Here's a number that surprises most business owners: running an AI assistant over roughly 10,000 customer support conversations costs about $37 in model usage. That's Anthropic's own worked example for its Haiku model, from its pricing documentation.
So why do AI projects still stall on cost? Because model usage is the smallest of three very different cost layers — and most articles about "AI pricing" only talk about the cheap one. This post walks through all three with sourced, current numbers, and is honest about the layer where no reliable published numbers exist.
All prices below were checked against the vendors' own pricing pages on August 15, 2026. AI pricing changes often — Google has already scheduled an increase for January 2027 — so treat the linked pages as the source of truth.
Layer 1: Model usage (the cheap layer)
Every AI feature ultimately calls a model and pays per token — roughly, per word processed. Current pricing for the tiers small-business workloads actually use:
| Provider | Sensible default model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|---|
| Anthropic | Claude Haiku 4.5 | $1.00 | $5.00 |
| OpenAI | GPT-5 mini | $0.25 | $2.00 |
| Gemini 2.5 Flash | $0.30 | $2.50 |
For context: a typical support conversation runs a few thousand tokens. At these rates, a small business handling hundreds of enquiries a month is spending single-digit dollars on model usage. Document processing — extracting data from invoices, summarizing contracts — lands in the same range. Techniques like prompt caching and batch processing cut these numbers by 50–90% for repetitive workloads.
The takeaway: model usage is almost never the reason an AI project is expensive.
Layer 2: Platform subscriptions (the predictable layer)
Most small businesses don't call model APIs directly — they use tools that package them. Current pricing, from the vendors' own pages:
| Tool | What it does | Current pricing |
|---|---|---|
| Zapier | Workflow automation + AI steps | Free tier (100 tasks/mo); Professional from ~$20/mo |
| Make | Workflow automation | Free tier; paid from $9–29/mo |
| n8n | Automation, self-hostable | Cloud from €20/mo; self-hosted from a ~$10 VPS |
| Chatbase | Website AI chatbot | Free tier; paid $32–400/mo |
| Tidio + Lyro AI | Chat + AI agent | From ~$24/mo + Lyro from ~$33/mo |
| Intercom Fin | AI support agent | Seats from $29/mo plus $0.99 per resolved conversation |
A realistic small-business AI stack — one automation platform plus one AI chatbot — runs $50–150/month at current list prices. Note the newer pricing models: Intercom charges per resolution, which scales with success rather than seats. Read the metering carefully before committing; this is where surprise bills come from.
If you're choosing between the automation platforms, we've written a detailed comparison of Zapier, Make, and n8n.
Layer 3: Implementation (the honest layer)
Here's what most articles won't tell you: there is no reliable published benchmark for what AI implementation costs a small business. We looked. The "$5,000–$50,000 average" figures you'll find on Google trace back to content-farm sites with no methodology, no survey, and no named researchers. Treat every one of them as marketing copy.
What we can tell you from doing this work is what drives the cost, because that's what an honest quote is built from:
- How many systems the AI has to touch. A chatbot that only reads your website is a small job. An assistant that reads your CRM, writes to your ticketing system, and triggers your invoicing is a different class of work — each connection is scoping, authentication, testing, and failure handling.
- The state of your data. AI that answers from your documentation is only as good as the documentation. If the real project is "first, clean up five years of inconsistent records," that's where the budget goes — and no vendor demo mentions it.
- How wrong the AI is allowed to be. A drafting assistant a human reviews is cheap to ship. Anything that acts autonomously — sends the email, issues the refund — needs guardrails, logging, and escalation paths. The cost lives in the last 10% of reliability, not the first 90%.
- Maintenance. Models get deprecated, prices change, APIs move. Budget ongoing attention, not just a build.
Working on this for your business? Book a free 30-min strategy call →
What other small businesses are actually doing
Survey data with real methodology, for calibration:
- 50% of US small-business workers now use AI at work, and 58% of those use it regularly — but only 6% use it to automate workflows with minimal human involvement. Most usage is still personal productivity: drafting, summarizing, brainstorming. (US Chamber of Commerce Foundation / Ipsos, May 2026, n=1,070)
- 38% of small businesses report actively using AI across multiple functions — led by marketing, written communication, and customer-service assistants. (Verizon Business / Morning Consult, 2026, n=600)
- Adoption scales with company size: 43% at 2–9 employee firms vs 59% at 100–249. The gap is capacity, not desire — and it's exactly the gap an implementation partner closes.
- The most-cited barrier isn't cost — it's privacy and security concerns (47%). Legitimate, and addressable with scoping: what data the model sees, where it's processed, what's logged.
The picture: adoption is mainstream, but workflow-level automation — the kind that actually returns hours — is still rare. That's the opportunity, and it's also why "we tried ChatGPT and it didn't stick" is so common: using AI and integrating AI are different projects.
When AI integration isn't worth it (yet)
We tell prospects this regularly, so it belongs here too:
- Low volume. If the task takes two hours a month, automate nothing. The setup will never pay back.
- No stable process. AI integration automates a process. If the process changes weekly, fix that first.
- The data isn't there. If the answers you want the AI to give don't exist in writing anywhere, the first project is documentation, not AI.
For everything else, the math is checkable: our free automation ROI calculator turns your hours, rates, and tool costs into a payback period — with the assumptions stated, not hidden.
If the numbers look good and you want the realistic version for your specific stack, that's what we do.