented, performance should be evaluated under conditions similar to deployment, and human oversight should be defined and assessed. Apply that directly:

Report results by scenario, field, vendor format, and image condition. Do not publish one blended “accuracy” percentage that hides a failed duplicate or missing-receipt case.

The tool passes when it reduces entry work without turning uncertain documents into certain accounting records. Keep hard stops in place, review drift with new vendor layouts, and rerun this pack after changes to extraction, matching, permissions, accounting integration, or approval policy.