AI request ceiling
The AI request ceiling tab of the System screen. It caps how fast paid requests to the AI provider may be made, so that one person — or one runaway screen — cannot spend the whole month's budget in an afternoon.
The count is kept per account, not per network address: all of this happens behind a sign-in, so the address identifies nobody. Above the ceiling a request is refused with a message saying when to try again, and no money is spent on it.
Requests per minute
Requests per minute per account is the ceiling itself: a whole number of 0 or more. The default is 120, which sits well above ordinary work — for a sense of scale, «translate into all languages» sends about six requests at once, bulk image generation arrives at roughly six a minute, and a batch of store reviews yields eight to twenty-four a minute.
Zero switches the ceiling off — requests stop being counted at all. Anything that is not a whole number of 0 or more is refused with Enter a whole number of 0 or more (zero switches the ceiling off)..
A change takes effect on the next request; no restart is needed.
What the ceiling covers
What the ceiling covers has two positions:
- Every paid request — every paid request. This is the default.
- Push campaigns only — push campaigns only.
Every paid request is the default deliberately: push campaigns are not the only thing money goes on, and a ceiling covering them alone would leave the most expensive path — the assistant — unwatched.
The ceiling sits at the single point every AI request passes through, not on individual screens. That is why it also covers requests made by the assistant's own tools rather than by a person pressing a button.
What stays outside the ceiling
Two paths are not capped, and each for its own reason:
- The scheduled analysis of creatives. No person stands behind it, and it paces itself on every tick. Its spend still shows up in the AI spend overview.
- Turning text into vectors for search by meaning. That is a separate layer with a setting of its own — see Semantic search. By default vectors are computed on this server and nothing is paid for outside, so this appears neither in the ceiling nor in the spend ledger.
Transcribing a voice message, on the other hand, is inside the ceiling: it counts against whoever sent the message, under the same limit as everything else. When the sender cannot be matched to an account, a cloud recognition is refused and that message gets no transcript — the engine running on this server costs nothing, so in that position it transcribes anyway.
Press Save to store both settings at once. Pressing it without a change answers No changes.