Skip to content
Pilot phase – not yet a commercial offer. The first 500 users join free of charge and share feedback to help us improve. The first pricing round starts in May 2027
Firmspace
Back to blog

The AI Budget That Runs Out Before Lunch: Who Spends a Firm's Daily Requests

At 11:40 on a Tuesday the panel goes quiet. A manager has asked the assistant to prepare three invoices and gets one sentence back: the daily AI limit is used up and resets tomorrow. Twenty minutes later the automation journal shows a rule that was supposed to draft a follow-up after a new deal — status: error, same reason. Two screens, two people, one counter.

Firmspace AI comes with a daily allowance of requests, not with a token budget and not with a monthly pool. The number belongs to the plan and to the firm, not to a single employee. A plan without the AI section has no allowance at all — the section is simply absent there. A plan with the section carries a fixed number of requests per day, and the assistant's panel shows how many are left.

The same counter serves three doors. The chat panel in the CRM, the Telegram bot and the AI action inside an automation rule all call one function before they speak to the model. A robot that wakes at three in the morning spends from the same bucket as the owner who asks for a summary at noon. There is no separate robot budget and no borrowing from tomorrow.

The bookkeeping is deliberately boring: one row per firm per day. A request increments the row only while the count is below the limit, in a single conditional update, so two people who press Send at the same moment cannot both take the last slot. The slot is claimed before the model is called. If the call then fails, nothing is refunded — the slot is already spent.

One request is not one model step. A message that the assistant splits into a queue of tasks costs one request, even if the queue runs for minutes and each task needs sixteen steps. A single automation firing costs one. Several tool calls, a long document, a search that finds nothing — all inside the same request. The unit is the conversation turn, not the work.

At zero the refusal is explicit. The chat answers with one short sentence and the panel shows nothing left; Telegram replies with a fixed sentence; an automation rule does not pretend it worked — the journal records an error with the same reason. Other rules triggered by the same event still run. Only the AI step of this rule stops.

There are two different stops, and they are easy to confuse. The first is the plan: if a firm downgrades and the autonomous AI step is no longer included, the rule fails with a message about the plan, not about the allowance. The second is the day: the plan includes the step, but the allowance is gone. The demo cabinet has a third, smaller ceiling, because the platform pays for those requests itself.

The day in the counter is the server's UTC day. In Berlin that means the reset arrives at one or two in the morning, depending on summer time, not at midnight. A batch that starts at 00:30 local time still belongs to the previous day; a batch at 02:30 belongs to the new one. “It resets tomorrow” is therefore true, but the tomorrow is the server's, and a night job can meet the old counter.

The ceiling exists for two reasons. A model call costs money, and a rule can fire on every new lead, every incoming message and every stage change; without a ceiling a loop or a busy Monday turns into a bill nobody planned. The second reason is predictability: one number for chat, Telegram and robots can be checked on one screen, instead of being split into three budgets that quietly compete.

What a firm can do with this knowledge is unglamorous. Look at the remaining count before a batch. Prefer one message with several tasks over several messages, because the queue costs one request. Do not attach an AI step to an event that fires hundreds of times a day. Read the automation journal for errors instead of assuming a silent rule did nothing. And if the allowance runs out regularly, the lever is the plan, not a second account.

None of this makes the assistant less useful. It makes it a colleague with a timesheet. A firm that knows where the line is can plan the work around it — and the assistant, unlike a tired employee, will be back at the same hour tomorrow.

Let us automate your next process together.

Describe a recurring workflow to us. Together we will check whether and how it can be automated sensibly.