What AI Agent Tools Actually Cost a Small Business

Most quotes for AI work open with a monthly number. That number is usually the least interesting thing in the document. The model itself, the actual AI doing the thinking, often costs a few dollars a month for a small shop. Everything else in the quote is the plumbing, the review time, the privacy work, and the parts nobody scoped.

Here's what the vendors actually publish, checked against their own pricing pages in August 2026, and where the money really goes.

What the model itself costs

AI tools bill by the token. A token is roughly three quarters of an English word, and you pay on both the text you send in and the text that comes back.

Published list prices per million tokens, as of August 2026 (input / output):

• Claude Haiku 4.5: $1 / $5. Sonnet 5 is $2 / $10, Opus 5 is $5 / $25 (Anthropic pricing docs)
• gpt-5.6-luna: $0.20 / $1.20. The mid tier, gpt-5.6-terra, is $2.00 / $12.00, and gpt-5.6-sol is $5.00 / $30.00 (OpenAI API pricing)
• Gemini 3.5 Flash-Lite: $0.30 / $2.50, with a free tier for low volume. Gemini 3.6 Flash is $1.50 / $7.50 (Gemini API pricing)

Per million tokens is hard to feel. So here's a worked example straight from Anthropic's own pricing page: handling 10,000 customer support conversations, averaging about 3,700 tokens each, on Claude Haiku 4.5, works out to roughly $37 total.

Ten thousand conversations. Thirty seven dollars.

That is genuinely where inference sits in 2026. If your shop fields 300 inquiries a month, the model bill is rounding error. Which also means that if someone quotes you two thousand dollars a month, almost none of it is the AI.

The per-use extras that scale with traffic

Token pricing is the predictable part. Tools are where it gets lumpy, because search, file retrieval and code execution bill separately.

Anthropic charges $10 per 1,000 web searches on top of tokens. OpenAI publishes $10.00 per 1,000 web search calls, $2.50 per 1,000 file search calls, and $0.10 per gigabyte per day just to keep those files sitting there. Google's grounding with Search gives you 5,000 free requests a month on Gemini 3.x models, then charges $14 per 1,000.

An agent that searches the web on every single customer question has a very different cost shape from one that answers out of your own documents. At 2,000 searches a month you're looking at roughly $20 to $28 in search fees, which is nothing on its own, but it climbs with traffic at the same time your token bill does.

Running agents can also bill for wall clock time. Anthropic's Managed Agents adds $0.08 per session hour on top of tokens, charged only while the session is actually working.

Two levers cut all of this hard, and they're worth asking your vendor about directly. Cached input costs 10% of the normal input price. Batch processing, for anything that doesn't need an instant answer, is 50% off both input and output.

The costs that don't appear on any pricing page

Integration is the real invoice. The model can draft a great reply. Getting that reply into your booking software, your CRM and your invoicing tool, and keeping it there when one of those vendors changes an API, is the actual work. This is where the hours go, and it's the line item most proposals bury inside a vague setup fee.

Licence stacking. Seat based AI is usually an add-on, not a standalone product. Microsoft's own licensing documentation is explicit that Microsoft 365 Copilot is an add-on plan requiring one of a list of qualifying Microsoft 365, Office 365, Teams or Exchange plans first (Microsoft Learn). If your team sits on a plan that doesn't qualify, the base upgrade comes first, per seat, every month, indefinitely.

Human review. Every AI system that touches a customer needs someone checking a sample of the output, especially in the first eight weeks. That's staff time. It's real, it's recurring, and pretending it away is the fastest route to a bad automation story.

Privacy and governance. In Canada this isn't optional. The federal, provincial and territorial privacy commissioners published joint principles for generative AI covering lawful authority for the personal information you feed a model, meaningful consent, transparency when AI is involved in a decision, access and correction rights, and safeguards against threats like prompt injection and model inversion (OPC). For a Vancouver or New Westminster business holding client records, that means a written policy, a real answer to "where does my data go," and a named person accountable. You build it once and maintain it forever.

Training and process change. Statistics Canada found that 44.4% of businesses using AI had already changed their training or staffing practices because of it, with 32.0% providing AI training to existing employees (StatCan, Q2 2026). Nobody puts that in a quote, and everybody pays it.

Model churn. Names and prices move faster than budgets. Claude Sonnet 5's introductory rate of $2 / $10 was made permanent rather than rising as originally scheduled, which is a rare price drop in the right direction. But Anthropic also notes that Claude 4.7 and later models use a newer tokenizer producing roughly 30% more tokens for the same text, so identical work can cost differently on a newer model even at the same headline rate. Budget for a re-pricing conversation once a year.

Where this doesn't apply

Cost is not actually the main thing stopping small businesses, and the data is fairly blunt about it. In the second quarter of 2026, 19.2% of Canadian businesses reported using AI to produce goods or deliver services, up from 6.1% two years earlier. Only 10.6% named cost as an obstacle, rising to 15.1% among firms with 20 to 99 employees. Cybersecurity and privacy concerns ranked higher, at 13.4%. And 41.4% of businesses with one to four employees said AI simply wasn't relevant to what they do (StatCan, Q2 2026).

That last figure deserves respect rather than a counter-argument. Plenty of small operations genuinely don't have a repetitive, high volume, text shaped process worth automating.

The productivity case is also weaker than the marketing suggests. A Statistics Canada study by Jiang Li and Huju Liu found that once you adjust for complementary capabilities such as R&D, cloud computing, data analytics and ICT training, the association between AI adoption and productivity falls to 5.1% and becomes statistically insignificant, with no significant relationship to short term productivity growth (StatCan, April 2026). The honest reading is that firms which got gains were already investing in digital capability. AI rode along with the rest.

So hold off if your process changes every week, if the task needs judgment more than throughput, if your data lives in three incompatible places nobody has cleaned, or if you can't spare anyone to review output for the first two months. In those cases the tokens are cheap and the project still fails.

And note what the adoption numbers don't tell you: the survey measures whether businesses use AI, not whether it paid off. Nobody has published good return figures for shops your size. Treat any specific ROI promise, including a vendor's, as an estimate rather than a measurement.

A sane way to budget

Assume the model is close to free. Assume integration is most of the build. Assume review time is permanent and smaller each month. Ask your vendor to break the quote into build, licences, per-use fees and ongoing support, and ask which of those four goes up when your volume doubles. If they can't answer that, the number isn't a price. It's a guess.

Sources

• Anthropic. *Pricing* (Claude Developer Platform documentation), 2026. https://platform.claude.com/docs/en/about-claude/pricing
• OpenAI. *API Pricing*, 2026. https://developers.openai.com/api/docs/pricing
• Google. *Gemini API pricing*, 2026. https://ai.google.dev/gemini-api/docs/pricing
• Microsoft. *License Options for Microsoft 365 Copilot*, updated 2026. https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-licensing
• Statistics Canada. *Analysis on artificial intelligence use by businesses in Canada, second quarter of 2026*, June 11, 2026. https://www150.statcan.gc.ca/n1/pub/11-621-m/11-621-m2026010-eng.htm
• Li, Jiang and Huju Liu. *Artificial intelligence adoption and productivity in Canadian firms*, Statistics Canada, April 22, 2026. https://www150.statcan.gc.ca/n1/pub/36-28-0001/2026004/article/00002-eng.htm
• Office of the Privacy Commissioner of Canada. *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/

If you want a straight answer on what an AI setup would actually cost your business, with the build, licences and per-use fees separated out, book a free call with Autana Solutions. We're in Burnaby, we work with small and mid-sized businesses across Metro Vancouver, and we'll tell you plainly if the numbers don't justify the project.