Appearance
AI settings
This page is for admins: decide where AI runs in your workspace, which provider it uses, what it costs and what the assistant may do.
For what members see, read AI in Connectdary. Everything here is under Settings, AI. Only owners and admins can change it.
INFO
On the Free plan there is no AI. Where plans are on, it starts on the Team plan; your own provider needs Business. Servers without plans have no limits.
Turn AI on or off
Marco wants AI for the team, but not call transcripts yet.
- Open Settings, AI.
- Check that AI in this workspace is on. Turning it off stops every AI feature at once, and nothing is sent to a model.
- In the list of features, switch off Call transcripts and notes. The others keep working.
The features you can switch are catch-up summaries, ask the workspace, smart inbox, decision detection, topic summaries, call transcripts and notes, composer assist, and the TL;DR of long messages and files.
Each person can also turn AI off for themselves with Use AI for me. That wins over the workspace setting, but a person turning it on can't override an admin who turned it off. If someone turns AI off for themselves, what AI already wrote from their words is no longer served to others.
Choose an AI provider with your own key
By default Connectdary uses the AI provider of the server it runs on. You can bring your own, so every AI answer in your workspace comes from it and the prompts go where you chose. This needs the Business plan.
Ana, the owner of Acme, wants Anthropic with the company's own key.
- Open Settings, AI, then Provider.
- Choose Anthropic. The other choices are OpenAI, OpenRouter and an OpenAI-compatible server. The server may allow only some of them.
- Paste the key under API key. It is sealed when saved, and afterwards only the last four characters show.
- Press Load models. Pick or type a model for each of Quick (fast, cheap work such as summaries and TL;DR), Smart (harder work such as Ask, decisions and the assistant), Embeddings (search by meaning) and Transcription. They start on the provider's default.
- Press Save provider. Connectdary tests the key first. If the provider refuses it, nothing is saved and you're told which model failed.
For an OpenAI-compatible server you also enter its Server URL. It must be a public https address. Private addresses and plain http are refused.
What changes for members:
- Everyone can read which provider the workspace uses on their privacy page, and on the AI trust card, which also shows hosting and retention. It never shows your key or the provider's errors.
- AI answers now come from your provider, and requests that run on your own key cost no credits.
Good to know:
- If your key later stops working, AI stops in your workspace. Connectdary doesn't switch to the server's provider behind your back. Members are told "Your workspace's AI provider isn't working. An admin can check it in Settings, AI." You see the latest provider error under Provider.
- The server's operator can forbid workspace providers or allow only self-hosted AI servers. You're told when that applies.
- Things your provider can't do, such as call transcription, may still use the server's provider.
- On the phone, the provider is shown but changed on the web.
AI credits
Where plans are on, what people ask AI for uses a monthly pool of credits shared by the whole workspace:
| Plan | Credits per billable member each month |
|---|---|
| Team | 300 |
| Business | 1,000 |
Acme is on Team with 8 billable members, so it has 2,400 credits this month. Under Settings, AI, Marco sees a line like "1,240 of 2,400 AI credits used, resets 1 November", and the month's use per feature. When Chen joins, the pool grows to 2,700 straight away. The pool resets on the 1st.
What costs credits:
- a quick action, such as a TL;DR: 1 credit;
- a smart action, such as asking the workspace a question: 3 credits;
- call transcription: 1 credit per started minute, so a 4-minute call costs 4.
Free of charge: what AI does by itself (catch-up, inbox triage), search, and a failed request. Ask on your own AI key costs no credits.
When the pool is empty, what people ask for stops until the 1st, and they're told to ask an admin. A smart action stops when fewer than 3 credits are left, while a TL;DR can still fit. Automatic AI keeps working. Members see a warning once 80% is used. Other members can't see usage.
TIP
If credits run out often, either bring your own provider key (Business), or switch off the features you use least.
Which actions the assistant may take
The assistant can propose actions for people, such as creating tasks, setting reminders or scheduling messages. Each person confirms before anything happens, unless they let it run simple actions on its own. You choose which kinds are allowed.
Marco allows tasks but not scheduled messages:
- Open Settings, AI, What the assistant and agents may do.
- For each kind of action, choose On or Off. It saves at once. All kinds start On.
- Set scheduled messages to Off.
Now the assistant stops offering scheduled messages, AI agents that members connected are told it is off and can't do it, and proposals already made can't be confirmed. Ana sees it "turned off by Acme's admins".
The same choice applies to AI agents connected over MCP. An agent whose kind is on acts directly as the person it works for. It shows with its name in that person's activity, with Undo where there is one. Members can set an agent to ask them first. See Apps and integrations and AI agents.
TIP
Sending and scheduling messages always ask the person first.