What AI Actually Costs a Small Business in 2026

Short version: AI costs a small business money in three places. A chat assistant for each person who uses one runs $20 to $25 a person per month on the business plans of ChatGPT or Claude, as of October 2026. AI features inside software you already pay for are sometimes included and sometimes an add-on. And AI built into your own software is billed per use — for ordinary paperwork, a fraction of a cent per document. The cost that shows up on none of those bills is the time someone spends checking the AI's work.

Layer 1: a chat assistant for each person

This is what most people mean by "using AI": someone opens ChatGPT, Claude, Gemini, or Copilot to draft an email or summarize a contract. You pay per person, per month. Prices as of October 2026, from each vendor's own pricing page:

• ChatGPT. Plus is $20 a month for one person. Business, for teams of 2 to 200, is $25 per user per month billed monthly or $20 billed annually, for a standard seat. (ChatGPT pricing)
• Claude. Pro is $20 a month, or $17 a month billed annually. Team, for teams of 2 to 150, is $25 per seat per month billed monthly or $20 billed annually, for a standard seat. (Claude pricing)
• Gemini. It comes inside Google Workspace. Business Starter is $7 per user per month on a one-year commitment ($8.40 month to month) and includes Gemini in Gmail and the Gemini app. Business Standard is $14 ($16.80 month to month) and adds Gemini in Docs, Meet, and more. (Google Workspace pricing)
• Microsoft 365 Copilot Business. $21 per user per month, paid yearly, on top of a qualifying Microsoft 365 business plan. (Microsoft Copilot pricing)

A five-person shop on ChatGPT Business or Claude Team pays $100 to $125 a month, depending on billing. Both have free tiers, and both say on their pricing pages that the business plans don't train on your business content by default. More on that in whether it's safe to put customer data into ChatGPT or Claude.

My advice: don't buy seats for everyone on day one. Buy one or two for the people who write the most, and see whether they're still using it in sixty days. A seat nobody opens still costs $25 a month.

Layer 2: AI inside tools you already pay for

This layer shows up without anyone deciding to "adopt AI." The software you already use for email, accounting, and scheduling keeps adding AI features. Sometimes they're folded into your plan, the way Google now includes Gemini in Workspace. Sometimes they're a paid add-on per user, the way Microsoft sells Copilot. Before you pay for a separate AI tool, check whether something you already pay for does the job. And when a renewal notice mentions AI, read what you're actually being charged for.

Layer 3: AI inside your own software, billed per use

This is the layer we build. When your software uses AI to read a receipt or draft a reply, it calls an AI model directly — through what's called an API — and pays only for what it uses. The bill is counted in tokens.

What a token is

A token is a small chunk of text — roughly three-quarters of an English word, per Anthropic's pricing page. You pay for the tokens you send in (your instructions plus the document) and, at a higher rate, for the tokens the model writes back. Pictures count too: Claude cuts an image into 28-by-28-pixel squares, and each square is one token. Big photos are shrunk first, which caps the cost — on Claude Haiku 4.5, one image tops out at 1,568 tokens (Anthropic's vision documentation).

What the models cost

Anthropic's API prices per million tokens, as of October 2026 (Anthropic's API pricing):

• Claude Haiku 4.5: $1 in, $5 out
• Claude Sonnet 5.5: $2 in, $10 out
• Claude Opus 5.5: $4 in, $20 out

Worked example: a shop that reads 300 receipts a month

Say your shop photographs 300 receipts and vendor invoices a month, and the software reads each one into an expense line — vendor, date, tax, total, line items — for someone to check. Assumptions, set on the generous side:

• The photo: 1,568 tokens, the most Claude Haiku 4.5 counts for one image.
• The instructions: 500 tokens telling the model what to pull out.
• The answer: 400 tokens of extracted fields.

The math on Claude Haiku 4.5:

• In: 1,568 + 500 = 2,068 tokens × $1 per million = $0.002068
• Out: 400 tokens × $5 per million = $0.002000
• Per receipt: $0.004068 — about four-tenths of a cent
• Per month: 300 × $0.004068 = $1.22

If my assumptions are off by a factor of three, it's still under $4.

Which model is good enough is something you test on your own documents, not something you assume. In my own system, the low-cost model wasn't good enough to read numbers on its own. If testing says you need the stronger one: Claude Sonnet 5.5 charges twice Haiku's rate, looks at images in finer detail (up to 4,784 tokens per image), and, like other newer Claude models, counts the same text as roughly 30% more tokens. Worst case, that's 4,784 + 650 = 5,434 tokens in × $2 per million ($0.010868) plus 520 tokens out × $10 per million ($0.0052) — about $0.016 a receipt, or $4.82 for 300. Still small, but about four times the bill. At higher volumes, how you set up the job matters more than the sticker price.

Real numbers from a system I run

I built WillIFit, a site that tells RV and truck drivers whether their vehicle clears a parking garage. It verifies clearance heights by pulling street-level photos of garage entrances and having Claude look for the clearance sign. It's my own product, so I can show you the bill.

• In August, a full verification sweep of 1,050 parking garages across 87 cities cost $51.73 in AI and image fees — about 5 cents a garage.
• An earlier audit of 1,545 scans, which cost $103, found that 61% of the spend bought the answer "no sign found." I was paying full price to look at photos with no sign in them.
• The fix was to split the job. A low-cost model only answers "is there a clearance sign in this photo?" A stronger model reads the sign only when there is one. In a pilot, that cut the cost per garage from about 18 cents to about 7 cents.

The waste wasn't the price per token. It was the model doing work that didn't need doing. And the low-cost model was low-cost for a reason: when I let it read the numbers by itself in a small test, it got 4 of 11 wrong. That story and the guardrails it led to are in AI reads the paperwork, you approve it.

How we bill AI usage

Our standard rates, adjusted to each project and confirmed in a written quote before work starts:

• Ordinary usage is included. The monthly fees cover hosting, monitoring, backups, security updates, bug fixes, and email support, and they include third-party services — AI usage among them — at ordinary levels.
• Heavy usage comes to you first. If usage materially exceeds the baseline — large-scale AI processing, for example — we tell you first, then either pass the extra through at cost or agree on new pricing in writing.
• Spending limits. Our client portal meters and caps the AI it runs for each client; AI inside your software runs under a spending limit stated in your quote, so a stuck job can't quietly run up a bill.
• Text messaging is billed separately.

Each staff login is $49 a month on top of the platform fee, so a five-person shop comes to $650 + 5 × $49 = $895 a month. Larger builds may carry a one-time build fee, quoted up front. The rest is in what custom software costs a small business.

The cost nobody puts on the invoice

Review time. Every receipt the AI reads still needs a person to look it over: photo on one side, extracted numbers on the other, approve or fix. Say that takes 20 seconds. For 300 receipts, that's 100 minutes a month — at $20 an hour, about $33 of someone's time against $1.22 of AI. It's worth paying; the real question is whether 20 seconds of checking beats typing each receipt in by hand. Good software keeps the check short: the source document next to each number, flags on anything odd, one click to approve.

Wrong answers. A receipt total read as $1,240 instead of $12.40 is a quick fix if someone catches it and a bookkeeping problem if nobody does. That's why we build AI to draft and a person to approve, and why the software, not the AI, does the arithmetic. If an AI feature touches money or customers and nobody reviews its output, the per-token price is the least important number in the deal.

Still working out where AI belongs in your business at all? Start with our practical guide to AI for small business and the list of 12 jobs AI can do for a small business today, and 4 it shouldn't.

Frequently asked questions