AP automation in 2026: what the research says, and what a 20-person business should do about it

Every autumn brings a fresh stack of automation surveys. Most are vendor marketing with a chart attached. A few are worth reading. We went through the 2026 crop with a specific reader in mind: the owner or operations lead of a business with 10 to 200 people, no finance department, and a supplier invoice problem they have been meaning to fix. Here is what the numbers say, where they mislead, and what we would actually do.

The numbers that hold up

• Manual keying is still the norm. The Institute of Financial Operations and Leadership, in its 2026 Accounts Payable Automation Trends research run with SAP Concur, finds most AP teams have not fully automated their core workflow. Roundups of the same survey by Dokka and DocuClipper put the share of teams still keying invoices into the accounting system by hand at 66 to 68%.
• Intent is high. About 1 in 5 teams describe themselves as fully automated, and 41% say they plan to automate payables within 12 months. That gap between intent and action has barely moved in 3 years.
• Small businesses have crossed the AI line. Business.com's 2026 Small Business AI Outlook puts the share of US small businesses investing in AI at 57%, up from 36% in 2023. Upwork's State of AI in SMBs reports the fastest growth in companies of 10 to 100 people.
• Errors are the hidden cost. Several 2026 benchmarks put the share of manually processed invoices containing at least 1 error near 4 in 10. The exact figure varies by study. The direction does not.

Where the numbers mislead

The cost-per-invoice figures that vendors love, manual at 12 to 15 dollars against automated at 2 to 5, come from IOFM and Aberdeen benchmarks built on mid-market and enterprise AP departments. They assume a team, an ERP, and an approval chain. A 20-person business has an owner, a bookkeeper 2 days a week, and a shared inbox. The saving is real, but it is not measured in dollars per invoice. It is measured in the bookkeeper's Thursday afternoons and in the price list that is finally current when someone quotes a job.

The adoption figures mislead in the other direction. "Investing in AI" in a survey includes buying a subscription that drafts emails. It says nothing about whether any recurring process actually runs without a person. In our own client base the honest number is much lower: most businesses that describe themselves as using AI have not yet automated a single end-to-end workflow. That is not a failure. It is where the leverage is.

What has actually changed since 2024

2 things. First, reading a messy document is now a solved step. A model reads a supplier invoice in any layout into structured line items, with a confidence score per field, for a fraction of a cent. The template parsers that broke every time a supplier changed their PDF are obsolete. Second, the floor under small-business tools got stronger: Google Sheets now holds 20 million cells, Apps Script is a Workspace core service with regional data, and no-code flows can call real code. The pieces for a proper pipeline exist inside tools a 20-person business already pays for.

What has not changed is the part that decides whether the project survives week 2: duplicates, layout drift, timeouts, and silent failure. A model reads the invoice. Rules decide whether to trust it, where the row goes, and who gets told when it does not fit. That is the whole design, and it is why we say systems first and AI where it counts.

A 3-step plan that fits a business without a finance department

• Step 1, 1 week: pick the single document type that costs the most Thursday afternoons, usually supplier invoices, and count them for a week. Sources, layouts, where they land now, who touches them. Most businesses are surprised the count is under 100 a month. That is small enough to automate fully and large enough to matter.
• Step 2, 2 to 4 weeks: build the pipeline for that 1 document type. Intake from the inbox or a folder, a model reading each document into rows with a confidence score, checks for required fields and duplicates, a review tab for anything under the confidence line, and a run log. Land it in the Sheet or accounting system you already use. Do not buy an AP platform to do this. You own the code, it runs in your account, and there is no per-invoice fee.
• Step 3, ongoing: watch the review tab for a month. The rows that land there tell you which suppliers need a rule and which fields need a check. Then add the second document type. Delivery notes against orders and statements against the ledger follow the same pattern.

The survey question is "have you automated accounts payable". The useful question is "which document still gets typed in by a person every week, and why". Answer the second and the first takes care of itself.

We built exactly this for an independent retailer with 14 suppliers and 14 invoice layouts. It reads every invoice from Gmail, lands the lines in per-supplier tabs, and rebuilds a master price list of more than 4,200 items every day. It has run unattended since early 2026. The write-up is on this blog under AI invoice processing that actually ships, and the pattern transfers to any business that receives documents it re-types.