AI operating platforms are a new kind of product, and the way they are charged for can feel confusing. A software licence used to be a fixed price per person. AI has a running cost every time it is used, so most AI products charge, at least in part, for use.
Here we try to demystify the main ideas: what you are paying for, why it varies, and the ways AI companies put a price on it.
Two analogies for an AI platform
The car. The platform is the car, and AI credit buys the fuel. With David, credit buys the wiper fluid too: transcription, search and the other services that are charged by use.
The mobile phone contract. The platform is the line rental, and AI credit buys the minutes and the data.
In both, there is a fixed part you pay for whether or not you use it much, and a running part that goes up and down with how much you use it.
How AI credit works: the travel card
AI credit works like topping up a travel card. You load money, each journey deducts a fare, longer journeys cost more, and when the balance runs low you top up.
What is AI credit?
AI credit is what you spend when you use an AI-powered platform or tool. It measures how much AI "thinking" your requests use.
The services behind a platform like David are charged by use: AI models, transcription, search and more. A monthly plan usually comes with an allowance of credit to cover them.
Why AI credit exists
Different AI tasks use wildly different amounts of computing power:
- A short text reply is cheap.
- A ten-second high-definition video, or a long reasoning task such as building a tool, is expensive.
Credit lets heavy and light use be charged for fairly, without anyone having to think in server time or tokens.
How credit is spent
Each request run through an AI model takes a fare in credit. The size of the fare depends on:
- How much content the AI processes: the words it reads and writes, or the seconds of audio and video.
- Which model does the work: small models cost less credit; advanced ones cost more.
- How many steps it takes: a single reply costs less than a multistep agent process that searches, reads and checks before it answers.
Which model does the work
The model is the biggest lever on what a task costs. The most advanced models can cost many times more than small ones for the same amount of text, and many everyday tasks, such as sorting an email or reading a date off an invoice, don't need them.
A platform that picks the model for each task can do the same work for a fraction of the cost of one that sends everything to its most advanced model. It is worth asking any supplier how that choice is made, and who benefits from it.
What's a token?
Models read and write text in small pieces called tokens. A token is often part of a word: in English, a token averages about three quarters of a word, so 1,000 tokens is roughly 750 words.
AI providers price their models per million tokens. What the model writes (output) usually costs several times what it reads (input). Images, audio and video are either turned into tokens or charged by the second.
Tokens are precise, but hard to budget with. That is one reason many platforms convert them into credit.
Ways AI companies charge
Most AI products use one of these, or a mix:
- Per seat: a fixed monthly fee per person, often with a fair-use limit on the AI behind it. Predictable, but you pay for people who rarely use it, and your heaviest users hit the limit.
- By usage: you pay for exactly what is used, at a price per million tokens. Fair, but hard to predict and to read.
- Credit: usage across many models and services converted into one unit, so a month's spend reads in one currency.
- Plans: a platform fee with an allowance of credit bundled in, and top-ups if a month needs more.
Questions to ask any supplier
- Is AI usage passed on at cost, or marked up? By how much?
- Is the allowance per person, or shared across the organisation?
- Does unused allowance expire at the end of the month, or roll over?
- What happens when the allowance runs out: does the service stop, slow down, or charge more?
- Can you set alerts or limits, so a busy month doesn't surprise you?