What AI Actually Costs a Small Business Per Month

Most AI quotes hide the same thing. You get one monthly number, and underneath it two completely different pricing models are running. One charges by the word. The other charges by the person. If you don't know which one you're buying, you can't tell whether your bill doubles next quarter or stays flat forever.

So here's the real AI cost for small business, with published numbers from the vendor pages themselves. Note that everything below is in USD on the pages I read, so Canadian buyers pay the exchange rate on top.

Token pricing: you pay by the word, roughly

API pricing is metered per million tokens, split between input (what you send) and output (what the model writes back). A token is about four characters, or roughly 0.75 English words, according to Anthropic's pricing documentation.

As of August 2026, Anthropic's published rates are $1 per million input tokens and $5 per million output for Claude Haiku 4.5, $2 and $10 for Claude Sonnet 5, and $5 and $25 for Claude Opus 5 (Anthropic pricing docs). As of the same date, OpenAI publishes $0.20 input and $1.20 output per million for gpt-5.6-luna, $2 and $12 for gpt-5.6-terra, and $5 and $30 for gpt-5.6-sol (OpenAI API pricing). Both vendors discount asynchronous batch work by about 50%.

Those figures stay abstract until you attach them to actual work. Anthropic's docs give a worked example: a customer support conversation averages about 3,700 tokens, and processing 10,000 support tickets on Claude Haiku 4.5 costs roughly $37 (Anthropic pricing docs). That's about $0.0037 per ticket.

Now scale it to a real shop. Say you handle 400 customer conversations a month. On Haiku 4.5 that's around $1.48 in tokens. Sonnet 5 is priced at exactly double Haiku on both input and output, so about $3. Opus 5 is five times Haiku, so about $7.40. These are illustrative: I'm scaling Anthropic's own published per-ticket figure by their own published price ratios, and your real number moves with how long your conversations actually run.

Read that again, because it's the part nobody says out loud. The raw model cost for a small business answering a few hundred conversations a month is often under ten dollars.

Seat pricing: you pay by the person

Then there's the other model, and for most small teams it's the bigger line by far.

Microsoft 365 Copilot Business is listed at $18 per user per month on an annual commitment as of August 2026, discounted from $21 through September 30, 2026, or $25.20 if you pay month to month. Microsoft 365 Business Standard with Copilot runs $23.50 per user per month annually. Every one of these requires a separate qualifying Microsoft 365 licence underneath (Microsoft 365 Copilot pricing).

Five staff on the Copilot Business add-on is $90 a month before the base licences. Compare that with the roughly $3 of tokens above for a few hundred conversations.

Seat pricing is predictable, and that's its genuine virtue. Finance departments like it. But it scales with headcount rather than with work done, so if two of your five people barely open the thing, you pay full freight for them anyway. Token pricing scales with usage, which means it's nearly free when you're small and it climbs when you succeed. Neither is a trick. They're just different bets.

The costs nobody puts on the quote

This is where budgets break. Every item here comes off a vendor page, not a guess:

• Retries and reruns. A wrong answer costs you the tokens twice, and nobody quotes a failure rate.
• Conversation history. Long threads resend the whole context every turn. Prompt caching fixes most of it: a cache hit costs 10% of the standard input price, though the write costs 1.25x or 2x depending on cache duration (Anthropic pricing docs).
• Server-side tools. Web search on the Claude API is billed at $10 per 1,000 searches on top of tokens, as of August 2026. Web fetch carries no extra charge beyond tokens.
• Tokenizer changes. Anthropic notes that Claude 4.7 and later use a newer tokenizer producing approximately 30% more tokens for the same text. Same words, more billable units. Upgrading your model can raise your bill even at an unchanged sticker price.
• Geography premiums. Pinning inference to the US costs a 1.1x multiplier on all token categories, and regional endpoints on Bedrock or Google Cloud carry a 10% premium over global ones (Anthropic pricing docs). Data residency isn't free.
• Runtime, not just tokens. Claude Managed Agents bill session runtime at $0.08 per session-hour separately from token usage.
• Everything that isn't the model. Telephony minutes, CRM seats, calendar integrations and hosting are billed by other vendors entirely. They're often the largest recurring cost in an AI phone setup, and they never show up in a per-token comparison.

The human cost is the one that actually bites

Statistics Canada surveyed businesses from April 1 to May 6, 2026 and found 19.2% had used AI to produce goods or deliver services over the previous 12 months, up from 6.1% in the second quarter of 2024. Among AI-using businesses, 44.4% made training or staffing changes, and 32.0% trained existing employees. For the smallest firms, those with one to four employees, 24.0% trained employees and 8.2% hired AI-skilled workers (Statistics Canada, Q2 2026).

Training time is a real cost. It just lands on your payroll instead of your software invoice.

There's a compliance line too. Canada's federal, provincial and territorial privacy authorities published nine principles for generative AI, covering legal authority and consent, appropriate purposes, necessity and proportionality, openness, accountability, individual access, limiting collection, accuracy and safeguards. Organizations are expected to maintain a clearly defined internal governance structure, tell affected parties when an AI tool is used in decisions about them, and set retention schedules for personal information (Office of the Privacy Commissioner of Canada, 2023). If you're handling customer records in New Westminster or anywhere else in Metro Vancouver, that's policy work someone has to do.

Where this doesn't apply

Plenty of businesses shouldn't buy AI this year, and the evidence backs them up.

A Statistics Canada study by Jiang Li and Huju Liu, released April 22, 2026 and forthcoming in Canadian Public Policy, found AI adopters showed 16.8% higher productivity than non-adopters. That premium fell to 10.2% once pre-adoption productivity was controlled for, then to 5.1% and statistical insignificance once complementary capabilities like R&D, cloud computing, data analytics and ICT training were added. The authors conclude there's no statistically significant direct association between AI adoption and productivity in the short term, and that AI adoption in isolation is likely insufficient to deliver transformative productivity gains (Li and Liu, 2026).

That's a finding worth sitting with. The firms that got faster were mostly already good at digital systems.

The survey data points the same direction. Two in five Canadian businesses, 40.0%, said AI simply isn't relevant to what they do. Cost was a barrier for 10.6%, and cybersecurity or privacy concerns for 13.4% (Statistics Canada, Q2 2026).

Practically: if you take twenty calls a month, setup and seat fees will dwarf any savings. If your booking process lives in someone's head instead of a system, automate the system first. If your data is messy, AI will confidently repeat the mess.

What to ask before you sign

Ask which pricing model you're on. Ask what happens to the bill if volume triples. Ask whether prompt caching and batch discounts are being used on your workload. Ask what the setup fee covers and what it doesn't. Then ask for a small pilot with real numbers instead of a projection.

Sources

• Anthropic. *Pricing* (Claude platform documentation), 2026. https://platform.claude.com/docs/en/about-claude/pricing
• OpenAI. *API Pricing* (developer documentation), 2026. https://developers.openai.com/api/docs/pricing
• Microsoft. *Microsoft 365 Copilot Plans and Pricing*, 2026. https://www.microsoft.com/en-us/microsoft-365-copilot/pricing
• Statistics Canada. *Analysis on artificial intelligence use by businesses in Canada, second quarter of 2026*, released June 11, 2026. https://www150.statcan.gc.ca/n1/pub/11-621-m/11-621-m2026010-eng.htm
• Li, Jiang, and Huju Liu. *The Role of Complementary Capabilities in AI Adoption and Productivity: Firm-Level Evidence from Canada*, Statistics Canada, released April 22, 2026, forthcoming in Canadian Public Policy. https://www150.statcan.gc.ca/n1/pub/36-28-0001/2026004/article/00002-eng.htm
• Office of the Privacy Commissioner of Canada and provincial privacy authorities. *Principles for responsible, trustworthy and privacy-protective generative AI technologies*, December 7, 2023. https://www.priv.gc.ca/en/privacy-topics/technology/artificial-intelligence/gd_principles_ai/

Want a straight answer for your own numbers? Autana Solutions builds AI employees for businesses around Burnaby and Metro Vancouver, and we'll walk through your actual call volume, seat count and integration costs before anyone talks about a monthly fee. Book a free call and we'll tell you honestly if the math doesn't work yet.