AI usage and limits

Your plan includes an amount of AI work each month. This page explains how that is measured, so the number on your billing page is one you can predict rather than one you discover.

Tokens, briefly

AI providers do not bill per question. They bill per token — roughly three quarters of a word, so about 750 words to 1,000 tokens. Every question your assistant answers uses tokens twice:

• Input tokens are everything the model reads: the question, the passages retrieved from your knowledge base, the instructions you have set, and the recent conversation.
• Output tokens are what it writes back.

Output costs about five times as much as input on every model we offer, which is why your billing page shows them separately rather than as one number. A short answer over a lot of context is cheap; a long answer is not.

What counts towards your allowance

Everything the AI does for your workspace:

Two things do not count: a conversation a visitor opens without asking anything, and a question that is refused before it reaches the model (outside your knowledge base, or over a limit).

How adding documents uses tokens

When you add a document, we split it into passages and turn each one into a list of numbers — an embedding — so that we can find it later by meaning rather than by keyword. That happens once per document, and embedding is far cheaper than answering: roughly a hundredth of the price per token.

A thousand-page manual costs less to index than a busy afternoon of questions. Re-indexing happens when you change the embedding model or the chunk settings on a knowledge base, and the app warns you before it does.

When it resets

At the start of each billing period, shown on your billing page as Resets. If you are on a monthly subscription, that is the same day each month, not the first of the calendar month.

Unused allowance does not roll over.

What happens when you reach the limit

An answer already being written always finishes. After that, new questions are refused until the period resets, and your visitors see a short message rather than an error.

Nothing is charged for going over, and nothing is cut off mid-sentence. You will see a warning on your billing page from 80% onwards.

If you need more before the reset, changing plan takes effect immediately.

Which models use more

Every model reads the same context, so the difference is price per token rather than tokens used:

You can change the model per knowledge base under Settings, and the choice only affects questions asked after you change it. Which models are available depends on your plan.

How your retrieval settings affect usage

This is the lever most people miss. Your assistant reads the passages it retrieved before it answers, so how many passages you retrieve is most of your input cost.

On each knowledge base, under Retrieval:

• Top K is how many passages are read per question. Going from 8 to 5 cuts roughly a third of the input on every single question, and usually improves the answer, because weak matches are noise.
• Similarity threshold drops passages that are not a good match. Raising it slightly has the same effect, and for the same reason.
• Chunk size decides how big each passage is. Smaller chunks retrieve more precisely; very large ones mean each of the K passages carries more text.

If your usage is higher than you expected, look here before you look at the model.

Bring Your Own key

On the Bring Your Own plan, you connect your own Anthropic or OpenAI key under Settings → Your own AI keys. From that moment:

• Every answer and every embedding for the workspace runs on your provider account, and that provider bills you directly for it.
• Your WelcomeAI allowance no longer applies — there is nothing of ours to run out of, so nothing is refused for being over a limit.
• Your billing page still shows tokens and an estimated cost, so you can see what your assistants are costing you before your provider invoice arrives.
• You can bring a key for one provider and not the other. If you supply an Anthropic key but not an OpenAI one, both are needed — answers run on yours and embeddings have nowhere to go — so add both.

Your key is encrypted, is never shown again after you save it, and is checked against the provider before it is stored. Remove it at any time and the workspace goes back to our accounts on your next question.

Reading your billing page

Billing and usage shows, for the current period:

• AI tokens / month — used against included, as a bar.
• Input and Output separately.
• Allowance — what is left.
• Resets — when it goes back to zero.
• By model — where it went, most expensive first. This is the table to open when a number surprises you.

The cost shown is an estimate at current provider rates. It is what your usage costs to serve, not a charge to you — unless you are on Bring Your Own, in which case it is a close estimate of what your provider will bill.