Using Claude in a small business: what we build on it, and why
Claude comes as four things: a chat plan, a team plan, an enterprise plan, and an API. A small business needs to know which one it is buying, what each does with its data, and when the right answer is a system that runs on the API instead of a person typing into a window. We build on it, and this is the reasoning.
A small business uses Claude in one of four ways: the Pro plan at $17 to $20 a month for one person, the Team plan at $20 to $25 a seat a month with central billing and no training on your content by default, the Enterprise plan with a HIPAA-ready option, or the API, priced per million tokens, for a system that runs a process on its own (Anthropic pricing). Anthropic signs a Business Associate Agreement for the API and HIPAA-activated Enterprise, not for the consumer or Team plans (Anthropic). Our AI Finder runs on the API.
The four surfaces, and what each costs
| Surface | Price | What it is for |
|---|---|---|
| Free | $0 | Chat on web, phone, and desktop with basic features |
| Pro | $17 a month billed annually, $20 monthly | One person: chat, Claude Code, Cowork, Design, and Science |
| Max | From $100 a month | One person with 5 or 20 times Pro usage and priority access |
| Team, Standard seat | $20 a seat a month annually, $25 monthly | A company: central billing, admin controls, SSO, enterprise search |
| Team, Premium seat | $100 a seat a month annually, $125 monthly | The same, with more usage per seat |
| Enterprise | $20 a seat plus usage at API rates | SCIM, audit logs, compliance API, custom retention, HIPAA-ready option |
| API | Per million tokens; see below | A system that calls the model on your data on a trigger |
Plan prices as published on Anthropic's pricing page, checked September 5, 2026.
The first six rows are a person typing. The seventh is a system running. That distinction decides most of what follows.
What each one does with your data
Two facts, and they are different facts. First, Anthropic's pricing page states that there is no model training on your content by default across the Team and Enterprise tiers. Second, a Business Associate Agreement, the contract that makes a vendor responsible for protected health information, is available for two surfaces only: the first-party API on a HIPAA-ready organization, and Enterprise plans where the Primary Owner has activated HIPAA compliance and accepted the agreement (Anthropic). It does not cover Free, Pro, Max, or Team, and Claude Code and Cowork are excluded (HIPAA-ready Enterprise plans).
For a business without patient data, the first fact is the one that matters, and the Team plan has it. For a practice with patient data, only the second fact counts. No training on your content is a setting. A BAA is a contract. The four HIPAA checks walk through the difference.
Which surface for which job
- One person drafting, summarizing, and thinking. Pro. A small firm's owner reworking a proposal or a first draft of a policy is the Pro plan doing its job.
- A team that needs one bill and an admin. Team. Central billing, SSO, and the no-training default on every seat.
- Any workflow that touches protected health information. Enterprise with HIPAA activated, or the API under a BAA. Nothing else.
- A process that should run without a person typing. The API. Intake forms that shortlist, documents that get read and routed, records that get matched. Two small CPA firms told the Journal of Accountancy in August 2026 that they built an intake form with Claude and a client app with Claude Code in an afternoon; that is the API doing a job a person used to do by hand.
What it costs to run a process on the API
The API is priced by usage, per million tokens in and out, and the model chosen sets the rate. As published on Anthropic's pricing page on September 5, 2026: Opus 5 at $5 in and $25 out per million tokens, Sonnet 5 at $2 and $10, Haiku 4.5 at $1 and $5, with batch processing at half price and cached prompt reads at a fraction of the input rate. A million tokens is roughly the text of several long books, so a process that reads a form and writes a short result costs cents per run.
The number to estimate is not the rate. It is the volume: how many runs a week, how much text each one reads, and how much it writes. A week of real counts before go-live turns the per-token price into a monthly figure, and published guides put typical small-business API usage at $100 to $500 a month when a tool calls a model on your data (AIessentials). Choosing the smaller model for the routine step and the larger one for the judgment step is where most of the cost control lives.
See a system running on the API.
The AI Finder asks four questions about the business and the bottleneck, then returns a shortlist of existing tools with the honest catch on each. It runs on Claude through the API, and it is free.
Why we build on it
Our AI Finder runs on Claude through the API. A business answers four questions, the system reads them against a catalog of tools, and it returns a shortlist with the catch on each. That is the shape of most of what we build: a defined input, a model call in the middle, a structured result, and a person deciding what to do with it.
Three reasons it is the model we reach for. The data terms are legible: no training on content by default, and a BAA path through the API when a client needs one. The API is built for the system-running case rather than the person-typing case, with batch pricing and prompt caching for the repeated work a business actually has. And the model handles the judgment-shaped middle step, reading a messy form or a document and producing something structured, well enough that the exception path stays short. We do not publish accuracy figures for the AI Finder, because we have not measured them in a way we would stand behind.
When it is the wrong answer
When the data may not leave the building. Our patient presentation automation for an orthodontic clinic runs entirely on the clinic workstation and sends nothing to any cloud model, because the boundary required it. The same is true of the dental payment integration, where patient identity stays inside the clinic and the matching is local logic, not a model call. A cloud model is the wrong tool for a workflow whose first rule is that the data stays home, and no plan or contract changes that.
It is also the wrong answer for a process nobody has defined. A model in the middle of an undefined process produces confident output from unclear input, and that is the expensive mistake. Define the process, decide what the system may not choose, then pick the model.
How to start
- Name the surface. A person typing, or a system running? Pro or Team for the first; the API for the second.
- Check the data facts. No-training default for ordinary business data; a BAA for patient data; local-first when the data cannot leave.
- Count a week. Runs, input size, output size. That turns the per-token rate into a monthly number.
- Build one process. A trigger, one model call, a structured result, an exception path to a person, and a log.
The plan is a person. The API is a system. Buy the one that matches the job.
What to hold on to
- Four surfaces, two kinds. Free, Pro, Max, Team, and Enterprise are a person typing. The API is a system running.
- Two different data facts. No training on your content by default on Team and Enterprise. A BAA only on the API and HIPAA-activated Enterprise.
- Usage pricing rewards counting. Per million tokens by model, half price in batch, cheaper on cached reads. A week of real counts gives the monthly number.
- We build on it because the terms are legible. Our AI Finder runs on the API. We publish no accuracy figure for it.
- Wrong answer when data stays home. Both of our clinic builds run local with no cloud model, because the boundary required it.
Questions owners ask us
How much does Claude cost for a small business?
Pro is $17 a month billed annually or $20 monthly for one person. Team is $20 a seat a month annually or $25 monthly for a Standard seat, with Premium seats at $100 or $125. Enterprise is $20 a seat plus usage at API rates. The API is priced per million tokens by model: Opus 5 at $5 in and $25 out, Sonnet 5 at $2 and $10, Haiku 4.5 at $1 and $5 (Anthropic pricing, September 5, 2026).
Does Claude train on my business data?
Not by default on the Team and Enterprise plans, according to Anthropic's pricing page. That setting keeps your content out of model training. It is not a Business Associate Agreement, which is a separate contract available only for the API and HIPAA-activated Enterprise plans.
Is Claude HIPAA compliant?
Anthropic signs a BAA for two surfaces: the first-party API on a HIPAA-ready organization, and Enterprise plans where the Primary Owner has activated HIPAA compliance (Anthropic). Free, Pro, Max, and Team are not covered, and Claude Code and Cowork are excluded (HIPAA-ready Enterprise plans). Compliance is then the practice's setup: minimum data, where it lives, and what the tool may not decide.
When should a small business use the Claude API instead of a plan?
When a process should run without a person typing: an intake form that shortlists, documents read and routed, records matched on a schedule. The plans are a person working faster. The API is a system with a trigger, a model call, a structured result, an exception path, and a log. Two CPA firms told the Journal of Accountancy they built exactly that in an afternoon.
What has Prometheus built on Claude?
The AI Finder: a business answers four questions about itself and its bottleneck, the system reads them against a catalog of tools through the API, and it returns a shortlist with the catch on each. Our two clinic builds do not use a cloud model at all, because patient data had to stay on the workstation.
How do I estimate what a Claude API workflow will cost per month?
Count a week of real work: how many runs, how much text each reads, and how much it writes. Multiply by the per-million-token rate for the model you chose, with batch at half price for anything that can wait. Published guides put typical small-business usage at $100 to $500 a month when a tool calls a model on its data (AIessentials).
Sources
Prices and data terms are quoted from Anthropic's own pages on the date checked; the firm examples from the Journal of Accountancy; our own use from the systems we run. Prometheus is an Anthropic customer, not a partner. Checked September 5, 2026.
- Anthropic, Claude pricing. Plan prices, API rates per million tokens, batch and caching discounts, and the no-training-by-default statement.
- Anthropic, "Business Associate Agreements (BAA) for Commercial Customers" and "HIPAA-ready Enterprise plans". Coverage by product and the activation steps.
- Journal of Accountancy, "Real-life ways small firms use AI". August 1, 2026. Agate CPA and High Rock Accounting.
- AIessentials, "AI Consultant Cost (2026)". Updated August 21, 2026. Typical API usage costs.
- AI Finder, patient presentation automation, and dental payment integration, our own systems.