Enterprise
Deployment options
Where OpsAI can run, what each option means for data movement and latency, and what an enterprise network deployment actually involves.
The only question that matters first
What leaves your network?
Not the feature list. Every option below is described by what crosses the boundary, because that is what an architecture review opens with and what a tier table hides.
Options
3
Where nothing leaves
2
Three shapes
Described by what crosses your boundary, not by a feature column.
A deployment page is usually a pricing table wearing technical clothes. What an architecture review needs is what leaves the network under each option, what running it actually involves for your team, and the tradeoff — which is the part a tier table never has.
- 01
Managed
available- What leaves your network
- Action metadata and decision records. Never a credential, never the contents of a read, never a model prompt or completion.
- What it involves
- An API key and a connection per system. No infrastructure, and rotation and upgrades are ours.
- The tradeoff
- Decision records sit in our tenancy. For most estates that is acceptable and for some it is the whole objection, which is why the other two exist.
- 02
Private cloud
available- What leaves your network
- Nothing. The control plane runs in your own cloud account, and records never leave your tenancy.
- What it involves
- A managed deployment into your account, with upgrades applied on a cadence you approve rather than on ours.
- The tradeoff
- You carry the infrastructure and the upgrade window. Latency improves, because the boundary sits next to the systems it governs.
- 03
Self-hosted
available- What leaves your network
- Nothing, including telemetry. Air-gapped operation is possible.
- What it involves
- You run and upgrade it. Expect to own capacity planning, backup and the evidence-store retention policy.
- The tradeoff
- The most control and the most operational load. Worth it where a regulator requires it and rarely worth it otherwise.
Two questions that get conflated
Where OpsAI runs and where the model runs are different decisions.
This confusion wastes an entire review cycle when it goes unnoticed. Self-hosting the control plane does nothing whatsoever about a model that answers from a vendor API — the request content still leaves, and it leaves from a path OpsAI is not on.
- Where OpsAI runs
- A deployment choice, and the one this page is about. It determines where decision records sit and how far an evaluation has to travel.
- Where the model runs
- A property of each model, not of OpsAI. 4 of 7 models in the sample estate never send request content to a third party — self-hosted or on a private endpoint.
- Where the data sits
- A third question again, and a fact about your infrastructure. The sample estate spans 3 residencies, which is the figure a data-protection review asks for.
- Why this matters commercially
- A team that self-hosts to satisfy a residency requirement, while every model call goes to a vendor API, has taken on operational load and solved nothing.
What OpsAI never holds, under any option
A credential. Held in the vault under managed, held by you under the other two, and returned by no endpoint in either case.
The contents of a read. What is recorded is that a read happened and what it was permitted to do.
A prompt or a completion. OpsAI is not on the inference path — it decides whether an action may proceed.
A copy of your data. There is no sync, no index and no cache to go stale or to leak.
Which means the managed option moves less than a reviewer usually assumes, and is worth establishing before the conversation defaults to self-hosting. Why nothing is copied.
What deployment does to the decision
The boundary is in the path of a user-visible action, so distance costs something.
An evaluation happens while somebody waits. That is why no model call sits in the decision path — and it is also why where the control plane runs relative to the systems it governs is an architectural decision rather than a preference.
Decision latency
11ms
the canonical action, evaluated
Model calls in that path
0
deterministic evaluation
Options where records stay in your tenancy
2
private and self-hosted
Options requiring a code change
0
none of them
IllustrativeThe canonical action from the OpsAI sample estate. How the trace is built.
A control fast enough that nobody is tempted to route around it.
This is the real reason latency belongs on a deployment page. A control that adds visible delay gets bypassed — not maliciously, but by somebody under pressure who adds a code path that skips it. Placement is how you avoid ever having that conversation.
Status, not claims
Trust center
Certification status stated plainly, the properties behind each claim, and what OpsAI will not do.
How it is built
Security at OpsAI
Where the control points sit, and how identity, credentials and execution are constrained.
For an engineer
Security documentation
Authentication, workload attestation, tenant isolation and the data boundary, as mechanisms.
Where to start
Establish what actually leaves before choosing where to run it.
Most reviews reach for self-hosting first and discover afterwards that the managed option moved less than assumed — and that the model, not the control plane, was the thing sending content outside.