Cost Control: What This Stack Should Cost You

The bill for a small-business AI stack rarely arrives as one number. It arrives as four subscriptions nobody remembers signing up for, one of which quietly bills per unit of work rather than per person. Every figure below was read off the vendor's own pricing page on 31 July 2026. Check the date before you trust any of it, because two of these prices already have expiry dates published.

Three meters run at once and they bill on different things

Per seat is the one everybody understands. You pay a fixed amount per person per month, and a person who never logs in costs the same as one who lives in the tool. Assistants and office suites bill this way.

Per unit of work is the one that catches people. Automation platforms count runs, and the definitions differ enough to change your bill by a factor of ten. Zapier's pricing page defines its unit as follows: "A task is counted whenever Zapier successfully completes a unit of work for you. Failed actions are not counted." Triggers, polling and built-in tools such as Formatter and Filter are excluded from the count. n8n counts differently, and its pricing page says so directly: "An execution is a single run of your entire workflow." Step count and data volume do not change it.

So a twelve-step workflow that runs a thousand times a month is one thousand executions on n8n and can be several thousand tasks on Zapier, depending on how many of those steps are billable actions. That is not a criticism of either. It is the single most important thing to work out before you commit to one.

Per token is the third meter, and it only applies if someone built you something that calls a model directly. Anthropic's platform documentation lists Claude Sonnet 5 at $2 per million input tokens and $10 per million output tokens under introductory pricing, Haiku 4.5 at $1 and $5, and a Batch API discount of 50 percent on both input and output. It also publishes a worked example for a support-agent workload: roughly 3,700 tokens per conversation on Haiku 4.5 works out at about $37.00 per 10,000 tickets.

Assistant seats, read off the pricing pages on 31 July 2026

Claude publishes its numbers plainly. Pro is listed as "$17 Per month with annual subscription discount ($200 billed up front). $20 if billed monthly." Team Standard is "$20 Per seat / month if billed annually. $25 if billed monthly", described as being for teams of 2 to 150. Team Premium is $100 per seat per month annually, or $125 monthly. The page also carries its own caveat: price and plans are subject to change at Anthropic's discretion.

ChatGPT is the awkward one to cite, and the reason is worth stating rather than papering over. OpenAI's pricing page at openai.com/chatgpt/pricing lists the plan names Free, Go, Plus, Pro, Business and Enterprise, but the fetched page text contains no dollar amounts next to any of them. The prices load client-side behind the buttons. The one figure OpenAI states in plain text is in its help centre article on ChatGPT Plus, which gives "$20/month (billed monthly)". For Business and Enterprise seat prices you will have to start a checkout or talk to sales, which is a deliberate choice on OpenAI's part and one you should factor into how easily you can budget.

Microsoft 365 Copilot has a prerequisite that is easy to miss and expensive to discover late. Microsoft's business page states: "A Microsoft 365 Business plan is required to purchase Microsoft 365 Copilot Business." The add-on itself is listed at "$18.00 user/month, paid yearly" as promotional pricing, or "$25.20 user/month" on a monthly commitment. Bundled alternatives exist, with Business Standard with Copilot at $32.00 per user per month annually for the Premium tier and $23.50 for the Standard tier.

Google Workspace bundles rather than adds on. Its pricing page lists Business Starter, Standard and Plus with Gemini AI features included in every tier rather than sold separately.

The price you see depends on where you are billed from

This matters more than it sounds, and it is why a price you read in a blog post from another country is not a number you can budget against.

Fetched from this location on 31 July 2026, Google Workspace, Notion, n8n and Grammarly all served euro pricing, while Claude, Microsoft and Zapier served dollars. Workspace listed Business Starter at 3.40 euro per user per month on an annual commitment (6.80 euro flexible), Standard at 6.80 euro (13.60 euro flexible), and Plus at 10.55 euro (21.10 euro flexible). Notion listed Plus at 9.50 euro and Business at 19.50 euro per seat per month, with the Notion Agent, Meeting Notes and Enterprise Search sitting in the Business tier. n8n listed Starter at 20 euro per month billed annually for 2,500 executions, Pro at 50 euro for 10,000, and Business at 667 euro for 40,000.

Those are the same pages that show dollar figures to a US visitor, and the conversion is rarely a straight exchange-rate calculation. Price your stack from the pages as they render for your own billing country, in your own currency, and ignore every third-party summary including this one for the final number.

Two dated price cliffs already on the calendar

Most vendor price changes arrive without warning. Two current ones are published with dates, which makes them plannable.

Microsoft's $18.00 per user per month figure is promotional pricing tied to an annual commitment, described as applying through September 30, 2026 and to the first year only, with automatic renewal at standard rates afterwards. If you sign up in September you are budgeting for a renewal at a higher number.

Anthropic's Claude Sonnet 5 API pricing carries an explicit note: introductory pricing of $2 and $10 per million input and output tokens is in effect through August 31, 2026, after which standard pricing of $3 and $15 per million takes effect. That is a 50 percent increase on both sides, dated, on a page you can read today. Anyone running a Sonnet 5 workload should be modelling September's bill now.

The overage rules that turn a small bill into a large one

Per-seat plans cannot surprise you. Usage plans can, and Zapier publishes exactly how far the surprise can go. Exceeding your limit with pay-per-task enabled bills overages at 2.5 times the base rate on monthly plans and 1.25 times on annual plans. There is a ceiling, stated on the page: "Usage stops once it reaches 3x your task subscription, until the next cycle."

Read that as your genuine worst case. A $29.99 plan with pay-per-task on can bill roughly three times its subscription before it stops, and then the automation stops running, which for a business relying on it is the more expensive half of the problem. Decide in advance which failure you prefer, and set the toggle accordingly rather than discovering the answer during a busy month.

A defensible starting budget, and where to cut first

For a three-person business, one paid assistant seat for the owner plus free tiers for everyone else is a legitimate opening position. On the numbers above that is roughly $17 to $20 a month, not the several hundred a stack diagram implies. Add the office suite you were already paying for, where Google now includes Gemini in every Workspace tier rather than charging separately for it.

Add automation only once a workflow has run manually for a month and you can name how many times it fired. Zapier's 100 free tasks per month exists precisely for that measurement, and it is enough to answer the question without paying anything.

When it is time to cut, three checks find the waste faster than a spreadsheet. Look for seats assigned to people who have not opened the tool in thirty days, which is the single biggest overspend in per-seat software. Look for two products doing one job, which happens naturally when a suite you already pay for adds a feature you bought separately last year. And check whether you are paying monthly on anything you have used for a year, since the annual rates above are consistently 15 to 30 percent cheaper on the same product.

Then put the September dates in your calendar. The Microsoft promotional rate and the Sonnet 5 introductory rate both change within about five weeks of this page being written, and neither vendor is obliged to remind you.