The cost of AI for a smaller business: AI pricing for subscriptions, usage, hardware and staff time

The cost of AI for a smaller business comes from four places: seat subscriptions for tools such as ChatGPT or Copilot, usage fees when your own software calls a model, hardware if you run a model yourself, and the time your staff spend setting it up and checking its work. Subscriptions are priced per user per month, usage per million tokens, and a rented GPU server per hour. This guide gives the formula for each part, one table of vendor list prices checked on September 28, 2026, and a worked example for a company with 20 office staff.

Derik Lawlis Founder, ThriveAI

The four parts of the cost of AI

Each part is billed differently, so estimate each one on its own and add them up. The table gives the formula for each part. Vendor prices sit in a single table further down, so you can update the figures in one place when they change.

A loaded hourly cost is a person’s wage plus benefits, payroll taxes and overhead, divided by the hours they work.

Seat subscriptions

A seat is one named user on a business plan, such as ChatGPT Business, Claude Team, Microsoft 365 Copilot or Google Workspace with Gemini. You pay the seat price whether or not the person uses it that month. Most plans cost less per month with an annual commitment, and some set a minimum number of seats. OpenAI, for example, requires at least two paid seats for a ChatGPT Business workspace.

Some seats also leave usage open. Anthropic’s Enterprise plan charges a seat fee that “covers access only,” and bills all usage separately at API rates, so its monthly cost depends on how much your team uses Claude. AI workspace compares what each business plan includes, and ChatGPT alternatives for business in Canada compares the assistants on storage and model choice.

API usage

When your own software or an agent calls a model through an API, the connection one program uses to call another, you pay for tokens. A token is a small piece of text, often part of a word. Input tokens are what you send, such as instructions and a document. Output tokens are what the model writes back. Providers list prices per million tokens, and output costs more than input.

The formula is the one in the first table. Three things move the result:

• Tokens per job. A two-page purchase order and a 60-page contract differ by an order of magnitude. Measure your own documents on a small test before you estimate.
• The model. A provider’s smallest and largest models differ in price by a factor of ten or more, as the list price table shows. Test whether the cheaper model gets your jobs right, as AI evals explains.
• Discounts for repeated or delayed work. On most Claude models, reading a cached prompt prefix costs a tenth of the base input price, and the Batch API, which processes requests that can wait, charges half the standard price for input and output (Anthropic).

Claude API pricing covers Claude models and discounts in US and Canadian dollars.

Hardware for a local model

A local model is an open-weight model you run on a server you control, so prompts never go to an outside provider. The server needs a GPU, the graphics chip that does the model’s arithmetic, and the GPU’s memory limits the size of model it can run. Local LLM covers which models fit on which hardware.

You can buy the server or rent one. DigitalOcean, for example, rents GPU servers by the hour and lists NVIDIA H100, L40S, RTX 4000 Ada and RTX 6000 Ada machines in its Toronto data centre. Its pricing page says a GPU server is billed per second, and that a powered-off server is still billed “until you destroy the instance” (DigitalOcean). A server left running all month runs about 730 hours, so the formula is the hourly rate × 730.

A bought server has no hourly charge. Its cost is the purchase, the power, the space and the person who patches it, backs it up and replaces parts, spread over the years you use it. Private AI for business compares running your own server with a hosted model on what each protects.

Staff time

Staff time does not appear on any vendor invoice. How to build an AI agent notes that integration and testing take more of an agent’s budget than model usage does. Estimate staff time task by task:

• Setup: choosing the tool, connecting single sign-on, cleaning up folder permissions so the AI shows each person only what they may see, and writing the rules for staff.
• A test set: collecting 30 to 50 past jobs with the right answers for the first job you automate, as AI evals describes.
• Training: hands-on sessions on your own documents. AI training covers the format.
• Review: the time a person spends approving what the AI drafts. Human in the loop covers which actions need it.
• Upkeep: rerunning tests when a model or prompt changes, updating connections when a system changes, and moving to a new model when the old one is retired.

Multiply each task’s hours by the loaded hourly cost of the person who does it. For review time, compare the hours with the hours the job takes today, because the review replaces part of that work when the AI does the drafting.

List prices checked on September 28, 2026

This is the only table of prices on this page. Each row comes from the vendor’s own page as it stood on September 28, 2026. Vendors change prices often, so confirm the figure on the linked page before you budget.

A worked example with stated assumptions

This example is generic. It uses the list price table and the formulas above, and every assumption is written out so you can replace it with your own.

The company: a manufacturer with 20 office staff who would use an AI workspace, and one job to automate: reading supplier invoices and matching them to purchase orders, as AI invoice processing describes.

1. Seats. Assume all 20 office staff get a Standard seat on ChatGPT Business or Claude Team. The formula gives 20 × US$20 to US$25, or US$400 to US$500 a month before tax.
2. API usage for the invoice job. Assume 500 invoices a month, each using about 10,000 input tokens and 1,000 output tokens. That is 5 million input tokens and 0.5 million output tokens a month. On the Claude models in the table, the formula gives (5 × US$1 + 0.5 × US$5) to (5 × US$10 + 0.5 × US$50), or US$7.50 to US$75 a month, before any caching or batch discount.
3. A local model instead of the API. If the invoices must never leave your control, assume one rented GPU server in Toronto left on all month. The formula gives 730 hours × US$0.76 to US$4.41, or about US$555 to US$3,219 a month, depending on the GPU the model needs.
4. Staff time. Assume 40 to 80 hours of setup across the owner of the invoice job and IT, including the test set and training, and 2 to 4 hours a week of review and upkeep afterwards. These hours are placeholders for your own estimate. Multiply them by each person’s loaded hourly cost.

In this example, the usage fees for the invoice job are the smallest cash cost, well below the seats. How to build an AI agent makes the same point: model usage is usually the smallest line in an agent’s budget. Your own figures depend on your document sizes, your volume and the model your tests show you need.

Costs that are easy to miss

• Currency. Many AI prices are in US dollars. OpenAI says its listed Business prices “are in USD and may vary by country and currency” (OpenAI), Anthropic lists US prices for the Team plan, and DigitalOcean says it bills “exclusively in USD.” Convert at the rate your bank charges you.
• Tax. Anthropic and Google list prices that exclude tax. Add the sales tax your company pays on these services.
• Minimums and commitments. Minimum seat counts and annual terms set a floor under the bill, even in a month when few people use the tool.
• Usage past the seat. Claude Team sells usage credits on top of the seat, and Claude Enterprise bills all usage separately. Read how your plan handles a heavy month before it arrives.
• Idle servers. A rented GPU server bills while it sits idle, and on DigitalOcean even while powered off, until you delete it.
• Model changes. Providers retire models on a published schedule, and moving to a new one takes a round of testing. Budget staff time for that testing.

Funding programs can share part of the build cost for some projects. How to build an AI agent lists the programs that apply to manufacturers and distributors in Canada.

How ThriveAI helps

ThriveAI is an AI engineering company in Ottawa. It builds private AI systems on the client’s own data for manufacturers and distributors in Ontario and Quebec. Derik Lawlis, the founder, leads every project and stays close to the build.

The platform is designed to keep each client’s data on its own server in Canada. You choose the model: one that runs on that server, or a hosted model under a written zero data retention agreement, under which the provider keeps no copy of a request or its answer. A hosted model may process requests outside Canada, so the contract names the model. A named person at your company approves every action before anything is sent or saved. Enterprise AI platform shows how the pieces fit together, and About ThriveAI covers the company.