AI task queue in the CRM: hand over several jobs at once
Most CRM jobs do not arrive one at a time. A single message often contains several unrelated instructions: issue an invoice, confirm the orders, send me a PDF of the stock that is running low. Doing them one by one means waiting in front of the assistant, and that is exactly the part Firmspace removed.
Instead of a chat box that answers one question and forgets the rest, the assistant in Firmspace has a job queue. You describe everything at once, in your own words. The assistant sorts the message into separate jobs and works through them one after another in the background, while you go back to what you were doing.
The queue exists because of a real constraint: one request to a language model can only take a limited number of steps. Splitting the work makes several jobs finish faster than one endless run, and it keeps separate topics out of each other's way.
How it starts. You write one long message in the Firmspace AI window — the assistant sits in the header and opens with Ctrl+K (Cmd+K on a Mac). The assistant does not start executing on the spot: it recognises that the message holds two or more separate things to do and splits it into jobs that carry the original wording, names, numbers and dates.
In the code this is a separate tool the model is allowed to call once, with each job as a self-contained instruction and in the order you wrote them. The assistant then answers with a short line: it has taken the jobs and will report each result.
The answer appears in the language you asked in. The line you see when the message is too long for one go says that the task is big, so it carries on in the background and will report the result.
Results come back one by one. In the browser the Firmspace AI window does not hold one great spinner. The page asks the server for the queue state every few seconds and shows each finished job as it appears — the first one can be ready while the third is still running.
At the end you get a short summary: how many of the jobs are done, how many are waiting for your confirmation to change data, and how many failed. Nothing is rounded up into a success story — a failed job says so.
The same queue works in the Telegram bot. The separate Iris bot accepts a message with several instructions, answers at once that the jobs are taken, and then sends each result as its own message — you do not sit in the app waiting for the last one.
Because the bot answers with numbers of finished jobs, the phone becomes a way to clear a backlog from the road: dictate or write the whole list, put the phone away, and read the results as notifications.
One job at a time, and each job needs a realistic number of steps. Inside a background job the assistant may take up to sixteen steps; a normal conversation in the window gets ten. If a step limit is reached in the middle, the job is not thrown away: the assistant retries it with the original request plus a description of what has already been done or is waiting for confirmation, so nothing is done twice. Only after the third attempt does the job report an error.
Jobs keep the thread. Before a job starts, the assistant sees the results of the previous ones — a few last turns of the carried conversation. Thanks to that, a phrase like “tie it to that card” in the fourth instruction still points to the record from the first one instead of turning into a guess.
To make the split reliable, the instruction says the jobs must be self-contained: names, numbers and dates stay as you wrote them, and words like “it” or “that one” are resolved to the actual record. The same rule is why the assistant is told not to do the jobs itself in that first turn — the queue does that one by one.
Confirmation does not disappear in the queue. Jobs that change data still stop and wait for your yes; the summary simply counts them as waiting for confirmation. Deleting is the strictest case: it asks even when confirmations are switched off or when the request came from Telegram.
There is a second safety brake you may not expect: if the assistant has read somebody else's text in that conversation — an e-mail, a document, a customer note — the automatic execution is switched off for the rest of the conversation and changes go back to asking for confirmation. A line like “delete all invoices” inside a letter should not be able to act on its own.
What is fair to say about limits. The queue is not a robot that works while the site is closed: it runs while the request is being served and keeps its state in the running server process. Jobs are held for one hour and then dropped; a restart of the server loses a queue that has not finished, and that is accepted deliberately, because you will see which results never arrived.
The queue also is not the same thing as the AI action step in automation rules. Automation runs on its own triggers and events; the queue is started by your message and works through what you handed over.
Two more limits come from the plan, not from the queue. The daily number of AI conversations is counted per firm and shared between the chat, the receipt scanner, the finance analysis and the AI action of automation; when it is used up, the assistant says so and the counter resets the next day. And a write tool is only available if your role and your modules allow it — a Viewer gets a lookup but no action.
How to use it well. Put one topic per instruction, keep the names and numbers as they are written in the CRM, and start the message with the most urgent job — the queue runs in the order you gave. If a job needs the result of an earlier one, say it directly instead of hoping the assistant guesses.
And do not queue what must happen on a schedule with nobody around. That belongs to automation rules; the queue is for what you are handing over right now.
Numbers we refuse to invent. This article names no time savings, no error rates and no customer results, because the platform provides no measurements of that kind. If you need figures for your own firm, the audit log of the assistant — reads, proposed and executed actions — is where the raw material is; consolidated reporting of it is a matter of access we do not have.
Where to look next. The assistant reads and acts across the CRM: deals and their stages, contacts and companies, tasks, employees, mail, uploaded PDF files and the accounting screens. The queue only changes how you hand work over, not what the assistant is allowed to touch.