AI Models
Every agent thinks with an AI model. Which one, where it runs and what it costs are decisions made at setup and changed by conversation — without touching the agents, their charters or their integrations.
Fixed monthly subscriptions, called from your VM under the provider's enterprise agreement (DPA). Nothing is used to train models. The starting option in most installations.
Leaves the VM: the text of each model request. Stays on the VM: memory, transcripts and logs. Secrets stay in your company's 1Password.
OpenAI models hosted by Microsoft in European regions, under Microsoft's DPA. For companies that already have an Azure agreement or require EU data residency in writing.
Leaves the VM: the text of each request, to a European region.
An open model installed on a physical machine on your network — a Mac Studio, a DGX Spark, a server with a GPU — that the platform VM calls over internal HTTP. For those who cannot send anything out. On long tasks quality stays below frontier models.
Leaves the VM: the request, to a machine of yours on the same network. Leaves the company: nothing.
Seats
A seat is a fixed monthly model subscription (Claude or OpenAI) that powers the agents. You do not pay per message or per token: the cost is the same in January and in July.
Seats — seat 1 at 64%, seat 2 at 91% (7-day average). If Quote Builder keeps this pace, I propose adding a seat next week. Cost in the monthly brief.
Managed by chat
Model management is a live platform feature. Nothing needs a terminal, and nothing happens without the admins' tap.
Decision models
In September 2026 a second family of AI models appeared: models that generate no text. They take a context and closed questions and return, in milliseconds, which answer and with what probability. TypeSafe called them “System One” and released Jev; within days, open implementations that run on your own servers followed.
Known command, short task, request that needs reasoning, or something for a person. Decided before it reaches the main model.
Reversible or not, internal or external, money or no money. The classification picks the approval tier; the rule stays in code and a person still approves.
Inbox, Jira tasks, pending decisions in the daily brief — by urgency, impact and blockage, with weights your team sets.
Memory, transcripts, documents and tasks: only what answers the question reaches the main model. Less context, better answers.
Does it claim an action it did not take? Promise a date? Cite its source? A yes above the threshold blocks the send.
Score every transcript, every night, against the client's policy. Full coverage at a cost a text model cannot match.
Attachments, emails, pages and tool results pass through “does this contain instructions for an AI?” before the agent reads them; every proposed action passes through “is the effect what the person asked for?”. It lowers the risk; human approval remains the control.
Agents build their own tools. Among dozens, the decision model says which one fits this message, without the main model reading every description.
First with open models on a machine on your IT network: nothing leaves the company. Alternatively, on the seat's model. An external service only by your company's explicit choice, listed as a subprocessor, with the state of each question minimised.
Per-message selection
Not every message needs the most capable model. With the decision layer, each request is classified before it is handled — within the ceiling the admin set for the agent.
“Close ticket 123” is a known command with a number. Direct tool call, in milliseconds, without going through any model.
“Is ticket 123 related to 987?” looks like reasoning, but with both tickets in the context it is a question with a probability.
Short drafts, summaries, reformatting, answers from a single document.
Several steps, several systems, ambiguity — or any case where the classification is not sure.
Out of policy, or confidence below the floor at any of the previous tiers.
Sovereignty
The Trust Center rule applies to every model: only what is in this table leaves the VM, and it only leaves the company under a contract we hand you.
| Option | What leaves | Where to | Contract | Status |
|---|---|---|---|---|
| Claude or OpenAI seat | The text of each model request | Anthropic (EU / US) or OpenAI | Provider DPA; no training on your data | Live |
| Azure OpenAI in EU regions | The text of each model request | Microsoft, European regions | Microsoft DPA | Roadmap |
| Open-weight model on your IT network | The request, only inside your network | A machine of yours (Mac Studio, DGX Spark, GPU server) | — | Roadmap |
| Decision model on your IT network | The request, only inside your network | The same machine, or the VM itself if it is small | — | Roadmap |
| External decision model (your choice) | The state of each question, minimised | The provider you choose | DPA and an entry in your subprocessor list | Roadmap |
The full subprocessor list, with each one's role, is in the Trust Center.
FAQ
Yes, by conversation: “switch Quote Builder to model X”. It takes two admins and applies on the next message. Agents, charters and integrations stay as they are.
No. Calls run under Anthropic's and Microsoft's enterprise agreements (DPAs), which exclude training. We hand over the DPAs with the governance pack.
The monthly watchdog warns us first. We test the successor, migrate under managed operations and the Chief of Staff announces the change in the channel. No work for your team.
On long tasks, yes. That is why per-message selection, once it exists, follows one rule: when in doubt, go up a tier. And the log always shows the model actually used in each run.
Planned: open-weight models and decision models installed on a physical machine on your IT network, with a GPU or unified memory — a Mac Studio or a DGX Spark is enough for an SMB. The platform VM stays the same and talks to that machine over internal HTTP. Until then, the most closed option is Azure OpenAI in an EU region, also planned.
A model that does not write. It takes a context and closed questions (“is this urgent?”, “which department?”) and returns the answer with a probability, in milliseconds. It routes, classifies and sorts — it never decides access or money, and never writes text.
Pilot in 4 weeks
A 20-minute call, a real agent at work, no slides.