What AI Really Costs to Run Each Month: A Small Business Guide

Most small businesses pay for AI two ways: a flat monthly fee per user, and pay-per-use charges billed by the token. Seats are predictable. Usage is what spikes, driven by agents that take many steps, reasoning modes, top-tier models, long chats and public-facing bots. Estimate the cost per task, set caps, and review the bill monthly.

For a lot of owners, AI started as a $20 subscription. Then someone added an AI feature to the CRM, a chatbot went on the website, and a contractor built a workflow that calls an AI model in the background. The bill stopped being one line.

That's where surprises come from. In Pax8's Q3 2026 SMB AI Pulse (October 12, 2026), 49% of businesses that paid for AI in the past year got at least one bill or usage charge higher than they expected. In the United States it was 57%. Pax8 says usage-based charges are the hardest to forecast, and that they grow as businesses build more agents and workflows. A separate July 2026 survey of 363 US owners by Clarify Capital found the average monthly AI bill rose 343% in a year.

Here's how the billing works, what makes it jump, and how to keep it in a range you chose on purpose.

How do small businesses pay for AI?

There are two models, and more plans now mix them.

• Per seat. A fixed price per person per month, like ChatGPT Business or a Claude plan. You know the number in advance. Heavy users hit usage limits instead of a bigger bill.
• Per use. You pay for the amount of text the AI reads and writes, measured in tokens. This is how most website bots, automations and custom builds are billed, usually through the AI company's developer platform (its API).

The line between them is blurring. Anthropic's pricing page lists Claude Pro at $20 a month ($17 a month billed annually) and says paid plans can turn on usage credits at standard API rates when someone hits a limit. Its Enterprise plan is $20 per seat per month plus usage billed at API rates, "so cost scales with the models and tasks your team runs." OpenAI's business pricing page prices ChatGPT Business per user per month, notes that Business and Enterprise customers can buy credits for more access, and says that starting in December, plan usage will share one limit across ChatGPT and its Codex coding tool.

So even a "flat" plan can now carry a usage line. Check which of your tools can charge beyond the seat price, and who on your team can turn that on.

What is a token, and why does output cost more?

A token is a chunk of text, roughly a short word or part of a longer one. The AI is billed for the tokens it reads (input) and the tokens it writes (output), usually priced per million.

Output costs more. On Anthropic's published list, one of its mid-tier models, Claude Sonnet 4.6, is $3 per million input tokens and $15 per million output tokens. Its smaller Haiku 4.5 is $1 and $5. That 5-to-1 ratio between output and input shows up across its list. The practical point: a task that writes a lot costs more than one that reads a lot, and the model you pick can change the price several times over.

What makes an AI bill spike?

In my experience the jump almost never comes from a person chatting more. It comes from how a system is set up. These are the usual causes, each backed by the vendors' own documentation:

1. Agents that take many steps. An agent that searches, reads, decides and acts may call the model ten or twenty times to finish one task. Each call is billed. Pax8's finding that usage charges grow with agents and workflows fits this.
2. Long conversations and growing context. The model doesn't remember on its own; each new turn sends the earlier conversation again. OpenAI's conversation state guide says all previous input tokens in a chain are billed as input tokens. A bot that keeps a long history, or an agent that keeps pasting in tool results, pays for that history again on every step.
3. Reasoning modes. "Thinking" models write hidden reasoning before they answer. OpenAI's reasoning guide says those reasoning tokens are billed as output tokens and can run from a few hundred to tens of thousands per request. Anthropic's documentation says the same about its thinking tokens.
4. Using the top model for everything. A frontier model can cost several times more per token than a smaller one. Most routine work, like sorting emails or pulling fields from a form, doesn't need it.
5. Paid tools inside the task. Some features are billed on top of tokens. Anthropic, for example, lists web search at $10 per 1,000 searches, plus the tokens for the results.
6. Public-facing bots with no limits. A website or phone bot answers whoever shows up, including spam, bots and people who chat for twenty minutes. If you're considering one, Should a Small Business Let AI Answer the Phone? covers the setup side.

What does that look like in dollars?

Here's an illustrative example. The volumes are made up to show the math; the prices are Anthropic's published Claude Sonnet 4.6 list prices as of November 2026. Your tool, model and volumes will differ.

The math for the bot: 5 million input tokens at $3 per million is $15, plus 0.5 million output tokens at $15 per million is $7.50. For the agent: 200 million input tokens is $600, 10 million output tokens is $150, and 3,000 searches is $30.

Same number of jobs, about 35 times the bill. Nothing went wrong in the second case. That's simply what multi-step work costs, and it's why a budget built on the first number gets blown by the second.

How do you estimate AI costs before you build?

Estimate per task, not per month. You need four numbers:

1. How many times a month the task runs. Use real counts from your inbox, phone log or job system.
2. How many model calls each task takes. One for a simple answer; often five to twenty for an agent.
3. Roughly how many tokens go in and out per call. Your developer or vendor can measure this from a small test run. Both OpenAI and Anthropic report token counts on every response.
4. The price per million tokens for the model you'll actually use, from the vendor's current pricing page.

Multiply it out, then add a margin for growth and retries. Run the real workflow for a week with a small group before you trust the estimate. Then put the result in your annual plan. Your 2027 AI Plan and Budget on One Page shows where it goes.

How do you cap and review the bill?

Caps first, then habits.

• Turn on spend controls. OpenAI lists usage analytics and spend controls on ChatGPT Business, and Anthropic lists user and organization spend controls on its Team and Enterprise plans. On the developer side, set a monthly limit and an alert below it.
• Limit what one request can spend. OpenAI's reasoning guide recommends setting a maximum on output tokens so a single request can't run away.
• Match the model to the job. Use a smaller model by default and send only the hard cases to the expensive one.
• Trim the context. Start fresh conversations, summarize long histories, and send the AI only the records it needs.
• Use batch pricing for work that can wait. Anthropic and OpenAI both offer a 50% discount for requests that can run in the background, like overnight report drafts or bulk tagging.
• Keep a human approval step on agents. Draft-and-approve limits how far an agent can go before someone looks. Your First AI Agent explains the guardrails, and How to Connect ChatGPT or Claude to Your Business Data covers keeping access narrow.
• Review monthly. Put every AI charge on a business card, list them by tool and workflow, and compare each one with what it saved. How to Measure AI ROI in a Small Business gives a simple way to do that.

How Caprock Agility helps

We don't resell AI subscriptions or take a cut of usage. We help you see where your AI money goes and build workflows with a cost you can predict.

Our AI Readiness Audit ($2,500, 3–5 business days) inventories the AI you're already paying for, including seats, add-ons and usage charges, and maps which workflows are worth the spend. If you're ready to build, EDGE Starter (starting at $8,500, 4–6 weeks) sets up one workflow end to end with the right-sized model, spend caps, human approval and a Day 30 outcomes review.

You can compare every option on our services and pricing page.