CAI for business
The sensitive materialnever leaves the building.
Client files, case notes, patient records, source code — the data your obligations are written around stays on hardware you own, because the model runs on it.
The problem
Most AI asks you to send the file away.
Every regulated trade has the same shape of rule: the sensitive material stays where you can account for it. A legal practice has privilege. A clinic has patient records. An accountant has client finances. A software team has code it is contractually forbidden to disclose.
- A data-processing agreement is a promise. It describes what a vendor intends to do with your data once they have it. It does not change the fact that they have it.
- "We don't train on your data" is a policy. Policies are set by whoever owns the servers, and they can be revised.
- The upload is the event you can't undo. Once the file has left, retention, jurisdiction and breach exposure stop being yours to decide.
Running the model on your own machine isn't a stronger promise. It removes the need for one.
The architecture
Nothing leaves, because nothing has to.
These are properties of how CAI is built, not settings you have to remember to switch on.
- Inference runs on your hardware. Local models through Ollama, on a box in your building. There is no cloud dependency unless you deliberately add one.
- Routing is deterministic — no model decides. Which model handles a request is settled by rules you can read, with per-slot budgets, cooldowns and failover chains. There is no LLM anywhere in the routing path.
- Credentials live in an encrypted vault. API keys are sealed and scoped per user. Extensions and packs call through the vault and never see the secret itself.
- Row-level security in the database. Postgres RLS is enabled and forced on every user-scoped table, with schema ownership held by a separate role from the application.
- Audit trails on privileged actions. Stopping or deleting an application is recorded, not silent.
- Speech is processed on the box. Local text-to-speech and transcription. No audio leaves your network.
Your models can still be cloud models if you want them to be — CAI speaks to Anthropic, OpenAI, Gemini and Groq as readily as to a local one. The point is that it is your decision, per agent, not the platform's.
The Enterprise pack
Governed by default.
CAI's capabilities ship as packs. The Enterprise pack is the one built for organisations, and what it leaves out matters as much as what it includes.
01 Collaboration
Shared project spaces
Work that belongs to a team rather than to one person's account, with the isolation still enforced underneath.
02 Retention
Governed data stores
No-delete stores, so an agent can write to the record and never quietly remove from it.
03 Documents
Sheets, slides and docs
Agents produce real deliverables — spreadsheets, presentations and documents — rather than text you then have to reformat.
04 Integration
HTTP integrations
Reach the systems you already run, over your own network, with the credentials held in the vault.
05 Reporting
Database reporting
Query your own databases and turn the answer into something a person can act on.
06 Scheduling
Scheduled work
Jobs and reminders that run on a schedule, so the routine reporting happens without anyone remembering to ask.
The Enterprise pack ships without shell and code-execution tools. That is a decision, not an omission.
Those tools exist — they are in the Developer pack, and they are the right thing for an engineer's workstation. They are the wrong thing for an agent with access to a shared corporate data store.
Separating them by pack means the governed toolset is the default for organisational use, and granting more is a deliberate, visible act rather than something that arrived switched on.
Scaling
One box, a rack, or several sites.
The architecture doesn't change as you grow — the hardware under it does.
01
Small
A single machine in the office running local models for a handful of people. The whole platform fits on one box.
02
Medium
Rack hardware with the GPU capacity to keep larger models resident, serving a department without queueing.
03
Enterprise
Per-user isolation, per-user credentials and audit trails across a larger estate — the controls that matter once "the team" is several teams.
At every size, CAI reads the VRAM, RAM and CPU it has been given and tells you what to expect from a model — including when the answer is "not on this hardware" — rather than failing quietly once it is running.
Start a deployment
Talk to us about your estate.
Tell us the shape of it — how many people, what they handle, and what has to stay inside — and we'll scope the hardware and the deployment around it.
Want to try it before you talk to anyone?
Download CAI free and run it on one machine →
Need the network, cameras and cabling too?
That's the installation side →
Start a deployment
Tell us the shape of your estate and we'll scope it. No form, no sales sequence — an email reaches a person.