# Supplier invoices matched to POs before anyone retypes them

> An invoice processing pipeline reads supplier invoices from an email inbox, extracts the vendor, totals, tax and line items, checks them against your purchase orders and vendor records, sends anything that fails a check to a reviewer and posts approved bills to QuickBooks Online, Xero, NetSuite or Sage Intacct. The Method Crew builds one in 6 to 9 weeks for $15,000 to $35,000.

- Page: https://methodcrew.com/blueprints/invoice-processing-pipeline
- Last updated: 2026-10-28
- Publisher: [The Method Crew](https://methodcrew.com/)
- Blueprint: Invoice processing pipeline

Invoices arrive in a shared inbox and leave as checked, approved bills in your accounting system. People only see the ones that fail a check, with the PDF and the reason side by side.

## Key facts

- Build time: 6 to 9 weeks
- Price: $15,000 to $35,000 per project. One entity with 2-way PO matching sits at the low end. Extraction fees and model usage billed at cost.
- Typical stack: n8n or Make, Gmail or Microsoft 365 shared mailbox, Azure Document Intelligence, Google Document AI or Amazon Textract, OpenAI or Anthropic vision models for difficult layouts, QuickBooks Online, Xero, NetSuite or Sage Intacct, Retool or a small Next.js review app, Postgres for the audit log, Slack or Microsoft Teams for approvals
- Runs on your accounts; model and tool usage billed at cost

## How it runs, step by step

1. **Collect the invoices.** A shared AP mailbox in Gmail or Microsoft 365 and an upload folder feed one queue. Multi-invoice PDFs are split, and statements, reminders and marketing email are set aside.
2. **Extract the fields.** An extraction model reads vendor, invoice number, dates, totals, tax, PO number and line items. Each field comes back with a confidence score and its position on the page.
3. **Match the vendor.** The vendor is matched to your vendor master by tax ID, name and remit-to details. A new vendor or changed bank details stops the invoice for review.
4. **Check against the PO.** Quantities and prices are compared with the purchase order and, where you track them, goods receipts. Differences inside your tolerance pass; anything outside becomes an exception.
5. **Catch duplicates and errors.** The pipeline looks for the same invoice number or amount from the same vendor, totals that do not add up, wrong tax and dates outside the open period.
6. **Review the exceptions.** Exceptions land in a review screen with the original PDF beside the extracted fields and the failed check named. The reviewer corrects or approves, and the change is logged.
7. **Approve and post.** Clean invoices follow your approval rules by amount and cost center, then post as bills with the PDF attached. Payment stays in your existing payment run.

## What the system is made of

- **Intake mailbox:** One address suppliers send to, watched by the workflow.
- **Document classifier:** Splits bundled PDFs and filters out anything that is not an invoice.
- **Extraction engine:** Azure Document Intelligence, Google Document AI or a vision model, chosen by testing on your invoices.
- **Validation rules:** Vendor, PO, duplicate, tax and arithmetic checks written in plain code.
- **Exception queue:** A review screen in Retool or a small web app, showing the PDF and the failed check.
- **Approval router:** Amount and cost center rules that mirror your delegation of authority.
- **Posting connector:** Creates bills with attachments through your accounting system's API.
- **Audit log:** Every document, field value, correction and approver, stored on your accounts.

## What could go wrong and what stops it

- **Changed bank details.** Invoice fraud often arrives as a genuine-looking invoice with new bank details. The pipeline never updates vendor records. Any change to bank details or remit-to address stops the invoice and asks finance to confirm by phone, using the number already on file.
- **Low-confidence fields.** Each field has a confidence threshold set during the pilot, stricter for totals, tax and PO numbers. Anything below it goes to review with the source region of the PDF highlighted. Values are never corrected silently.
- **Duplicate invoices.** Exact and near-duplicate checks run on every invoice: the same number with different formatting, or the same vendor and amount within 30 days. Suspected duplicates are held, never posted.
- **Posting errors and permissions.** Each bill is posted once, keyed to its document ID, so a retry cannot create a second bill. The pipeline's API user can create bills and attach files. It cannot pay, delete or edit posted entries.

## How does matching against purchase orders work?

Matching is where most of the value sits, and it is plain code. The model's job ends once fields are extracted and checked.

A 2-way match compares the invoice with the purchase order: same vendor, a valid open PO, unit prices and quantities within tolerance. A 3-way match adds the goods receipt, so you only pay for what arrived. Tolerances are yours to set, for example 2% or $50 per line, whichever is lower, and they live in a table your controller can edit.

Line matching is the hard part. Supplier descriptions rarely match your item master ("Widget blue 10pk" against "WGT-BLU-10"). The pipeline tries SKU and supplier part number first, then quantity and price, and only then asks a model to suggest a match, which goes to review the first time it appears. Once a reviewer confirms it, the mapping is saved and that pair matches automatically from then on.

Invoices without a PO (utilities, software subscriptions, professional fees) skip matching and go to the budget owner for that vendor or cost center. Recurring ones are compared with the last 3 months and flagged when the amount jumps.

If you do not use purchase orders at all, the pipeline still extracts, checks duplicates and routes approvals. That is a smaller build, near the low end of the range. More on the extraction side in [intelligent document processing](https://methodcrew.com/glossary/intelligent-document-processing).

## What does invoice processing automation cost to build and run?

The build is a fixed price between $15,000 and $35,000, inside our $12,000 to $40,000 band for [document processing](https://methodcrew.com/services/intelligent-document-processing). The low end covers one entity, one accounting system, invoices by email and 2-way matching. Three-way matching, several entities or currencies, line-level coding to cost centers and a backfill of open invoices push it toward the top.

Running costs are billed at cost on your accounts:

| Item | What drives it | Notes |
|---|---|---|
| Extraction | Pages processed | Published prices run from about $30 to an estimated $560 per 1,000 pages ([Parseur, September 2026](https://parseur.com/blog/idp-pricing)) |
| Model calls | Hard layouts and line matching | Cents per invoice |
| Workflow platform | Executions | Self-hosted n8n on a small server, or a cloud plan |
| Review app | Reviewer seats | Retool seats, or nothing extra for a custom app |
| Care plan (optional) | New layouts, vendor changes | $1,500 to $3,000 a month |

The cost that matters more is reviewer time. Thresholds start strict, so in the first weeks more invoices go to review. As vendor mappings build up, the share that passes untouched rises. We report that pass rate weekly, because it tells you whether the build is paying back. See [how to measure automation ROI](https://methodcrew.com/insights/measure-ai-automation-roi).

## When should you not build a custom invoice pipeline?

If you process fewer than a few hundred invoices a month, start with capture tools built for that volume. Dext, Hubdoc and the invoice capture inside BILL and Ramp read invoices into your ledger for a monthly fee, and a person approves each one. That is often all a 30-person company needs.

A custom pipeline is worth the cost when the checks are the hard part: PO and goods-receipt matching, several entities, unusual layouts, or approval rules those tools cannot express. It also makes sense when invoices arrive alongside other documents you want read the same way.

It is the wrong place to start if your vendor master is unreliable. Duplicate vendors and out-of-date bank details mean a large share of invoices gets flagged on day one, and reviewers learn to click through alerts. We clean the vendor list during the [Map stage](https://methodcrew.com/approach#stage-map), or ask you to do it first.

If invoices are one part of a bigger document problem (delivery notes, claims, contracts), scope it as [intelligent document processing](https://methodcrew.com/services/intelligent-document-processing) across those document families, so extraction and review tooling is built once. For contract terms, see the [contract review assistant](https://methodcrew.com/blueprints/contract-review-assistant). And if the slow part is getting approvals signed off, a [business process automation](https://methodcrew.com/services/business-process-automation) project on the approval chain may be the cheaper fix.

## Common questions about this build

### How accurate is AI at extracting invoice data?

It depends on your invoices, so we measure it on yours. In the pilot we run 200 to 500 of your past invoices, where the correct values are already in your ledger, and report accuracy field by field. Clean PDFs read better than phone photos. Fields below the confidence threshold go to review, so errors are caught before posting.

### Does the pipeline pay suppliers?

No. It creates bills in your accounting system and attaches the original PDF. Payments stay in your existing payment run, approved by the people who approve them today. The pipeline's API user has no permission to pay, delete or edit posted entries, which keeps the separation of duties your auditors expect.

### Which accounting systems can it post to?

The blueprint is designed for QuickBooks Online, Xero, NetSuite and Sage Intacct, which all have APIs for creating bills with attachments. Other systems work if their API can create a bill or draft with line items. Some on-premise systems need a file import instead of an API, which we check in the scoping call.

### What happens with invoices that have no purchase order?

They skip matching and go to the budget owner for that vendor or cost center, with the extracted fields and the PDF. Recurring invoices such as rent, utilities or software are compared with previous amounts and flagged if they change by more than the tolerance you set. Everything else about the flow stays the same.

### Can it read scanned or photographed invoices?

Yes, within limits. Extraction services and vision models read scans and photos, but accuracy drops with skew, low resolution and handwriting. Low-confidence fields go to review with the image beside them. If one supplier keeps sending poor scans, the practical fix is asking them to email the PDF their system produces.

## Services that build and run it

- [Intelligent document processing](https://methodcrew.com/services/intelligent-document-processing): Invoices, contracts and forms turned into clean, checked data. $12,000 to $40,000 per project.
- [Business process automation](https://methodcrew.com/services/business-process-automation): One cross-team process, such as client onboarding, rebuilt so the hand-offs run on their own. $5,000 to $40,000 per project.
- [Automation care plans](https://methodcrew.com/services/automation-care-plans): Monitoring, fixes and small changes for the automations you rely on. $1,500 to $6,000 per month.

## Sectors where it fits

- [Professional services](https://methodcrew.com/industries/professional-services): Your firm sells expert time, and too much of it goes to finding last year's clause, chasing clients for documents and retyping intake forms. We build the systems that give those hours back, inside the confidentiality rules you already work under.
- [Ecommerce and retail](https://methodcrew.com/industries/ecommerce-retail): If your support queue is mostly people asking where an order is, and your catalog was written in a hurry before launch, both are fixable with systems that read your order data and product data directly.

## Next step

Request a 30-minute scoping call: https://methodcrew.com/contact. Or email contact@methodcrew.com.