Enterprise / On-premise AI agents

On-premise AI agents: private AI, deployed where your data lives

For IT and operations leads in energy, banking and industrial operations whose documents and data cannot go to a public SaaS. Mirage deploys its AI agents on your premises, in your own cloud account, in a private or sovereign cloud, or at the edge and offline. They read your documents and systems where they are, write only what each workflow allows, and Mirage's access stays limited to what the deployment needs.

Being deployed inside Mass Céréales Al Maghreb's own operational environment, across two port terminals.

The reality

The documents and data your operation runs on are covered by policies that keep them inside your network: contracts, operating records, supervision and plant data, customer files. Most AI agent products are delivered as a public SaaS that takes the document to the vendor's cloud. So the security review stops the project, or the work stays manual, retyped by the people allowed to see it.

What Mirage does

What an on-premise deployment of AI agents involves

Three questions decide whether a private deployment holds: what runs inside your perimeter, how the agents reach your systems, and what they are allowed to write.

What runs inside your perimeter

In a deployment in your infrastructure or your own cloud account, the agents' data and processing stay in the environment you control, and nothing is transferred outside it. The audit record stays there with the rest of the data, and the encryption keys are yours. Each deployment is isolated: its own environment and its own data, shared with no other client.

How agents reach your systems and documents

Through the interface each vendor provides: an API, a file exchange, eAdaptor for CargoWise, data entities over OData for Dynamics 365. Email inboxes, PDF and scans, spreadsheets and EDI files are read where they arrive; telemetry and supervision systems through the approved interfaces of the operating environment. Where a system has no API, Mirage builds the connector during the deployment. Nothing is migrated.

Approval before anything is written

What an agent may write on its own, and what waits for a named person on your team, is set for each workflow, with access rights by role. Every action is logged with its inputs, its sources, the rule it applied, its result and, where a person decided, the approver's name.

Who operates it, what Mirage can reach, how it starts

The questions an IT team asks before a private deployment, answered from what Mirage commits to on its security and enterprise pages.

Who operates it

Forward-deployed engineers build each workflow inside your systems, with your team, then Mirage operates, monitors and improves it in production. Your team works the review queue and approves what is written.

What Mirage's access is

Strictly limited to what the deployment needs, or removed after deployment, depending on the architecture you choose. Your administrator decides who holds each role, and you can remove Mirage's access.

How it is scoped and piloted

A non-disclosure agreement on request, and answers to your security questionnaire for the architecture you are considering. Then a proof of concept on your own documents and data one week after the first meeting, a fixed-price pilot on one workflow that you can stop after, and the first workflow in production in 4 to 6 weeks.

When on premises is the wrong choice

If your policy accepts a cloud account your organisation controls, data, processing and keys stay with you there too, with no servers of your own to run on site. If the only concern is training, no deployment trains a model on your data, cloud included. On premises earns its place when documents may not leave your site; edge and offline, when the site has no reliable connection.

DEPLOYMENT OPTIONS

Which deployment option answers which constraint?

The six options Mirage lists on its security page. Who holds the data and the keys follows what the site commits to: in your infrastructure or your own cloud account, they stay with you. Where the site makes no commitment, the row says so; the option is chosen with you, against your requirements.

Which deployment option answers which constraint?
OptionWhere the agents runData and keysThe constraint it answers
Cloud, hosted by MirageAn environment Mirage hosts, isolated for your deploymentSet out for this architecture in the answers to your security questionnaireYour policy allows documents to be processed in a vendor-hosted environment
Your cloud accountYour organisation's own cloud accountStay in your account; the encryption keys are yoursCloud is accepted, but only in an account you control
Private cloudA private cloud, within the infrastructure you controlStay in your environment; the encryption keys are yoursProcessing must run in a private environment, not a shared public cloud
Sovereign cloudA sovereign cloudSet out for this architecture in the answers to your security questionnaireData must stay with a provider under a given jurisdiction
On premisesServers on your own siteStay on your premises; the encryption keys are yoursDocuments and data may not leave your site
Edge and offlineAt the operating site, able to run without a connectionStay at the site; the encryption keys are yoursA remote or isolated site with little or no connectivity

In every option: your data never trains a model, each deployment is isolated, every agent action is logged, and access rights are set by role. Where a deployment calls a model provider, that provider works under a zero-retention agreement.

The output

An on-premise deployment, component by component

An illustrative deployment checklist, with invented values: what runs where, who controls it, and what is still waiting on a decision.

ComponentWhere it runsWho controls itStatus
AgentsServer in your data centre, site 1Your IT teamRunning
Connector to the ERPYour network, through the ERP's APIYour ERP administratorAccess granted
Connector to the document archive↳ The archive offered no API. The connector was built during the deployment, so nothing was migrated.Your network, connector built during the deploymentYour IT teamBuilt
Inbox for incoming documentsYour mail serverYour IT teamConnected
Review queue and approvalsYour environment, behind role-based accessAn approver named for each workflowApprover named
Audit recordYour environment, with the rest of the dataYour organisationRecording
Encryption keysYour key managementYour security teamHeld by you
Mirage engineer access↳ Mirage's access is limited to what the deployment needs. Your administrator decides whether it stays limited or is removed after go-live.Through your remote access processYour administrator, removableYour decision at go-live

Proof

Deployments inside the client's own environment

No Mirage case study describes the hardware or hosting of a deployment, and none is published as an on-premises server installation. The two nearest, labelled as what they are: at Mass Céréales Al Maghreb, the platform is being deployed inside the client's own operational environment, across two port terminals; at Transwin, five systems run in production, including a private assistant grounded in Transwin's own documents and operational records, with no hosting described.

Mass Céréales Al Maghreb (deployment in progress)

Deployed inside the client's own operational environment

2 port terminals, Casablanca and Jorf Lasfar

1 control layer across both terminals

0 core operational systems replaced

See the case study

Transwin (hosting not described in the case study)

5 systems in production

A private assistant grounded in Transwin's own documents and records

70% less manual data entry on freight documents

See the case study

An on-premise deployment, scoped on your constraints

Tell us which systems and documents the agents would work on and what your policy allows. We reply within two business hours, and at the first meeting we name the deployment option, what runs inside your perimeter and the access Mirage would need.

Attach your security questionnaire (recommended)

We use your email to reply to you and, now and then, to tell you about Mirage. You can opt out at any time. Privacy policy

How a deployment runs: Observe, Map, Build, Run, Expand. See the method

FAQ

Frequently asked questions

  • An AI agent that runs in infrastructure your organisation controls instead of a vendor's public SaaS. It reads your documents and systems where they are, prepares the work and writes back only as your workflow allows, with data and processing staying in your environment. Mirage agents can run on your premises, in your own cloud account, in a private or sovereign cloud, at the edge and offline, or in Mirage's cloud.

  • In a deployment in your infrastructure or your own cloud account, data and processing stay in the environment you control and nothing is transferred outside it. Your data never trains a model, ours or anyone else's, and any model provider a deployment calls works under a zero-retention agreement. This page names no model or hardware: the architecture is chosen with you, against your requirements.

  • Through the interface each vendor provides, such as an API or a file exchange, and through the inboxes and files where documents arrive. Where a system has no API, Mirage builds the connector during the deployment. Your systems stay in place and nothing is migrated.

  • Mirage's access is limited to what the deployment needs, or removed after deployment, depending on the architecture you choose; your administrator decides who holds each role. Mirage's engineers build each workflow with your team, then operate, monitor and improve it in production within that access.

  • Only where your workflow allows it. What an agent may write on its own, and what waits for a named person on your team, is set for each workflow, and every action is logged with its inputs, sources, rule, result and approver.

  • Mirage has not published a case study of an on-premises server installation. At Mass Céréales Al Maghreb, the platform is being deployed inside the client's own operational environment, across two port terminals. At Transwin, five systems run in production, including a private assistant grounded in its own documents and records; the case study does not say where it is hosted.

  • When your policy accepts a cloud account your organisation controls: data, processing and keys stay with you there too, with no servers to run on site. And when the only concern is training: no deployment trains a model on your data, cloud included. On premises fits documents that may not leave your site; edge and offline fits a site without a reliable connection.

  • A proof of concept on your own documents and data one week after the first meeting, and the first workflow in production in 4 to 6 weeks. A fixed-price pilot on one workflow first, which you can stop after. Then an annual subscription priced by volume, quoted in the first conversation. Licences for your own systems, and hardware where a deployment needs it, are named in the quote.