Security
The first question a serious business asks is where its data sits and what a colleague is allowed to do with it. Here are the answers, described structurally rather than as a promise.
A workforce is only worth having if you can trust it with the business. Trust here is built into the structure of the platform — isolation, default-deny tools, approvals and an audit trail — not bolted on as a policy you have to take on faith.
01 — Isolation
Atelier is multi-tenant with a database per customer. Your firm's data lives in its own database — not a shared table with a filter and a promise in front of it. The isolation is structural, so one tenant's work, knowledge and audit trail are physically separate from the next.
Where data must stay on your own infrastructure, a self-hostable edition keeps the whole fabric inside your walls.
02 — Least privilege
Every colleague starts with zero access. A tool is available only once it has been explicitly assigned — there is no fallback that quietly grants everything, and no way for an unassigned colleague to reach a system you did not intend. Access is something you grant deliberately, and can withdraw the same way.
03 — Approvals
Sensitive actions do not happen unattended. A colleague proposes the step, a named person approves or declines it, and only then does it run. The workforce carries the work up to the point of consequence, and stops there for a human decision — every time.
04 — Audit
Every message a colleague sends and every tool action it takes is logged and auditable, with the colleague, the context and the outcome attached. Administrators hold per-colleague controls, including the ability to pause a colleague entirely. Nothing the workforce does is off the books.
Practices
The platform is built to SOC 2 practices — least privilege, audit, change control and access governance carried through the way the product is designed and operated. We describe this as the practice we hold ourselves to, and will say "certified" only when it is.
MCP hardening
Colleagues act through your systems over MCP — open and inspectable rather than a black-box integration. Those connections are hardened with agent-bound tokens: a credential is scoped to a single colleague, so access can be reasoned about, attributed and revoked one colleague at a time.
Book a session with us and put the platform's isolation, approvals and audit model to the test against your own requirements.