How AI pricing works: tokens, context and cost, explained for lawyers
By the ChamberOne team 5 minute read
Most lawyers meet AI pricing as a line on a quote that they do not quite understand. It might say 'per seat', 'per document' or 'usage-based', and underneath sits a word that sounds like arcade money: tokens. You do not need to become an engineer to read a pricing page sensibly. A handful of ideas explains why costs behave as they do, and why some bills surprise people.
Key points
- Tokens measure text going in and coming out; longer documents and longer answers cost more.
- Ask what a unit of usage is, what the limit is and what happens when it is reached.
- Set caps and alerts, and test large jobs on a small sample before running them in bulk.
- Match the model to the task and still review every output.
What a token is
A language model does not read words as we do. It breaks text into small pieces called tokens. A short common word is often one token. A long or unusual word, such as a technical legal term or a party's name, may be split into several. Numbers and punctuation also count.
For planning, think of tokens as a measure of how much text goes in and comes out. More text means more tokens. The exact ratio varies by model and by language, so do not rely on a rule of thumb for budgeting. Ask the vendor to show real counts on a sample of your own documents.
Input, output and context
Providers usually charge separately for the text you send in (input) and the text the model writes back (output). Output tokens are normally priced higher than input tokens, which is why a request that asks for a long, detailed answer costs more than one that asks for a short summary of the same document.
The context window is the amount of text a model can consider at one time. A model with a large window can take a whole contract, or several, in one go. A smaller window forces the tool to split documents into pieces, which can cost extra effort and sometimes loses connections between distant clauses.
- The text you send in
- Charged separately
- Normally priced lower than output
- The text the model writes back
- Normally priced higher
- A long, detailed answer costs more than a short summary
Why long documents cost more
The reason is not mysterious. A hundred-page agreement contains far more text than a two-page letter, so sending it means far more tokens. If your workflow asks the tool several questions about the same document, and each question sends the whole document again, you pay for it each time. Some systems store the document once and send only relevant extracts, which can cut cost, but it also depends on how well the extract is chosen.
Scanned documents are another cost. If text must first be read from an image, that step has its own cost and its own error rate. Messy scans can also create nonsense text, which wastes tokens and produces unreliable answers.
How software plans usually bundle usage
Legal software vendors tend to package AI in a few ways. None is better in all cases.
- Included in the subscription: simple to budget, but check for fair-use limits and what happens when you reach them.
- Allowance plus overage: a set amount of usage is included and extra is charged. Ask the overage price and how you are warned.
- Pay as you go: you pay for what you use. Flexible, but harder to forecast.
- Bring your own key: you pay the AI provider directly and the software vendor charges separately for the software. This gives you visibility, and puts more administration on you.
Questions to ask before you sign
- 1
What counts as a unit of usage: a document, a page, a question, or a token?
- 2
What is the limit on the plan, and what happens when it is reached: a hard stop, a slowdown or an extra charge?
- 3
Can I see usage by user, matter or feature, and export it?
- 4
Can I set a cap or an alert so one person or one large job cannot use the whole allowance?
- 5
Which model is used for which task, and can I choose a cheaper one for routine work?
- 6
Are prices fixed for the contract term, or can the vendor change them?
- 7
Are prices shown with or without tax, and in which currency?
Bill shock: how it happens and how to avoid it
Imagine a mid-sized firm that tries an AI tool on a due diligence project. Someone uploads the entire data room, a few thousand files, and asks the tool to analyse every one. In a usage-based plan, a single enthusiastic afternoon can use a month's worth of allowance. Nothing was done wrongly; nobody simply knew that bulk work counts.
The remedies are dull and effective. Set caps. Appoint one person to watch usage in the first few months. Run large jobs deliberately, with a budget agreed in advance, and tell the client if the cost is to be passed on. Test on a small sample first and multiply.
- Set caps
- Appoint one person to watch usage in the first few months
- Run large jobs deliberately, with a budget agreed in advance
- Test on a small sample first
- Upload a whole data room and analyse every file in one afternoon
- Leave a usage-based plan without a cap
Cheaper does not always mean worse, or better
Models come in tiers. Smaller, cheaper models are often good enough for sorting, tagging or extracting plain facts such as dates and parties. Larger models cost more and are usually better at nuanced reading and drafting. A sensible firm matches the model to the task and checks quality on its own samples, rather than defaulting to the biggest model for everything.
Keep a simple log of what you tried, which model and how many tokens it used. After a month you will have your own numbers, which are worth far more than any rule of thumb in a vendor presentation.
Whatever the price, the output still needs a lawyer's review. Paying more reduces some kinds of error, but does not remove the need to check.
Passing the cost on
If AI usage is a real cost on a matter, think in advance about how it is dealt with in your fee arrangements. Some firms absorb it as overhead; others itemise a disbursement where the client has agreed. Check the guidance from your regulator or bar on what may be charged to clients and how it should be described, and say it plainly in the engagement letter so there is no argument later.
AI usage in ChamberOne
- Choose the model for each AI feature, GPT-5 or Claude, or switch AI features off for the firm.
- Every AI call is recorded in the audit log with the model and token counts.
What it does not do
- It does not make AI output correct. Output can be wrong or incomplete and does not replace a lawyer's review.
- With AI features on, the text needed for a request is sent to the AI provider for that feature, outside the country where ChamberOne is hosted.
This article is general information about the law, legal practice and technology, current when it was published. It is not legal advice. Laws and guidance change: check the sources linked above and take advice on your own circumstances from a lawyer qualified where you are.
Read next
AI and the billable hour: what actually changes and what does not
How faster drafting puts pressure on hourly billing, and which fee models suit which work.
5 minute readFixed-fee pricing with AI-assisted scoping
How to use past matter data and AI summaries to scope a fixed fee, define exclusions and set a review point.
5 minute readRecovering unrecorded time: how AI time capture works and its limits
How suggestions from calendar, email and documents help fill timesheet gaps, and why the fee earner must still confirm each entry.
5 minute read