Three different purchases, not one decision
The question always arrives in the same shape: how much does it cost to bring AI into the company? There is no single number because there is no single purchase. There are three, and they are different in kind: a per-person subscription, per-token consumption, and an engineering project. All three are below, with the arithmetic shown. Fixed assumption across every example: a 10-person company, 3 of them technical, list prices in dollars checked in September 2026.
Context helps calibrate the urgency. In Brazil, 52% of small-business owners said they had used AI in the two weeks before the interview, against 21% in the United States — a Sebrae survey of 7,182 respondents, run between 24 March and 8 May 2026. Your competitor is already in band 1. The advantage has moved from adopting to knowing what you pay, why, and what the invoice does not cover.
Band 1 — ready-made subscriptions: $317 a month for 10 people
A Claude Team seat costs $20 per person/month on the annual plan and $25 monthly, for teams of 2 to 150 people. GitHub Copilot Pro costs $10 and includes $15 in monthly credits; Pro+ costs $39 and includes $70 in credits, with each credit worth $0.01. Ten standard seats plus three Pro+ seats for the technical team land at $317 a month. That is the realistic floor of an AI stack in a small company — and the number most people overestimate before sitting down and adding it up.
What this band buys: writing, analysis, long-document reading, research and code, available on day one, with no project. What it does not buy: anything connected to your systems, any per-process record, and any guarantee that people will actually use it. There is also a pricing trap inside the band itself. The same 10 seats on a premium plan cost $1,117 a month, 3.5 times more. That only pays off for teams that demonstrably hit the standard plan's ceiling — and you establish that by looking at usage, not by asking for opinions.
Band 2 — API automation: the cheapest of the three, not the most expensive
Explicit assumptions, so you can redo the maths with your own volumes: 4,000 tickets a month, 4 model calls per ticket, 3,000 input tokens and 400 output tokens per call. That is 48 million input tokens and 6.4 million output tokens a month. With gpt-5-mini ($0.25 input and $2 output per million), the bill comes to $24.80. With Gemini 3.5 Flash-Lite ($0.30 and $2.50), $30.40. With Claude Sonnet 5 ($2 and $10), $160. With Claude Opus 5 ($5 and $25), $400. Same workload, a 16-fold difference, purely from the choice of model.
Two levers cut that bill before any negotiation. Routing: send the 80% of easy cases to the small model and reserve the strong model for the hard 20%; with that split, the bill drops to roughly $52. Prompt caching: on OpenAI, a cached input token costs 10% of the full price. And batching: Anthropic's Message Batches API processes within 24 hours — most batches in under one — at half price, which suits overnight classification and enrichment well.
The gateway is still missing. It centralises keys, caps spend per team, caches, and lets you switch vendors without rewriting code. Cloudflare's AI Gateway ships analytics, caching and rate limiting at no extra cost on any plan, with 10 million logs per gateway on the paid Workers plan; unified credit billing carries a 5% fee and passes inference pricing through with no markup. Adding routed tokens, observability and the gateway, the whole automation lands around $90 a month: $0.02 per ticket. Without routing and caching, the same ticket costs $0.10.
What band 2 buys: throughput. What it does not buy: correctness. Without an evaluation set — 200 hand-labelled cases with the right answer beside each one — you do not know whether swapping models improved or degraded the result, and your customer will tell you first. Building that set costs 3 to 5 days of someone who knows the process from the inside. It is the most expensive item in band 2 and the one almost nobody budgets for.
Band 3 — AI embedded in the product: the invoice misleads
This is where intuition fails badly. Multiply band 2's volume by ten, because now the customer is the user rather than your internal team: routed tokens go to roughly $520 a month. Add $200 of observability and evaluation on a team plan, and $200 of gateway, logs and staging environment. Total: about $920 a month in vendor invoices. Three times band 1, not forty.
Band 3's dominant cost appears nowhere in that sum: it is people and it is risk. Gartner projects that more than 40% of agentic AI projects will be cancelled by the end of 2027, on escalating costs, unclear business value or inadequate risk controls. What band 3 buys is a feature customers pay for. What it does not buy is exemption: logs holding customer data, prompt injection, retention policy and the data-processing agreement all stay yours. The invoice is the smallest line.
| Buys | Doesn't buy | |
|---|---|---|
| Band 1 · subscriptions | Writing, analysis, long-document reading, research and code | Nothing connected to your systems, no per-process logging or usage guarantee |
| Band 2 · API | Volume | Accuracy — without an evaluation set you don't know if it got worse |
| Band 3 · product | A feature the customer pays for | Exemption — logs with customer data, prompt injection, retention and contract remain yours |
The costs that never show up on the invoice
Five items appear in every project and on no price list. Integration: connecting CRM, ERP or a knowledge base is measured in engineering days, not tokens. Evaluation: every model swap means running the regression again, and models get new versions several times a year. Security: key rotation, log retention and prompt-injection review. Compliance: the data-processing agreement, the retention policy and a record of automated decisions. People time: training, process change, and the week your team works the old way and the new way at once.
Anyone building the business case on token prices alone is off by an order of magnitude. In bands 2 and 3, tokens are usually the smallest line in the budget.
When to move up a band — and the checklist before you sign
Two triggers, and only two. Move from band 1 to band 2 when the same task repeats more than 500 times a month with a stable input format; below that, one person with a subscription is cheaper. Move from band 2 to band 3 when a customer pays for the result, not just when the internal team uses it. A competitor moving up a band is not a trigger.
Checklist before you sign anything: what the task is, counted per month; what it costs today in hours; what the success criterion is and who validates it; how many seats went unused in the last 30 days; what one task costs on the small model and on the big one; where the logs live and for how long; and how you exit the vendor. If those seven answers do not fit on one page, the problem is not price — it is scope.
- 01What is the task
Counted per month
- 02What it costs today
In hours of work
- 03Success criterion
And who validates the result
- 04Unused seats
How many went unused in the last 30 days
- 05Cost per task
On the small model and on the big one
- 06Where the logs live
And for how long they're retained
- 07Vendor exit
How you get out of the contract if you need to