Key takeaways
- AI cost has four separate lines: seats, usage, setup, and the internal time nobody budgets for.
- Seat cost is predictable. Usage cost is the one that surprises people, and the one worth capping.
- Most overspending comes from a premium model doing routine work, not from too many users.
- A dormant seat costs exactly as much as an active one, so seat hygiene is real money.
The honest answer to "how much does AI cost a business" is that the licence is the small, predictable part, and everything else depends on decisions you have not made yet. This article breaks the cost into the four lines that actually appear on an invoice, shows what drives each one, and works through three examples so you can build a number you can defend.
Prices move, so treat the figures below as planning ranges rather than quotes. The structure of the cost, which is the part that matters for budgeting, has been stable for a while and is unlikely to change soon.
Line 1: seats
Every major provider sells a per-user, per-month subscription for its assistant, and business tiers cluster in a fairly narrow band. Expect roughly 25 to 60 US dollars per user per month at a business or enterprise tier, usually discounted for annual commitment. Some vendors bundle their assistant into a productivity suite you already pay for, which changes the arithmetic considerably and is worth checking before you sign anything separate.
The important question is not the headline price but which tier you actually need. Consumer plans are cheaper, and they are also missing the things that make a rollout defensible: a central admin console, organisation-wide billing, data retention controls, training opt-outs at account level, audit visibility, and single sign-on. If you buy consumer plans to save money, you are buying a governance problem at a discount.
Watch the tier ladder. Some providers put single sign-on and audit logging behind their most expensive tier. If SSO matters to you, price it before you shortlist, because the jump between tiers can be larger than the difference between two vendors.
Line 2: usage
Anything that runs through an interface rather than a person clicking in a chat window tends to be billed by consumption. This is where budgets go sideways, for a simple reason: consumption is a function of design choices, not headcount, and those choices are usually made by whoever builds the automation rather than whoever owns the budget.
Three factors dominate consumption cost:
- Model tier. The gap between a provider's fastest small model and its most capable large model is often an order of magnitude in price. Running everything on the flagship is the most expensive habit in the category.
- How much context you send. Attaching an entire document to every request when a relevant extract would do multiplies the cost of every single call.
- How often it runs. A process triggered on every record in a database behaves very differently from one triggered on demand, and the difference only becomes visible on the invoice.
None of these are hard to control. They just need deciding deliberately at setup rather than discovering in arrears.
Line 3: setup
Someone has to choose the provider, buy at the right tier, configure the admin console, set retention rules, connect identity, cap the spend, and teach people what to do. That work happens whether you pay a supplier for it, absorb it internally, or skip it and pay for it later in incidents and waste.
The market prices this two ways: hourly consulting against an open scope, where the final number is unknown when you sign, or a fixed fee agreed up front. For work whose scope is genuinely knowable in advance, and a first AI rollout usually is, a fixed fee is the better structure, because it puts the estimation risk on the party actually in a position to estimate.
Line 4: the cost nobody budgets
Internal time. The evaluation meetings, the pilot coordination, the person from IT who spends a morning on identity configuration, the manager who fields questions for a month. This line is invisible in finance systems and is often the largest of the four when a rollout drags.
It is also the line most affected by pace. A rollout compressed into 30 days consumes a fraction of the internal time of one that meanders across two quarters, which is a large part of why speed functions as a cost control rather than a risk.
Three worked examples
These are illustrative planning models, not quotes. They assume a business tier assistant licence, modest usage-based automation, and a one-off implementation.
| Scenario | Seats | Annual licence, indicative | Usage, indicative | What drives the number |
|---|---|---|---|---|
| 12 person professional services firm | 10 licensed, 2 admin only | Roughly 4,000 to 7,000 USD | Low, often under 100 USD a month | Almost all chat-based use, so cost is essentially seats. |
| 50 person operations business | 35 licensed | Roughly 13,000 to 25,000 USD | Moderate, a few hundred a month | Seats plus one or two automations touching document or email volume. |
| 200 person multi-site organisation | 120 licensed | Roughly 45,000 to 85,000 USD | Variable, and the line worth capping hardest | Seat hygiene and model selection matter more than the negotiated seat price. |
Notice the pattern. At small scale, cost is dominated by seats and is highly predictable. At larger scale the variance comes from usage and from seats that are paid for but not used, which is why the controls matter more than the negotiation.
Where the money actually leaks
Businesses that feel their AI spend is out of control are usually losing money in four specific places.
Premium models doing trivial work
Reformatting a list, drafting a two-line reply, pulling a date out of a document. None of this needs a flagship model. Mapping routine tasks to a cheaper tier is the highest-return cost change available, and it is usually invisible to end users.
Dormant seats
A licence bought in a burst of enthusiasm and unused for five months costs exactly as much as one used daily. A quarterly review of last-login data, with unused seats reclaimed rather than renewed, is dull work that pays for itself.
Duplicate subscriptions
Different departments buying different tools, often on personal cards, often for overlapping purposes. Consolidating to one governed provider usually reduces total spend even when the per-seat price goes up, because it removes the duplication.
Automations with no ceiling
A misconfigured loop can generate a genuinely alarming invoice over a weekend. A hard spending limit at the provider turns that scenario into a service interruption you notice on Monday and fix in an hour.
The single most valuable control: a hard cap, not an alert. Alerts tell you what you have already spent. A cap decides in advance what the worst case is allowed to be.
How to build a number you can take to finance
- Count the people who will genuinely use it weekly. Not headcount, not the whole department. Weekly users.
- Multiply by the business tier price for the provider you are leaning towards, at annual billing.
- Add a usage allowance for automation. If you have no automations yet, add a small placeholder rather than zero.
- Add the one-off implementation fee as a separate line so it does not distort the run rate.
- Set the hard cap at roughly 1.5 times your expected monthly usage line, then leave it alone.
- Diarise a 90 day review to reclaim dormant seats and check the model mapping is holding.
That gives you a defensible annual figure with a known ceiling, which is a far easier conversation than an open-ended one.
Frequently asked questions
How much does AI cost per employee per month?
Business and enterprise assistant tiers generally sit between 25 and 60 US dollars per user per month, cheaper on annual billing. Usage-based automation is billed separately and depends entirely on what you build.
Why did our AI bill go up unexpectedly?
Almost always one of three things: a premium model handling routine work, an automation running more often than anyone realised, or seats still being billed for people who have left or stopped using the tool.
Is the cheapest plan good enough?
Rarely, because the consumer tier lacks admin controls, central billing, retention settings and single sign-on. At the business tier you are not really buying a better model, you are buying the ability to govern it.
Should we negotiate on price?
At small scale the effort is better spent on model mapping and seat hygiene, which move the number more than a discount will. At a few hundred seats negotiation becomes worthwhile, and annual commitment is your main lever.
Get a cost model before you commit
The Clarity Package includes a provider and model direction plus the cost and risk traps to avoid, written up so you can take it straight to finance. The Implementation Package sets the hard limits and model mapping for you as part of the build.