What does AI automation cost to run each month?
Typically $50–$500 a month for a small business, covering four line items: model usage, software subscriptions, integration platform fees, and someone watching the system. The build is a one-time price; this is the bill that arrives every month after.
The build side of the market is noisy but at least discussed — we publish ours, an AI Readiness Audit at $3,500–$8,500 and an AI Implementation Sprint at $12,000–$45,000, fixed in writing before work begins. The running side is where buyers get surprised, because most sellers simply never bring it up. This guide walks through each of the four line items, shows what drives them up or down, and ends with the case nobody selling automation wants to mention: the workflows where the monthly bill exceeds the monthly value, and the right answer is to leave the work manual.
What does the model bill actually look like?
For most single-workflow systems, $20–$100 a month; document-heavy workflows can run several times that. Models charge by the token — roughly three-quarters of a word — metered in both directions, for what the machine reads and what it writes.
The shape of your workflow decides the bill. A chat-light system — an assistant that answers a few dozen staff questions a day against your procedures — reads and writes small amounts of text and lands at the bottom of the band, often a few dollars a day or less. A document-heavy system is different arithmetic. When the machine reads every 40-page contract, every intake packet, every invoice attachment, each document is thousands of tokens in and often thousands more out in summaries and drafts. A professional-services firm running proposals and engagement letters through a system will spend more per month on tokens than a shop using the same model to answer scheduling questions — not because anyone did anything wrong, but because tokens are the metered fuel. The Ampersand essay The Currency of the Machine explains the meter in plain English, and The Buildings Behind the Intelligence covers the physical infrastructure the meter is paying for.
You can rough out your own bill before anyone builds anything. Count the documents the workflow touches in a month and multiply by their typical length: two hundred 3-page invoices is a modest token load; forty 40-page contracts is a heavy one, even though it is fewer documents. Precision does not matter at this stage — what matters is knowing whether you are a $30-a-month workflow or a $300-a-month workflow before you commit to a build price, because the answer occasionally changes whether the build is worth doing at all.
What software subscriptions sit underneath a working system?
Plan on $30–$200 a month in tools around the model: the platform that hosts the workflow, the integration layer that moves data between your systems, and sometimes a specialty service for documents. Each piece is small; together they are usually the second-largest line.
A typical stack for one automated workflow looks like this: an integration platform (the plumbing that connects your inbox, your file storage, and your system of record) at $20–$100 a month depending on how many times it runs; a hosting or workflow tool at $0–$50; and, for document work, OCR or storage services at $10–$50. One caution worth flagging separately: per-seat AI add-ons — the $20–$30-per-user-per-month copilot licenses sold inside office suites — are a different bill entirely. Twenty seats of a copilot is $400–$600 a month before any workflow exists. That can be worth it, but do not confuse it with the running cost of an automated workflow; a scoped system usually costs less per month than blanket seats and does something specific.
Who watches the system, and what does that cost?
Someone has to own the system after handoff — either a named person in-house spending a few hours a month, or a maintenance retainer. Skipping this line item does not make it free; it makes it invisible until something breaks quietly.
Watching a system means concrete work: reviewing the exception queue where the machine flags what it could not handle, spot-checking outputs so drift gets caught early, and applying updates when a vendor changes an API or retires a model version. In-house, that is commonly two to six hours a month for a single workflow — a real cost even though no invoice arrives. If you would rather not staff it, our Managed Services covers exactly this: a monthly retainer, priced by what is running, with no lock-in — cancel any month and the system is still yours, because you owned it from handoff. Either answer works. Having no answer is the only wrong one.
Why do running costs shift over time?
Because the ground moves: model vendors retire versions, change prices, and update APIs, and each shift can require a few hours of retuning. The good news is the price direction — per-token costs have generally fallen, sharply, year over year.
The pattern to expect is occasional, not constant. A model deprecation forces a migration to the successor, and the successor never behaves identically — prompts get re-tested, edge cases get re-checked, and a workday or two disappears. An integration partner changes an endpoint and the plumbing needs a fitting replaced. Budget for two or three such events a year on a single workflow. This is most of what a maintenance retainer actually buys: not standby heroics, but somebody whose job it is to absorb these shifts before you notice them.
When do running costs exceed the value?
On low-volume workflows. A task performed a handful of times a month cannot repay even $50 in monthly costs plus the attention the system needs — so do not automate it.
Run the arithmetic before anyone runs a demo. If a task takes your bookkeeper two hours a month, full automation saves at most two hours a month — against $50 or more in running costs and a build price on top. That is a hobby, not an investment. A small retail shop reconciling four supplier invoices a month should keep reconciling them by hand; the same shop answering two hundred customer emails a week has a real candidate. Volume is the test, and it is why every audit we deliver marks some workflows “leave it manual” in writing. If a seller has never told you not to automate something, they are not scoping — they are selling. The free 20-point readiness checklist helps you sort your own list before anyone charges you to sort it.
There is a middle case worth naming, too. Some workflows are low-volume today and will not stay that way — the shop opening a second location, the firm about to take on a new contract type. For those, the right answer is usually “not yet, revisit at volume,” with a specific trigger written down: automate the invoice flow when it crosses fifty a month, not before. A dated threshold turns a vague someday into a decision that makes itself, and it costs nothing to write.
How do monthly costs compare across workflow types?
Three profiles cover most of what small businesses actually run. The ranges are typical figures for systems like the ones we scope — not quotes, because your volume decides your number.
| Profile | Typical monthly | What drives it | Watch for |
|---|---|---|---|
| Single chat-assist workflow — staff Q&A, drafting replies against your procedures | $50–$150 | Question volume; short texts in, short texts out | Per-seat add-ons creeping in beside it |
| Document-processing workflow — intake packets, invoices, contracts read and routed | $150–$400 | Document volume and length; OCR and storage on top of tokens | A single long-document type dominating the bill |
| Multi-agent back office — several workflows across departments, shared plumbing | $300–$500+ | Number of workflows and integrations; monitoring load | Maintenance hours growing faster than the token bill |
| Low-volume task — done a handful of times a month | Not worth it | Nothing — the costs exceed the value | Anyone trying to sell you this row |
Typical 2026 figures for small-business systems, stated as ranges — your document volume, run counts, and tool choices set the real number.
How do build costs and running costs fit together?
Think of the build as the mortgage and the running cost as the utilities: $12,000–$45,000 once for an implementation sprint at our published prices, then $50–$500 a month to keep it working. Both numbers should be on the table before you sign anything.
The full year-one picture, then: an audit at $3,500–$8,500 over 2–4 weeks, a sprint at $12,000–$45,000 fixed in writing over 4–12 weeks, and roughly $600–$6,000 in running costs across the first twelve months. For the build side of the market — hourly rates, boutique bands, Big Four minimums — the companion guide What an AI consultant costs lays out the whole shelf; the full price list is on one page. And one vetting question worth carrying into every meeting, including ours: ask for the projected monthly running cost next to the build quote. A builder who cannot estimate the monthly number has not actually scoped the workflow. Every engagement starts with a free 30-minute assessment, and we reply within 24 hours.