Where AI Actually Pays for Itself in an SMB Back Office (and Where Rules Are Cheaper)

Every tool your business already pays for grew an AI button this year. Your accounting package has one. So does your CRM, your helpdesk, and the scanner software nobody has updated since 2021. Somewhere in the middle of that noise is a real question that almost nobody answers straight: which parts of your back office genuinely get cheaper with a model, and which parts were always a job for fifty lines of ordinary code.

The gap between the marketing and the adoption is wide. The US Census Bureau's Business Trends and Outlook Survey, covering December 2025 through May 2026 and published in May, found between 17% and 20% of businesses reported using AI at all. Adoption climbs with size: 37% of firms with 250 or more employees, 32% of firms with 100 to 249, and under 20% for firms with fewer than 20 people. Most small companies are not behind. They are waiting for someone to tell them where it actually pays.

The question that settles most of it

Ask whether the input arrives in a predictable shape. If it does, and the decision follows rules a person could write on an index card, then ordinary software is cheaper, faster, and far easier to trust. AI earns its place when the input varies in ways you cannot enumerate, or when the task requires reading meaning rather than reading fields.

That single test resolves most back-office decisions correctly, and it is the test vendors never apply, because a rules engine is not a story they can charge a monthly per-seat fee for.

Four jobs, and what each one should actually run on

Reading invoices and receipts that arrive in forty layouts

This is the clearest win for machine learning, and it is not a new one. Every supplier formats an invoice differently, the total sits in a different place on each, and the moment you write a rule per layout you have signed up to maintain rules per layout forever. Trained extraction handles the variety without you enumerating it. If your accounts payable pile comes from more than about ten sources and they keep changing, this is where to spend.

Sorting and routing files

Almost always rules, and the answer is almost always boring. If the decision can be made from a barcode, a filename, a sender address, a form number in a fixed position, or which folder the scanner dropped it in, then a model is an expensive way to reproduce an if statement, and a less reliable one. We have built exactly this kind of pipeline for a print shop, sorting scanned documents to the right destination on a local server, and the useful part was never clever reading. It was the plumbing: watching a folder, handling the file that arrives half-scanned, sending the result to the right printer.

Matching payments to invoices

Mostly rules, with a small tail that is worth softening. Exact matches on reference and amount are trivial and should be automatic. The remainder, where a customer paid three invoices in one transfer or typed the reference wrong, is where fuzzy matching or a model earns its keep. Build the rules first and let the exceptions define what, if anything, needs to be smarter. Doing it the other way round means paying model prices to handle the easy 90%.

Drafting replies, summarising, and triage

Genuinely good, genuinely useful, and the one place where the hype is closest to true. Summarising a long email thread, drafting a first reply, classifying incoming messages by topic and urgency: these are language tasks with no rule set that ever worked. The condition is supervision. A draft a human approves is a time saver, while an unsupervised reply to a customer is a liability with a monthly fee.

What the AI costs when you buy it wholesale

Vendors sell AI as a plan tier. Underneath, most of it is a metered API with published prices, and the wholesale numbers are worth knowing before you agree to a per-seat uplift. These are Amazon Textract's published rates for US West (Oregon), checked this week.

What it does	Price per 1,000 pages
Detect Document Text, plain OCR	$1.50 (first 1M pages), $0.60 beyond
Analyze Expense, invoice and receipt fields	$10.00
Analyze Document Queries, ask for specific values	$15.00
Analyze Document Forms, full key and value pairs	$50.00

Source: aws.amazon.com/textract/pricing, August 2026. Language models are priced by token rather than by page, and the small ones are cheap: OpenAI lists gpt-4o-mini at $0.15 per million input tokens and $0.60 per million output tokens on its API pricing page. The flagship tier runs roughly thirty times that on input.

What this tells you is not "AI is cheap". It is that the model call is rarely the expensive part of an AI feature. When a subscription adds a meaningful monthly figure per user for AI, you are mostly paying for the integration, the interface, and the vendor's margin, not for the intelligence. Sometimes that is a fair trade. It is worth knowing which one you are buying.

The cost nobody quotes: the review queue

Accuracy is where AI projects quietly lose the money they saved. Suppose a system handles 400 documents a day and is right 95% of the time. That is an excellent result, and it still means twenty exceptions land on someone's desk daily, every one of which needs a human, a screen to fix it on, and a way back into the flow. That number is an illustration rather than a measurement, but the shape of it holds everywhere.

There is a second difference that matters more in a back office than any accuracy percentage. Rules fail loudly. A rule that expects a barcode and finds none throws an error, and you know immediately. Models fail quietly and plausibly: they return a total that looks entirely reasonable and is wrong by a decimal place. In accounting, a confident wrong answer is worse than no answer, which is why the design work around an AI feature matters more than the model choice. Confidence thresholds, a review queue that is genuinely pleasant to work in, and a hard rule about which values a human must always confirm.

How to tell a real AI feature from a relabelled one

Three questions, asked of any vendor or any developer, including us. What happens when it is not sure, and can we see and set that threshold? Does it improve from our corrections, or does it make the same mistake next month? And can we get the extracted data out into our own systems without buying another tier? A vendor who answers "it just works" is describing a marketing budget, not an architecture.

A sane order of operations

1. Pick the process that costs you the most hours, not the one that sounds the most modern.
2. Automate the predictable part with rules first. It is cheaper, it works this quarter, and it tells you exactly how big the messy remainder really is.
3. Measure what is left over for a month. Most teams discover the hard cases are a fifth of what they assumed, and occasionally that they are the whole job.
4. Apply a model only to that remainder, with a review path designed before the model is chosen.

This order also protects you from the most expensive mistake in the category, which is buying a platform to solve a process you have not yet written down. If you want the wider view of how the pieces fit, our guide to automating document processing for a small business covers the tooling landscape and what each option costs.

Where to point this

The useful version of this conversation is specific. Tell us the process, the volume, and where the documents come from, and we will tell you which parts are rules, which parts are worth a model, and roughly what each would cost to build. If the answer is that a subscription you already pay for does the job, that is what we will say. Start with our contact page and describe the process in a few sentences.

Agencies: if a client is asking for an AI feature and you need senior people who can tell the real cases from the theatre, that is what our white-label capacity is for.