How much does AI cost a business?

By Andrej Ruckij · November 5, 2026

TL;DR: It depends on which of two bills you’re on, and most small businesses are only on the first. Seats (ChatGPT Business, Claude Team, Copilot) cost a fixed $20-25 per person a month; you don’t pay more for heavy use, you hit a usage limit, and only pay extra if you buy a premium seat or switch on usage credits, which Anthropic bills at standard API rates. API usage is what you pay when you build AI into your own systems: per task, from about $0.50 to $50 for drafting 1,000 support replies depending on the model. Big bills come from the second one at scale, like engineers running coding agents all day (Apple just cut its cap from $300 to $200 a day per engineer). Separately from both, budget people’s time: one survey puts sorting out poor AI work at about $186 per employee a month.

Which bill are you on?

	Bill 1: seats	Bill 2: API usage
What it is	A subscription to the ChatGPT, Claude or Copilot app for each person	Paying per token when your own software calls a model (a support inbox, a CRM, a script, an agent)
What it costs	$20-25 per person a month, fixed	per task; see the worked examples below
What happens with heavy use	You hit a usage limit and wait, buy a premium seat, or switch on usage credits, which Claude’s pricing page says keep you working “at standard API rates” (Anthropic)	The bill grows with every request
Who’s on it	Almost every team that uses AI	Teams that automate or build with AI

The two don’t add up line by line: a team using only the apps has bill 1 and nothing else. The third cost, people’s time spent checking and fixing output, applies to both and never appears on an invoice.

Why this question is suddenly everywhere

For most of 2026 the message inside companies was “use AI as much as you can”. Now the bills have arrived. On November 4 Bloomberg reported that Apple cut its engineering teams’ daily AI budget from $300 to $200, that some allowances may fall to $100 a day, and that it is considering a limit as low as $25 a day outside engineering (TechCrunch). Uber earlier capped staff at $1,500 a month in token spending per AI coding tool, after it used up its 2026 AI budget in four months (Simon Willison, citing Bloomberg).

Those are big-tech numbers for engineers running coding agents all day, which is usage on bill 2. A 20-person company using the apps doesn’t spend like that. Here’s what each layer costs.

The three layers of what AI costs

1. Seats: the subscription per person

This is what most businesses mean by “AI cost”: a monthly plan per user.

Plan	Price per user per month	Source
Claude Team Standard	$25 monthly, $20 billed annually	Anthropic, official
Claude Team Premium	$125 monthly, $100 billed annually	Anthropic, official
ChatGPT Business Standard	$25 monthly, $20 billed annually	OpenAI, official
Microsoft 365 Copilot Business	$21 list (add-on to Microsoft 365)	Microsoft, official
Google Workspace Business Standard	€13.60, Gemini included in Gmail, Docs, Sheets and Meet	Google, official

Seats are predictable, which is why finance teams like them. Their catch is usage caps: standard seats on both ChatGPT Business and Claude Team have rolling limits. The ways past them are a premium seat at five times the price or, on Claude, usage credits billed at standard API rates, which moves you onto bill 2 (Anthropic). I compare the two plans in detail in ChatGPT Business vs Claude for business.

2. Tokens: paying per use (bill 2)

The moment you connect AI to your own systems (a support inbox, a CRM, a product catalogue) you usually pay per token through an API. A token is roughly three-quarters of a word. Prices are per million tokens, separately for what you send (input) and what comes back (output).

Model class	Example	Input / output per 1M tokens	Source
Frontier	Claude Fable 5.1, GPT-6 Astra	$10 / $50	Anthropic, OpenAI
Strong mid-tier	Claude Opus 5.5	$4 / $20	Anthropic
Everyday	Claude Sonnet 5.5	$2 / $10	Anthropic
Cheap	Claude Haiku 5.5	$0.10 / $0.50 (prompts up to 100K tokens)	Anthropic

What companies actually pay is lower than list. Ramp, which sees its customers’ card spend, found token costs averaged $0.72 per million tokens in April 2026, mostly because of caching (Ramp).

3. The costs that don’t show up on the AI invoice

Sorting out poor AI work. In a Stanford Social Media Lab and BetterUp survey of 1,150 workers, about 40% said they had received low-quality AI work from colleagues, which took about two hours to sort out each time. The researchers put the hidden cost at about $186 per employee a month (Entrepreneur). It’s a self-reported survey, but $186 is about nine standard seats.
Swings in usage. Ramp found month-over-month swings of 40% or more are common even when headcount stays flat, and advises a 40-50% buffer on token budgets (Ramp).
Building and maintaining the integration. I found no reliable figure for this. If you’re commissioning a custom setup, get a quote for build and upkeep before you count any savings.

What does a real task cost? Two worked examples

These use my own assumptions, stated so you can redo them with your numbers.

Drafting replies to 1,000 support emails. Say each request sends 3,000 tokens (the email, your instructions and a knowledge-base excerpt) and gets back 400. That’s 3 million input and 0.4 million output tokens.

Model	Cost for 1,000 emails
Frontier ($10 / $50)	3 × $10 + 0.4 × $50 = $50
Opus 5.5 ($4 / $20)	$12 + $8 = $20
Sonnet 5.5 ($2 / $10)	$6 + $4 = $10
Haiku 5.5 ($0.10 / $0.50)	$0.30 + $0.20 = $0.50

Summarising 500 thirty-minute sales calls. About 6,000 tokens of transcript per call and a 500-token summary: 3 million input, 0.25 million output. Opus costs about $17, Sonnet about $8.50, Haiku about $0.43. Transcription isn’t included.

Two things change these numbers. Models that “think” before answering bill that thinking as output: Anthropic’s docs say thinking tokens are billed as output tokens (Claude Code docs). If each email triggers 2,000 tokens of thinking, the Opus run goes from $20 to $60. In the other direction, batch processing halves the price at both Anthropic and OpenAI if you can wait for results.

So at small-business volume, a well-defined task usually costs less per month than one seat. Agents are where the bill grows.

Why AI costs so much for some companies

Because agents run in loops. A chat answer is one request. A coding or research agent reads files, calls tools, rereads its own history and tries again, all billed. Anthropic says Claude Code averages about $13 per developer per active day and $150-250 a month, and that agent teams use roughly seven times more tokens than standard sessions when teammates run in plan mode (Claude Code docs).

That’s how you get to the big-company numbers:

Apple: $200 a day per engineer after the cut, about $4,000 a month (TechCrunch).
Workday and Stripe: per-employee caps of about $2,000 a month, according to SemiAnalysis’s conversations with companies about token budgets (SemiAnalysis).
Shopify: an alert rather than a hard stop. Shopify’s engineering team told Bessemer: “We get alerts if someone spends more than $250 in tokens in a day”, and those cases are looked at, not shut down (Bessemer).

And the typical company is nowhere near that. Among companies that pay for tokens at all (bill 2), Ramp puts the median at $46 per employee per month, with the middle half between $3 and $352 (Ramp).

Prices can also rise without the price list changing. OpenAI’s new Images 2.6 kept the same per-token rates as 2.5, but in CellCog’s test it used about 4.2 times the tokens at low quality and 3.3 times at medium (CellCog). I’d check the monthly bill after any model upgrade.

Is AI cheaper than an employee?

As a tool, by a mile. The median US customer service representative earns $44,770 a year, about $3,730 a month before benefits and taxes (O*NET, BLS 2025 data). A $25 seat is under 1% of that.

Replacing a person is a harder sum. The best measured result is support work: in a study of 5,179 support agents, AI help raised issues resolved per hour by 14% on average and by 34% for novice workers, with minimal effect on experienced ones (NBER). For experienced developers, METR’s randomised study found they took 19% longer with AI tools while believing they were 20% faster (METR). A 2026 follow-up suggested developers may now finish tasks faster with AI, but METR itself calls the data “only very weak evidence” (METR).

My read: AI pays back fastest where the work is repetitive and the person is new to it. It pays back slowest where you’re already good at the job, and that’s also where people overestimate the gain.

Is AI cost effective? A test before you buy more

Ask three questions per use:

Is the task defined? “Answer refund emails using our policy” can be costed and checked. “Help with marketing” can’t.
Who checks the output, and how long does it take? If review takes longer than doing the work, you’re paying twice.
Is the volume high enough to matter? Saving 10 minutes on a task you do twice a month doesn’t justify a project.

What to do with this

For a team of 5-50 people, this week:

List every AI charge. Seats, API keys, and AI add-ons inside other tools such as your CRM, helpdesk and design software. Check your card statements for the last three months.
Pick one plan, not three. If you already pay for Google Workspace or Microsoft 365, test the AI included there before adding a separate subscription.
Set a budget per role, with an alert rather than a hard stop. Claude Team and Enterprise admins can set spend limits per organisation, group or member (Claude Code docs). If you build on the API, a gateway such as LiteLLM can cap spend per key or team (LiteLLM docs).
Route by difficulty. Use the cheapest model that passes on your own 20 test cases, and keep the expensive one for the hard cases. The worked example above shows a 100x price gap between Haiku and a frontier model on the same job.
Turn on caching and batching if you use the API. Cached input on Claude is billed at 0.1 times the normal price or less, depending on the model (Anthropic), and batch halves the price.
Measure the time spent fixing output for one month. That’s the number that tells you whether AI is cheap for you.

Who this doesn’t apply to: teams building AI products, where tokens are the cost of goods and need their own unit economics, and companies above a few hundred seats, where enterprise contracts change the prices.

What I couldn’t find

I found no reliable figure for what it costs a small business to build and maintain an AI integration. Seat prices were checked on OpenAI’s and Anthropic’s own pricing pages on November 5, 2026; both change often. Ramp’s figures cover its own customers, which skews toward US companies that already track spend closely.

Related: should you self-host an LLM? covers when open models cut this bill further, and wiki/glossary/prompt-caching explains caching. If you’re weighing AI phone answering specifically, see is an AI receptionist worth it for a small business?.

Sources

https://t.me/aidaily_business