AI Invoice Processing for Small Businesses: Build It or Buy It, With Real Per-Invoice Costs

AI invoice processing for small businesses

A bookkeeper opens the shared inbox on Monday and finds 140 vendor invoices. Each one gets opened, read, typed into the books with a vendor, a date, a total and a category, then filed. At four minutes apiece, that is nearly ten hours of the week spent copying numbers from one place to another.

This is one of the few AI jobs where the work is boring, the inputs are structured enough, and the cost of a model call is so small that the economics are not close. It is also a job where a confident wrong answer costs real money, so the safety checks matter more than the model. This guide covers what actually works, what to buy versus build, and the steps to run it on your own invoices this week.

What the numbers look like on real deployments

Two case studies on this site show the range. In the first, a $10M accounting firm rebuilt its accounts payable workflow with two non-technical accountants using Cursor and Claude Code. The firm reports that cost per invoice fell from $7 to $0.20, that the agent hit 80 percent accuracy in the first week, and that it reached 98 percent after about six months of edge case training (builtwithagents.ai/strategy/accounts-payable-ai-agent-7-dollars-to-20-cents). Those are the firm's own numbers.

In the second, an automation consultant built a three-agent Claude pipeline for a small chartered accountant firm in Gurgaon that reads vendor invoices from a shared Drive folder and posts them into Tally. The consultant reports about 2,047 invoices a day, 94.3 percent posted automatically at launch, and an exception rate that dropped to about 3.1 percent after prompt tuning, with model cost around $10 a day (builtwithagents.ai/strategy/claude-agent-2000-invoices-day-accounting-firm). That is self-reported and not independently verified.

Notice the pattern in both. Nobody got to full automation on day one. The first week was roughly 80 to 94 percent, and the work after that was handling the weird cases. Plan for that curve and you will not be disappointed by it.

Buy it or build it

Before you build anything, check whether your accounting software already does the extraction. For many small businesses it does.

Xero includes Smart Document Capture on all three of its US plans, which are listed at $27, $59 and $97 a month as of October 2026 (the entry plan caps you at 5 bills a month, so most businesses need Growing). Hubdoc, a capture tool that syncs to Xero and QuickBooks Online, is reported at around $12 a month by third-party sources, and Dext's US business plan is reported at roughly $25 a month billed annually for 250 documents. Those last two figures come from comparison sites, so confirm them on the vendors' own pages before you budget.

If you want extraction as a building block instead, Veryfi's API lists invoices at $0.16 each, with a free tier of 100 documents a month, as of October 2026. Its paid Starter plan has a $500 monthly minimum, which rules it out for most small shops.

So the honest answer is: if your volume is under about 250 invoices a month and your books are on Xero or QuickBooks, use the built-in or add-on capture and stop reading. Build your own when you have odd inputs (multi-page supplier statements, invoices in several languages, PDFs from a dozen portals), when you need custom coding rules (job numbers, cost codes, project splits), or when the accounting package is something the capture tools do not support.

How to build it: the steps

Step 1: Pull 50 real invoices and tag them

Export the last 50 invoices from your inbox, including the ugly ones: photographed receipts, scanned PDFs, statements that list several invoices. Put the correct answer next to each one in a spreadsheet: vendor, invoice number, date, subtotal, tax, total, and the account you coded it to. This is your answer key. Without it you cannot tell whether the agent is 80 percent right or 98 percent right, and that is the number that decides whether to trust it.

Step 2: Set up one intake inbox

Create a single address such as invoices@yourcompany.com and forward every vendor to it. In n8n, use the Email Trigger (IMAP) node or the Gmail trigger on that address, or watch a Google Drive folder if your team drops files there. One door in means one place to monitor. n8n Cloud's Starter plan is listed at 20 euros a month billed annually with 2,500 executions, and the self-hosted community edition is free if you are comfortable running it yourself, both as of October 2026.

Step 3: Extract to a fixed schema

Send each attachment to Claude with a prompt that demands JSON in one exact shape: vendor name, invoice number, invoice date, due date, currency, subtotal, tax, total, and an array of line items. Tell it to return null for anything it cannot read, never to guess, and to include a confidence note for each field. PDFs and images can be passed straight to the Claude API, so you do not need a separate OCR step for most invoices.

Step 4: Run checks in code, not in the prompt

This step is what separates a toy from something you can leave running. Add plain code nodes after the extraction that test the output. Do the line items add up to the subtotal? Does subtotal plus tax equal the total? Is the date in a sane range? Does the vendor match a vendor already in your books? Has this vendor and invoice number been seen before? A failed check never gets posted. It goes to a review queue.

Step 5: Map to your chart of accounts

Give the model your list of expense accounts and ask it to pick one for each invoice, plus a one-line reason. Better still, build a vendor lookup table: once a vendor has been coded correctly twice, use the stored account and skip the guess. Most of your invoices come from the same twenty vendors, so the lookup does most of the work and the model only handles new ones.

Step 6: Post as a draft and route exceptions

Use your accounting software's API or an n8n node to create the bill as a draft or unapproved bill, not a paid one. Post only invoices that passed every check and have a known vendor. Send everything else, with the model's notes, to a Slack message or a Google Sheet that a person reviews once a day. Do not let this agent pay anything.

Step 7: Run it beside your current process for 30 days

For a month, keep doing the work the old way and compare. Measure how many invoices the agent posted correctly, how many it routed to review, and how many it got wrong without flagging. That last number must be close to zero before you switch over. The Gurgaon deployment above reported its exception rate falling over time because the prompt was tuned against real failures, and yours will do the same.

What this costs

The model cost is tiny. As of October 2026, Anthropic lists Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, and Haiku 5.5 at $0.10 and $0.50, with a 50 percent discount for batch processing. A typical one-page invoice should take a few thousand tokens in and a few hundred out, which is my estimate, not a measured figure. That works out to roughly one cent on Sonnet and well under a tenth of a cent on Haiku. At 150 invoices a month, you are looking at about $1.50 or less in model fees.

The real costs are your time and your hosting. Budget a weekend to build and tag your sample, plus an hour or two a week of review during the first month. Add n8n at 20 euros a month if you do not self-host. If you hire someone to build it, ask for a fixed price that includes the 30-day parallel run and the exception queue.

What breaks

Duplicates. Vendors resend invoices, and a statement can list invoices you already entered. Your duplicate check on vendor plus invoice number is not optional.

Changed bank details. A fake invoice or a spoofed vendor email with new payment instructions is a common fraud. The agent should flag any invoice whose bank details differ from the vendor's record, and a human should verify that by phone. Keep payment approval entirely out of the pipeline.

Handwriting and poor scans. Accuracy drops fast on photos of crumpled receipts. Route low-confidence reads to review rather than accepting them.

Tax and compliance. Sales tax, VAT and withholding rules vary by place. The agent can extract the numbers, but your accountant should own the coding rules.

If you process fewer than 30 invoices a month, skip all of this. A built-in capture feature or a part-time helper costs less than the time you would spend building and checking a pipeline.

Frequently Asked Questions