You asked a vendor what their AI would cost to run and got a shrug and the word “usage.” You looked at a provider’s pricing page and found numbers like “$3 per million input tokens,” which is not a unit any business plans in. You have a budget meeting on Thursday and no number to put in it.
The AI cost for small business use is knowable. It breaks into four lines, three of which are boring subscriptions and servers, and one of which (the model usage) is metered in a way that needs one translation before it makes sense. This article does the translation and gives you rules of thumb for each line.
By the end you will be able to estimate a monthly running cost for any proposed AI system, and you will know the two costs that never appear in the proposal.
What this actually is
Running an AI system in a business costs money in four places. First, the model itself: every time text goes to the provider (OpenAI, Anthropic, Google) and an answer comes back, you pay by the amount of text, measured in tokens. Second, the platform that runs the automation around the model: a subscription to Make, Zapier, or n8n cloud, or a small server if you self-host. Third, storage and connections: a database, a phone number, an email sending service, a place to keep documents. Fourth, people: the minutes a day someone spends approving drafts and the hour a month someone spends reading logs and fixing what drifted.
The useful analogy is a delivery van. There is fuel (model usage, metered by distance), the lease (platform subscriptions, fixed monthly), insurance and parking (storage and connections, small and fixed), and the driver (people, the largest line and the one everyone forgets to count). Vendors quote you the fuel. Your budget needs all four.
AI cost for small business systems, line by line
1. Turn tokens into dollars per thousand conversations
A token is a chunk of text, roughly three quarters of a word. Providers publish two prices per model: one for input tokens (what you send, including instructions and any documents) and a higher one for output tokens (what the model writes back). Both are quoted per million tokens.
To make that usable, estimate one typical interaction. Picture a support conversation: the standing instructions are 1,500 tokens, the retrieved passages from your knowledge base are another 1,500, the customer’s messages and the running history add up to 1,000, and the model writes 300 tokens per reply across four replies. That is roughly 12,000 input tokens and 1,200 output tokens per conversation, because the instructions and passages are re-sent with every turn.
At the rates providers publish today for a mid-tier model (check the current pricing page; a rough range is single-digit dollars per million input and two to five times that for output), that conversation costs somewhere around three to six cents. So a thousand support conversations a month costs roughly $30 to $60 in model usage. A simple email classification is far less text (perhaps 800 tokens in, 20 out) and costs a fraction of a cent, so a thousand of those is under a dollar. This is the arithmetic. Do it once for each use and the “usage” line stops being a mystery.
2. Understand subscriptions versus usage
There are two ways to pay for a model. A chat subscription (Claude, ChatGPT, Gemini for individuals or teams) is a flat fee per person per month, typically in the $20 to $30 range for a standard seat and more for business tiers; check the current page. It buys a person a tool. It does not buy your systems anything, because your automations cannot use a seat.
Usage pricing through the provider’s API is what a built system uses. You pay for exactly the tokens consumed, with no seat fee, and the bill scales with the volume of work. For most small businesses the usage bill is smaller than one seat, because the automations use small, targeted prompts. The two are not alternatives; a business usually has a few seats for staff and a usage account for the systems.
3. Pick the cheapest model tier that does the job
Every provider sells a range: small fast models that cost well under a dollar per million input tokens, mid-tier models at a few dollars, and top-tier reasoning models that cost several times more. The difference in price is often ten to thirty times between the bottom and the top of the range.
Most business tasks belong at the bottom. Classifying an email, extracting fields from an invoice, deciding whether a call is urgent: a small model does these well. Drafting a careful reply to an upset customer or reasoning through an ambiguous contract clause belongs to a mid or top tier. In the systems we build, the cheap model handles the first pass and the expensive one is called only when the cheap one flags a hard case. That single design choice usually cuts the usage line by more than half. The full set of techniques is in LLM Cost Optimization in Production: A Practical Guide.
4. Budget the automation platform
The software that watches for triggers, calls the model, and writes results into your systems is either a hosted no-code platform or something self-hosted. Zapier and Make charge by a mix of monthly tier and operations (each step in each run counts), and the tiers most small businesses land on run somewhere in the $20 to $100 a month range, rising with volume. n8n has a cloud plan in a similar band and can also be run on your own server.
The trap is per-operation pricing at volume. An automation with eight steps that runs 5,000 times a month is 40,000 operations, and that lands you in a higher tier than the marketing page suggests. Estimate operations, not just workflows, before choosing. The comparison in n8n vs Make vs Zapier for a Business Without a Developer goes through the pricing models in detail.
5. Price voice separately, by the minute
AI phone answering is priced differently because it stacks three meters: the telephone minutes (Twilio or similar, a few cents a minute plus a dollar or two a month for the number), the speech-to-text and text-to-speech, and the model. Platforms like Vapi, Retell, and Bland bundle some of this and charge per minute of call, and the all-in figure is typically in the range of ten to twenty cents a minute once the model and voice are included; check the current pricing pages because these move.
So a business taking 600 after-hours calls a month averaging three minutes is 1,800 minutes, which is somewhere between $180 and $360 a month all in. That is a real cost, and it is also usually far less than a human answering service for the same hours. See AI Receptionist vs Answering Service: An Honest Comparison for that comparison.
6. Add the small server, if there is one
If anything is self-hosted (n8n, a small web app, a database), it runs on a virtual server from a provider like Hetzner, DigitalOcean, or a cloud hosting service. For a business this size, a machine with a few gigabytes of memory is enough and costs somewhere between $5 and $30 a month. Backups add a few dollars. A domain name is $10 to $20 a year.
A self-hosted setup also has a labor cost: someone has to apply updates and check backups. A rough rule is an hour a month. If nobody in the business will do that, use the hosted version of the platform and pay the subscription instead. Self-Host n8n for Your Business Without Being Held Hostage lays out what the job involves.
7. Count storage, search, and messaging
A system that answers from your documents needs somewhere to store and search them. If you already run a Postgres database, the pgvector extension adds meaning-based search for no extra fee. A hosted vector service like Pinecone has a free tier and paid plans starting in the tens of dollars a month. Sending email through a transactional service (Resend, Postmark) is a few dollars a month at small volume. A CRM or helpdesk you already pay for (HubSpot, Zendesk, Intercom) may charge extra for API access on some tiers; check yours.
8. Count the review time honestly
This is the first hidden cost. Every well-built AI system has a person in the loop somewhere: approving drafts before they go to customers, glancing at call summaries, handling the cases the system flagged as unsure. In our experience this runs ten to twenty minutes a day for a system handling a few hundred items, and it is real payroll.
It is also the best money in the budget. The review step catches the wrong answer before a customer sees it, and it is the source of the corrections that make the system better. A proposal that claims zero review time describes a system with no safety net.
9. Count the upkeep
The second hidden cost. Knowledge bases go stale. An upstream tool changes its export format. The provider retires a model version and the replacement behaves slightly differently. Someone has to read the logs monthly, refresh the material, and re-run the test questions. Budget two to four hours a month of a capable person’s time, or a maintenance arrangement with whoever built it. Systems with no upkeep budget decay quietly, and the owner finds out from a customer. Automation Error Handling for Businesses Tired of Silent Failures covers what goes wrong and how to catch it.
Picture a business like this one
The business below is a composite of the kind of company that writes to us, not a client. The numbers describe the shape of the problem, not a case study.
Picture a business like this one: an e-commerce store selling specialist kitchen equipment, eleven staff, about 4,000 orders a month, and a support inbox that receives around 2,500 emails a month. Two people answer them. Most are “where is my order,” “can I return this,” and “does this fit my model,” and each one takes four to eight minutes because the answer has to be looked up in the order system and the product database.
What gets built is an inbox pipeline: each email is classified by a small model, order-status questions are answered automatically from the order system with the tracking link, return questions get a drafted reply from the return policy with a citation, fit questions get a drafted answer from the product specs, and anything angry, ambiguous, or involving a refund over a set amount goes straight to a person with a summary. Every automated reply above a low confidence threshold is sent; the rest are drafts a human approves.
What changes is that the two support staff spend their day on the hard cases and on the drafts, replies go out in minutes instead of hours, and the store owner can read a monthly log of what the system answered and where it was unsure.
What it costs to run
For the store above, a realistic monthly sheet looks like this. Model usage: 2,500 emails classified by a small model is a few dollars; roughly 1,800 of them get a drafted or automated reply using a mid-tier model with retrieved passages, at perhaps 4,000 tokens in and 200 out each, which is somewhere in the $20 to $40 range. Platform: a Make or n8n cloud tier that covers around 20,000 operations, or a self-hosted n8n on a $15 server. Storage and search: pgvector on the existing Postgres, $0. Email sending: through the existing helpdesk, $0 extra. Total tooling somewhere between $40 and $120 a month, with the exact figure depending on the current pricing pages.
Then the people: roughly fifteen minutes a day of draft approval across the two staff, and about three hours a month of upkeep. That is the largest line, and still a fraction of the time the inbox used to consume.
The mistakes we see most
Most errors in estimating AI cost for small business systems come from the same handful of habits.
- Budgeting only the model usage. The platform, the server, the review time, and the upkeep are usually larger, combined, than the tokens.
- Using the top-tier model for everything. Sorting emails with a reasoning model is like sending a courier for every envelope. Use the tier the task needs.
- Ignoring per-operation pricing. A cheap-looking automation tier becomes expensive at 40,000 operations a month. Estimate volume first.
- Comparing a seat subscription to a built system. A $25 chat seat is a personal tool. It does not do the work of an automation, and its price tells you nothing about running one.
- Buying “unlimited AI” from a vendor without asking what model sits underneath. Unlimited usually means a small model with a low cap, and the fine print says so.
- No upkeep line. The proposal ends at launch. The costs do not.
When to bring in help
An owner can run the numbers in this article alone, and should, before any purchase. Off-the-shelf tools handle simple cases at low volume: a Zapier or Make automation with one AI step and a chat seat or two will cost under $100 a month and needs no developer.
A developer earns their fee when volume makes per-operation pricing hurt, when the cheap-model-first design has to be built in, when the system has to talk to your order system or CRM with proper error handling, and when the review and log steps have to exist rather than be promised. Those choices decide whether the running cost is $80 a month or $800.
Levelbrook builds these systems for businesses on a fixed price from a written scope, with every running cost line above listed in the scope before you sign, and everything running in accounts you own so the bills are yours and visible. If you want a running-cost estimate for a specific idea, the form below is where to ask.