Ask a business what AI costs and you will usually get a per-seat price for Copilot or ChatGPT. That is the line on the quote, so it is the line people budget for.
It is one of three main costs. The second is usage: tokens, credits and agents, which you pay for by how much AI does rather than by how many people have it. The third is training, and its biggest part never appears on an invoice, because it is your team's time.
They are not the only costs. Connecting AI to your other systems, getting your data into a state it can use, keeping automations working, and checking what AI produces all take time or money, and on a bigger project they can outweigh the rest. This note covers the three every business meets first. It gives the method and the arithmetic rather than a price list, because vendor prices move too often for a price list to stay true.
Licences: start with what you already have
Before buying anything, check what you already pay for. Eligible Microsoft 365 business subscriptions include Copilot Chat with enterprise data protection, so your staff can use AI on work material under your existing agreement with Microsoft. That is the place to start, not personal accounts on free tools, which put business information outside your control.
A paid AI licence is a monthly or annual fee per named person. Microsoft 365 Copilot, ChatGPT Business and Enterprise, and Claude Team and Enterprise all work this way, and Google bundles Gemini into its Workspace plans. What the fee buys varies: deeper access to your own files and mail, higher limits, more capable features, and admin controls. Business-grade data protection does not always need an extra seat, so check what your current plan already gives you before paying for more.
Four things move the real number away from the headline one.
- Minimums. Some plans have a minimum number of seats or only sell annual terms.
- Commitment. Monthly is usually dearer per seat than annual, but annual locks you in before you know who uses it.
- Add-ons. Copilot is added to an eligible Microsoft 365 plan, so check the plan underneath as well as the add-on.
- VAT and currency. Some vendors price in dollars and convert, and the figure on the web page may be before VAT.
The biggest saving is not in any of those. It is in not buying a seat for everyone. In a sixty-person business, the people who live in email and documents all day might be fifteen. Start there, and let everyone else work in the approved tools you already have until there is a reason to pay for more.
Our notes on Copilot for small businesses and ChatGPT vs Claude cover what each licence actually buys.
Usage: tokens, credits and agents
A token is the basic unit AI models are metered in. It is roughly three quarters of a word. Everything you send to a model is tokens in, and everything it writes back is tokens out. Output is priced several times higher than input.
You meet usage costs in three places, and they are not always counted in tokens:
- Anything built on an API. A chatbot on your website, an automation that reads invoices, or an assistant your IT partner built for you. These are usually billed per million tokens, straight from the model provider or through Azure, AWS or Google.
- Agents inside the tools you license. Microsoft meters many Copilot agents, and Copilot Cowork, in Copilot Credits. A credit is not simply a token under another name: the charge can cover the model, the time the agent runs, the context it works with and the tools it calls. Some simpler agents carry no extra charge, so check how the one you want is billed before you assume it is covered by the seat.
- Usage limits on seat plans. ChatGPT and Claude plans include a set amount of use. Heavy users hit the limit, and some plans let you buy extra usage beyond it.
The cost of one job is small. Summarising a 3,000-word report is about 4,000 tokens in and 500 out. At an assumed $3 per million tokens in and $15 per million out, typical of a mid-range model, that is about two US cents, or under two pence. A larger model costs more and a smaller one less.
The bill comes from volume and from agents. An agent that works through a task in many steps may process its context again at each one, and that repeated processing can push a single job into hundreds of thousands of tokens. Say each run costs about £2. Ten people running one each working day, over 220 working days, is £4,400 a year, or about £367 a month.
A licence is a fixed cost you can see in advance. Usage is a running cost you only see afterwards, unless you set a limit first.
So set one, and check what kind it is. An alert tells you when spend passes a figure. A hard limit stops usage, and even then some providers let a small amount through before it takes effect. Make sure the control you choose actually stops spending if that is what you need. Agree the limit before anything goes live, check actual use after the first month, and adjust from real numbers. Check current prices on the provider's own page, such as Claude's pricing, OpenAI's API pricing or Microsoft's Copilot Studio pricing, on the day you plan.
Training: the cost that makes the other two work
Licences without training are an easy way to waste money. People log in, try two prompts, get a weak answer and go back to how they worked before. The seat stays on the bill.
Budget training per person, and count two parts.
The fee. Group training is usually priced per session, so filling a suitably sized group lowers the fee per person. Size still matters: people need room to practise and ask questions, and a group too big for that is cheaper per head and worse at the job.
The time. This is usually the larger cost. A course of two half-days takes a working day per person. For someone on £35,000, a day is about £135 of salary, taking a year as 260 weekdays. Add employer National Insurance at 15% above £5,000 and a 3% pension contribution on qualifying earnings, and the fully loaded cost is nearer £155 a day. Across sixty people that is about £9,300 of time. It is worth spending, but it should be in the budget, not discovered later.
Businesses already spend this way. The government's Employer Skills Survey 2024 puts average employer spend at about £1,700 per employee a year across all training, not just AI, and about half of that is the wages of staff while they train. Time is already the biggest part of what employers invest in training. AI training is no different.
Two more lines are easy to forget. Tools change every few months, so budget a short refresher each year rather than treating training as done. And somebody has to own AI in the business from the start, with time set aside for it: watching the spend, answering questions, setting the rules on which tools are approved, and deciding what to license next.
Train more people than you license. The approved tools you already have reach everyone, the judgement about what to trust reaches everyone, and the people trained on those tools are the ones who will tell you who should get a paid seat next.
Putting it together
For a team of sixty, a first-year budget looks something like this. An owner for AI, with time allocated from day one. Licences for the fifteen or so people who will use them daily, annual only once a pilot has proved it. A usage limit for anything built or any agents, set low and raised on evidence. Training for everyone, with the time counted as a cost. A refresher in year two. And an honest allowance for the work around the edges: connecting systems, preparing data and checking outputs.
Then test the total properly. A seat that saves an hour or two a month can cover its equivalent cost, but saved time is not the same as saved money. Take off the time spent checking and correcting what AI produces, then ask whether the time released goes on something useful. Training works the same way. It pays off not because people use the tool, but when their work gets enough better to justify what you spent. If you cannot yet say either, you are not ready to buy at scale.
How to start
Train the team on the approved tools you already have. Pick two or three heavy users for a one-month paid trial. Set a usage limit before anything automated goes live. After a month, look at who used what, what it saved once checking is counted, and what that time went on, then decide with numbers.
That is the order we follow in training: the approved tools you already have first, then what the paid tools add, so your team leaves knowing which licences are worth paying for.
Last reviewed November 2026. Licence and usage prices change often and without much warning. The worked figures here are illustrations with their assumptions stated, not quotes. Check each vendor's current pages before you budget on them.
Written with Claude, then steered, cut and checked by me. I answer for it.
Explore Benchmark Lab’s free indices for personal AI use, organisational adoption and governance. All are in Research beta and can be used independently of training.