AI: automatic model choice, costs and the monthly limit
How Operacora picks the cheapest AI model that does each task well, what it costs, where to see it, and how to cap your monthly AI spend.
On this page
AI in Operacora is optional and runs on your organization's own AI account, usually OpenRouter (see Connections). You do not have to choose a model. For each task, Operacora picks the cheapest model that does that task well, by itself, and keeps a monthly limit so AI can never cost more than you allow.
AI only suggests; people decide. AI never assigns work, never changes a priority, never sends a message and never changes what someone wrote. Anything a manager sets by hand stays as the manager set it.
What AI does today#
| Task | What you see |
|---|---|
| Summary and tags for new requests | Reports from the phone app (and requests by text and email, once those are connected) get a short one-line title, suggested categories and a work type. The title is marked Suggested by AI. The original description is always kept exactly as sent. |
| Reading a floor plan | Find rooms on a plan picture drafts rooms for you to check (see Floor plans). |
| Voice notes | Turning a voice note into text, with your Groq connection (the cheapest), or OpenAI's if you have no Groq key. |
| Translation and thread summaries | Ready in the engine; the screens that use them are Coming. |
A request is always saved first, without waiting for AI. If AI is not connected, is over the limit, or gives an answer that does not fit, the request simply stays as it came in.
How the model is chosen#
Each task says what it needs: how much text it reads, whether it must answer in a fixed format, whether it reads pictures, and how good the model must be (basic, good or strong). Operacora reads OpenRouter's public price list once a day and, for each task, takes the cheapest model that:
- is on our checked list for that quality level, so "cheapest" never means an unknown, low-quality model;
- can do the task (enough room for the text, the right answer format, pictures when needed);
- stays under that task's price limit for a single call.
Free models may be used for simple tasks, always with a paid model behind them, because free models are often busy. Every call also names one or two backup models, and has a cap on how long the answer can be. Every answer is checked before it is used; an answer that does not fit is tried once more on a better model, and otherwise dropped.
Use one model for everything#
On the OpenRouter, Claude or OpenAI card, Model choice is Automatic (recommended). If you have a reason to, choose Use one model for everything and type the model's name. Every task then uses that model where it can (a model that cannot read pictures is not used for plans). If you had typed a model before this choice existed, it keeps working as your one model.
Claude or OpenAI instead of OpenRouter#
If you connect Claude or OpenAI directly and not OpenRouter, Operacora picks the cheapest suitable model from that provider in the same way.
See what it picks, and what it costs#
Open Settings › Connections and, under AI, open See what it picks. For every task it shows the model it would use right now and roughly what 1,000 of them cost at today's prices. A free model is priced at its paid backup, so the estimate is never too low. Below that you see this month's AI spend, the number of AI calls, and your monthly limit.
The amounts come from your provider's own usage figures where it gives them (OpenRouter does). Your provider's bill is always the final word.
The monthly limit#
Each organization has a monthly AI limit, $5 a month to start. Before every call, Operacora checks that the call still fits; if it does not, the call is refused with a clear message, nothing is charged, and everything else in the app keeps working. The limit starts again on the first of the month (UTC). Change it under See what it picks; set it to 0 to turn AI off completely.
Privacy#
- Your AI key is stored encrypted, never shown again and never sent to browsers.
- Operacora keeps a short record of each AI call: which task, which model, how many tokens, the cost and whether it worked. It never keeps the text that was sent or the answer.
- Text in requests is treated as data, never as instructions: a message that says "ignore your rules" cannot make AI do anything beyond suggesting a title, categories and a work type from your own lists.
- One organization's AI account, usage and limit are never used for or shown to another.