OdyControlOdyControl, the sovereign AI platform by Cybervize

The sovereign AI platform you run yourself.

OdyControl combines a browser-based AI workplace for all employees, language models in your own network and a control layer. You decide which requests stay in-house and which may go to a public model your organisation has approved. Calls are documented in a traceable way.

Local AI first. External models only on your terms.

Available

OdyControl runs in production at Cybervize and is installed in your environment. Which models, applications and data classes you connect is what the conversation is for.

Three building blocks, run in your own organisation

One platform instead of a patchwork of AI tools

Many organisations now have one AI tool per department, each with its own access and its own rules. OdyControl replaces this with one workplace, one access point and one control layer. It brings what your employees use every day, what does the computing in the background and what decides where a request may go.

AI workplace in the browser

A chat most people already know: draft texts, translate, analyse long documents, assistants per team. Sign-in runs through your directory, and even admin rights give no view into conversation histories.

Local language models

Current open models run on hardware in your network. There is no billing per request: you plan hardware and operation, and costs do not grow with every question. A new model only goes into operation once your organisation has reviewed and approved it.

Control layer with approvals

For each data class you decide whether a request stays local or may go to a public model your organisation has reviewed and approved. Calls are logged, to the extent you determine. In the “audit-proof” operating profile, OdyControl processes no request without an entry in the tamper-evident log chain.

Your own applications and agents reach the same models through an OpenAI-compatible interface, under the same rules. If nothing is to leave your organisation, you run the platform fully locally.

Not every AI request carries the same protection need

An internal contract, a personnel record, confidential source code and a draft for a trade fair text must not be treated under the same rules. Yet in many organisations they are: either every request goes through the same external service, or cloud AI is blocked outright.

Everything goes out

A single external service receives every question, regardless of what is in it. In the end nobody knows exactly what left the building.

Everything is blocked

Cloud AI is forbidden, and the ban rarely holds. Anyone facing a deadline pastes the paragraph into a browser. The ban turns into shadow AI.

The route in between

Whether data may leave the organisation becomes a routing decision you set once and that then applies to every request.

OdyControl turns data sovereignty into an executable policy. Local models are the preferred route; external models are not a default route but a deliberately approved addition.

Three data classes

Three routes, and you decide which one applies

Your organisation defines the classification. OdyControl enforces the resulting routing rules on the model path.

  • Local Only

    Stays inside your environment

    Trade secrets, confidential research data, proprietary source code, personal data needing special protection, unreleased product information, confidential contracts.

    Only models running on your own infrastructure are permitted: the local primary model, a second model server, a second cluster, local specialist models.

    If no permitted local backend is available, OdyControl ends the request in a controlled way. It is not quietly handed outside. Availability does not override the confidentiality boundary.

  • Trusted External

    Vetted providers only

    Sensitive data whose external processing is permissible for the specific case, for instance under a data processing agreement, a checked processing region and clarified training use.

    Your local models are permitted, plus a restricted pool of specially vetted providers, in an order you define.

    If local AI is unavailable, a new request may go to an approved provider, provided the policy explicitly allows it for this class.

  • Flexible Processing

    Free choice of model

    Publicly available information, general ideation, generic drafts without internal detail, synthetic test data, sufficiently abstracted content.

    Any connected model you have approved for this class is permitted. Selection follows your routing strategy, for example required capability, availability or local capacity.

    If the local model fails, new requests move straight to an external model. They stay answerable without sensitive data taking the same route.

Where the class comes from

Usually it comes from the calling application, the workflow, a fixed policy for the use case or an upstream classification service. In addition, your organisation can switch on optional content routing: a classifier proposes a type from your own taxonomy based on the content. If it is not sure, or the request contains parts it cannot read, the request stays in-house. OdyControl enforces the result.

An example

Two requests, one control point

Two requests at the same time, one from a chat interface and one from a business application. Both pass through the same control point. It decides which model is allowed to answer, spreads the requests across the permitted backends and documents what happened.

Path 1 · Local Only

Chat interface in the browser

An employee asks for a summary of a draft contract. The text contains trade secrets, so the classification is Local Only.

Path 2 · Trusted External

Business application

A business system calls a model through the usual interface to sort free text into categories. For this use case your organisation has reviewed and approved an external provider.

OdyControl

One endpoint that both requests pass through

Mapping
The model name in the request is mapped to a backend that is approved for this data class. The calling application does not need to know about it.
Load balancing
Several equivalent backends share the requests. If one stops responding, the next request goes to another permitted target.
Logging
Timestamp, calling application, data class, selected model and the decision taken go into the log, at the depth you have configured for this case.
Path 1 · stays in house

Local model

A model server in your own data centre answers the request. The contract text never leaves the network.

Your network
Path 2 · permitted, logged

OpenAI

In this example OpenAI is the provider your organisation has approved for this data class. Which one it is at your organisation is your decision. OdyControl checks the approval, permits the call and documents it.

Outside
The example shows two of the three data classes. If no permitted target is available, OdyControl ends the request in a controlled way instead of falling back to a model that is not allowed.
Example of a rollout

Introducing AI without documents leaving the network

How a rollout works in four steps, from the test days to production, using the example of an industrial company with high protection requirements.

To the rollout example

What the platform does

Local first

Process confidential AI workloads on self-operated models inside your own infrastructure.

Policy-based routing

Define which data stays local, which may use approved providers and which models are available for non-critical work.

An AI workplace for everyone

Give your employees an assistant for texts, translations and analyses, with their own assistants per team.

One central endpoint

Connect chat interfaces, agents and applications through a single OpenAI-compatible interface.

Provider freedom

Use local models and different external providers without tying your applications to any one of them.

Traceable use

Document model targets, routing decisions, operational events and, where needed, complete requests and responses.

Cyber Resilience Act

OdyControl and the Cyber Resilience Act

The Cyber Resilience Act (Regulation (EU) 2024/2847) sets security requirements for products with digital elements. Since 11 September 2026 the reporting obligations for manufacturers have applied, and full application follows on 11 December 2027. Anyone choosing an AI platform for their own network today should therefore ask how it is built and delivered.

Secure by default

The route to external models is closed until your organisation approves it for a data class. Without approval, no model request goes outside.

Bill of materials and signature

The delivery pipeline produces a software bill of materials (SBOM) for every version and signs the images. You can check what you are installing before you install it.

Protected access

The audit interface requires sign-in with a second factor and grants rights through roles. Exports and evidence packages need four-eyes approval.

Tamper-evident logs

Logs are written as chains, sealed and anchored with timestamps from two timestamping services. A later change becomes visible.

What is described here is how OdyControl is built. Full application of the CRA begins on 11 December 2027.

More about OdyControl

Your people already use AI. Just not yours.

Use case

The question behind it sounds much the same everywhere: data leaves my organisation every day because people use shadow AI. If I block it, I slow the business down. What am I supposed to do?

  • One in ten

    employees in Germany uses AI without their employer knowing about it.

    Bitkom, study report “Künstliche Intelligenz in Deutschland” 2026, survey of 1,005 people aged 16 and over in 2025

  • No. 3

    is where shadow AI ranks among the most common non-malicious activities related to data leakage.

    Verizon Data Breach Investigations Report 2026, based on data from 2025

  • 20%

    of the organisations studied had a breach connected to shadow AI.

    IBM Cost of a Data Breach 2025, around 600 organisations that had already suffered a breach

  • $670k

    was the average difference in breach cost between organisations with high and with low or no shadow AI use.

    IBM Cost of a Data Breach 2025, difference between groups, not a demonstrated cause

Allowing it is not enough on its own

The best-known case is Samsung in 2023. According to media reports, engineers entered confidential source code into ChatGPT shortly after the company had permitted its use; it then banned it again. The order of events is what matters: use was already permitted, and something still left that should not have. Permitting without a rule on which data may take which route produces the same outcome as no guidance at all. A ban, in turn, tends to move the usage rather than end it: anyone facing a deadline uses the tool that answers.

What your people could do instead

With the OdyControl AI workplace it typically looks like this, in each case on a model you operate yourself. The control layer decides which request may reach which model.

  • Write proposals

    The assistant knows your customers' requirements from curated knowledge bases and inserts standards, specifications and boilerplate from earlier bids, with a reference back to the source document.

  • Translate

    Correspondence, documentation and contracts between languages. As long as the data class stays local, the text touches no external translation service.

  • Draft emails

    Drafts, rewrites and tone adjustments, from a quick reply to sensitive customer communication.

  • Summarise

    Condense long documents, reports, meeting notes and mail threads into a briefing.

  • Work on documents

    Upload and analyse PDF and Office files, analyse data, and set up your own templates and team assistants.

  • Research the web

    Find and compile sources. Here the search query leaves the environment; the content of your documents stays inside as long as the workflow does not put it into the query.

Both halves on one platform

An answer to shadow AI has two halves. One is a workplace your people use willingly; whether it is good enough for them is decided by your tasks, not by a promise from us. The other is a control layer that enforces which request may reach which model, and that can prove what happened. OdyControl brings both halves.

Talk about a pilot

Introduction and operation inside your company: sovereign AI

A policy nobody can enforce is a document

Where it fits

Most organisations now have an AI policy. It sits in a folder while reality runs alongside it: anyone facing a deadline uses whichever tool answers. OdyControl is where the policy becomes a rule that applies to every single request.

Anyone building an AI management system to ISO/IEC 42001 or implementing the requirements of the EU AI Act describes which data may be processed by which systems. OdyControl enforces its outcome on the model path and records what actually happened. Both frameworks are in our catalogue, and OdySecure carries the evidence for them.

To AI security governance

An outage does not move a data boundary

Availability

Ordinary failover asks whether the next system is reachable. For enterprise AI that is not enough. A model can be running and still be impermissible for this request. OdyControl therefore checks two conditions: is the target technically available, and is it permitted for this data class. Only when both hold does the request go there.

A backend fails

OdyControl detects the state, removes the backend from active rotation and stops sending new requests to it.

Recovery

An automated restart of the model serving process is triggered, the state is checked again, and after a successful check the backend returns to rotation.

Second cluster

Where the local infrastructure has at least two independent clusters, new requests keep running on the healthy cluster during an outage, including Local Only ones.

Return

Once the local backend passes its health check it becomes the preferred target again, depending on the routing strategy. Local first in normal operation, controlled fallback in the exception.

What failover does not do

It catches new requests. A response already streaming can break off during an outage and has to be requested again. Redundant model clusters alone also do not add up to high availability: gateway, network, identity, storage and monitoring have to be sized and designed accordingly.

One entry point, behind which the models may change

Integration

Chat interfaces, agents, development environments, scripts and line-of-business applications address one central, OpenAI-compatible endpoint. Behind it you can run several local model servers, additional clusters, specialist models and approved external providers.

Your application's integration stays unchanged, even when

  • a local model is replaced
  • a new model cluster is added
  • an external provider is approved or withdrawn
  • a model is temporarily unavailable
  • departments use different models
  • a routing policy changes

How much adjustment a connection needs in a given case depends on which functions the client uses, for instance streaming or tool calls.

OdyControl runs in your environment: fully local, hybrid with approved external providers, or isolated with no outbound connection. Which topology suits your organisation is what the architecture call is for.

Document without recording everything

Traceability

Every model call that runs through OdyControl can be documented centrally, local ones as well as forwarded external ones. At three depths.

  1. 1

    Operational data

    Timestamp, calling application, user or group reference, data class, routing policy, chosen model, local or external target, provider, status, latency, errors, failover and recovery events. No content.

  2. 2

    Selective or redacted

    Specific fields are deliberately captured, omitted or pseudonymised.

  3. 3

    Complete

    The full request and response, for explicitly approved use cases. Including external calls routed through OdyControl.

The third level is not a default setting

A complete log can itself contain what the data class is meant to protect, which makes it a particularly sensitive store. Purpose, legal basis, data scope, access, retention, deletion, pseudonymisation, affected groups and co-determination requirements must be settled before it is switched on.

What OdyControl builds on

Prerequisites

Four points that are settled before the pilot.

It works with your data classification

The routing decision needs a data class or another unambiguous attribute. Optional content routing only assigns the types your organisation has defined; whatever it cannot assign with certainty stays internal.

It controls the model path

For web search, file transfer, retrieval outside the controlled path, external tools, code execution, speech services and identity services you set controls of their own.

It enforces your legal review

Whether a connected provider is permissible for a processing operation is your organisation's call: legal basis, contract, provider assessment and data minimisation. OdyControl enforces the outcome of data protection review, DLP and AI governance technically.

It is the only route to the models

Model requests run exclusively through OdyControl; your network blocks direct access to external providers. That makes OdyControl the control point for every request.

Frequently asked questions

What is a sovereign AI platform?

It is an AI platform that an organisation runs itself and whose rules it sets itself. The language models run in its own network, the organisation decides which requests may leave it, and usage is documented in a traceable way. OdyControl combines a browser-based AI workplace, local models and a control layer for this.

How can a company steer its use of AI centrally?

By routing every request to a language model through one control point, instead of each department connecting its own tool. In OdyControl this is the control layer: for each data class it determines which models a request may reach, enforces this on every request. Calls are logged. This works when direct access to model providers that bypasses the control point is blocked.

What happens if OdyControl cannot log a request?

That depends on the operating mode. In the “audit-proof” operating profile: if the entry in the tamper-evident log chain cannot be written durably, the request is refused. There, no entry means no processing. Alternatively, OdyControl can buffer entries and write them later. A mode that still answers despite a write error also exists, but it is not intended for regulated organisations.

Can OdyControl be used for VS-NfD (German restricted classified information)?

The BSI approves individual IT security products, not an AI system as a whole. A company releases its VS-NfD IT internally, through the person responsible for VS-NfD. OdyControl can be operated entirely inside your own network with no path to the outside. Whether such an environment is released for VS-NfD is decided by the person responsible for VS-NfD in your organisation.

Can OdyControl also use public language models?

Yes, if your organisation approves it. For each data class you decide whether requests stay local or may go to reviewed providers. Requests in the Local Only class stay in-house even during an outage; non-critical tasks can use whichever suitable model is available.

Does OdyControl conform to the Cyber Resilience Act?

Full application of the CRA begins on 11 December 2027. What we can evidence today is how it is built: the route to external models is closed without approval, every version is delivered signed and with a software bill of materials, access is protected and the logs are tamper-evident.

What happens if a provider changes its prices or models?

For local models, nothing changes as a result. They run on your hardware without billing per request, and a model is only replaced once your organisation has reviewed and approved the new one. For approved public models, the provider's prices and model versions apply; the control layer determines which tasks your organisation uses them for.

Does Cybervize use OdyControl itself?

Yes. OdyControl runs in Cybervize's own operations; we are our own first customer. What we propose to you, we use ourselves every day.

Does OdyControl keep all data local?

Fully local operation is possible. In mixed operation the configured data class decides whether a request must stay local or may use an external provider. Your organisation determines which AI workloads may leave it.

Can Local Only data be processed externally during an outage?

No. A Local Only policy remains binding during an outage. With no local backend available the request is refused in a controlled way or can be submitted again later.

Does OdyControl detect sensitive data automatically?

Optionally, within your rules. If your organisation switches on content routing, a classifier proposes a type from your own taxonomy based on the content. Uncertain cases and unreadable attachments stay in-house, and a content rule never opens a path to the outside that your policy has closed. Without content routing, OdyControl enforces the data class set by the application, workflow or policy.

Does OdyControl make us data-protection compliant?

No. OdyControl provides technical control mechanisms. Legal basis, contract design, provider assessment, data minimisation and the permissibility of the specific use case are assessed by your organisation.

Do existing applications have to be changed?

OpenAI-compatible applications can use the central endpoint. The effort depends on which API functions the client uses, for instance streaming or tool calls.

OdyControl and OdySecure

OdySecure, the security platform by Cybervize, brings ISMS, compliance and risk onto one data foundation. OdyControl is the AI platform alongside it: the workplace, the models and control over which request may reach which language model. Two products, the same stance: control stays with the organisation.

See OdySecure

No CISO of your own?

In a vCISO mandate, your vCISO works with OdySecure and the platform is included. From €3,600/month.

Talk to us about your AI platform

In an architecture call we work out what your employees should use AI for and which workloads have to stay local at your organisation. We also look at which providers are candidates for which cases and which topology fits.

Price on request. The scope depends on which models, applications and data classes are connected.