Give the whole team AI, and decide which AI

Most teams want everybody to have AI and nobody to have a surprise bill. Those two wishes fight each other the moment the only option is one vendor's flagship on every seat. Profusia takes a different shape: the AI inside it answers from what your team actually keeps here — the documents, the projects, the live tables — and an admin decides which models it may use to do that, for whom, and how hard they may think.

What an admin controls

An approved-model list for the workspace. It is a filter over the models the deployment offers, so approving never adds a model and every model on it has a published price beside it.

The same list, narrowed for a group and narrowed again for a person. Select any number of groups and people and add or remove models for all of them in one move. A person can only ever have a subset of what their groups have, which is a subset of what the workspace has.

An effort range per model, for the models that can be asked to think harder. Hold everyday work to medium and grant high to the people whose questions warrant it.

Whether people see what each answer cost — for everyone, for a group, or for one person. When they do, every answer carries a line naming the model and the rough cost, which is the cheapest way I know to make a team economise.

What a person gets

A conversation drawer on every module — documents, projects, live data, audience — that reads what is in front of them and says which documents it actually read. If they have more than one approved model they pick one; if the model takes an effort they pick that too. A pick outside what they are allowed is refused in a sentence, never swapped.

The same assistant they already pay for, if they have one. Claude, ChatGPT and other MCP clients connect to the workspace directly and do the publishing, filing and planning from the outside. The in-app AI is for the people reading; the connected assistant is for the person making.

What it is not

Not a benchmark. The page shows each model's published rates, a modelled cost for one everyday answer, what the workspace actually spent this month by model and by person, and a handful of computed recommendations. It also shows independent market prices, drawn from OpenRouter's public list and named as theirs, with links out to independent comparisons — because I would rather point at a good third-party page than maintain a worse copy of it.

Not a way to route around the operator. The deployment decides which providers are connected at all; the workspace decides who may use which of them. Nothing a workspace admin does can make a model appear that the deployment has not connected.

Not a lock on your data. A model on a provider tier that may train on what it is sent is refused for a workspace that has not accepted that, and the refusal says so.

Which plan

The workspace-level approved list, the effort ceilings, the cost switch and the model picker are on every paid plan. Narrowing by group and by person starts with the Team size, because a one-person workspace has nobody to narrow. Members are per seat on Team and Business, so there is no ceiling to hit. How much in-app AI is included is the tier: Efficient for a team that connects the assistant it already has, Equipped for a real monthly allowance with nothing to manage.