AI compliance in WORKSPACE.PM

This page summarises how WORKSPACE.PM uses artificial intelligence. It is written for everyone who reviews or approves its use: data protection officers, IT security, works councils, internal audit and auditors. The full version with all details, templates and appendices is available as a PDF above.

We also state what WORKSPACE.PM cannot do today. For your assessment that matters just as much as the controls that exist.

WORKSPACE.PM ships without a preset AI provider and without a bundled access key. Without explicit setup by your administration, no AI processing takes place and WORKSPACE.PM passes no data to a model.

The MCP interface, through which external AI applications connect, has its own switch. It does not call a model itself and does not require an AI configuration. In new tenants it is switched off; tenants that existed before this default was introduced kept it switched on. There, it is worth checking the switch.

WORKSPACE.PM is fully usable without AI. There is no step that cannot be done without it.

Except for the MCP interface, all features require a configured AI setup and are controlled through the generativeAi permission. They run within the signed-in person's permissions and when that person triggers them; the only exception is the project summary. There is no switch per feature; the features in Microsoft Teams are off by default.

Not an AI feature: search, reports, analytics, dashboards, workflows and notifications work rule-based on the stored data, without a language model. Document generation does not call a model either.

With your own provider, you choose the AI provider yourself and conclude your own agreement with them. Our data processing agreement does not cover this transfer; the AI provider is not a subprocessor of AMNAU GmbH. If the provider is outside the EU, this is a third-country transfer that you need to safeguard. For Anthropic, OpenAI and Mistral you set the processing region; DeepSeek's interface is operated in China.

With a locally hosted model, questions about data processing agreements, third countries and training at the provider do not arise. For the assistant's tools the model needs reliable function calling; otherwise the assistant works as a plain chat without tools.

WORKSPACE.PM itself trains no models and does not pass customer data on for this purpose. Whether a chosen provider uses transmitted content for its own training depends on its terms and your plan.

Who is responsible for what under data protection law

The AI always works on behalf of the signed-in person and sees at most what that person may see in the interface. Every tool call, in the assistant as well as via MCP, runs through the same tenant separation and permission check as a click in the interface. Tenant separation is enforced at database level.

For your assessment this means: what counts is the scope of permissions of the people allowed to use AI. Broad permissions give the AI broad access. So do not connect accounts with administration rights to the assistant or to MCP.

What is sent to the model

• The person's input, and in the chat also the conversation so far
• In the chat the person's first name, name and role as well as the response language; not the email address, personnel number or manager
• Results of tool calls, limited by the person's permissions
• When creating a project from a PDF only its text, at most 120,000 characters; the file itself never leaves the browser
• When analysing meeting minutes their text, which regularly contains employee names

What is not sent to the model

• Uploaded files themselves, only the text read in the browser
• Credentials, passwords and session keys
• Data of other tenants and data outside the person's permissions
• Consumption data and the audit trail as such; the project summary only receives the status and phase changes of the last 14 days

Through the open standard MCP (Model Context Protocol), your employees connect external AI applications such as Claude, ChatGPT or Microsoft Copilot to WORKSPACE.PM. Each person signs in individually instead of through a shared account; their permissions apply, plus a separate permission level for MCP. Your administration sees all connections and can revoke individual people's access at any time.

Two settings to make deliberately

1. 1Set MCP to read access. With the permission level integrationMcp: view every write attempt is rejected, even for people who may change everything in the interface. In new tenants the PMO and Team roles can only read and the Admin role can write; in older tenants all built-in roles have write access. Write actions via MCP are not confirmed by WORKSPACE.PM but, depending on its settings, by the external application.
2. 2Do not connect accounts with administration rights. For such accounts the upper limit would be "everything", which effectively cancels out the protection of the permission model.

What the external application does with retrieved data is outside WORKSPACE.PM; your agreement with its vendor applies. Which applications are allowed is something you regulate organisationally; there is currently no technical allowlist.

More on connecting via the MCP server

1. 1Provider or local model? For high protection needs: local. For Anthropic, OpenAI and Mistral you also set the processing region.
2. 2Who may use AI? A dedicated role with narrow permissions instead of broad existing roles.
3. 3Read only or write as well? Our recommendation: start with read access.
4. 4Which external AI applications are allowed? You regulate this organisationally. Until the target configuration is in place, MCP is best left switched off.
5. 5How long are chat histories and consumption data kept? For consumption data you set the period; for chat histories you currently define it organisationally.
6. 6Explicitly rule out employee-related analyses.
7. 7Involve the works council before approval. We assume co-determination applies.

Recommended baseline for highly regulated environments

The review and approval checklist in the documentation lists these points for ticking off; the policy template contains matching wording.

What an auditor can verify directly in WORKSPACE.PM, without having to rely on assurances:

The documentation provides pre-written building blocks for your data protection impact assessment: description of the processing, necessity and proportionality, and risks with measures. Only you can assess your specific purpose of use.

Indicators for proportionality: the features are optional and can be switched off, the scope of data is limited by the person's permissions, and operation without transfer to third parties is possible with a local model. Less intrusive alternatives are restricting to read access and a narrow group of users.

Once the recommended baseline and the organisational rules are in place, we consider the residual risk manageable. Three points require an explicit decision by you: retention of conversations, exclusion of employee-related analyses and the handling of external AI applications via MCP.

What you add

• Purpose of use and affected areas
• Group of people with AI permission
• Chosen AI provider and legal basis for the transfer
• Retention rule
• Interaction with other systems
• Outcome of the works council involvement

The following points describe limits of the current state. Take them into account in your assessment and in your organisational measures. We track them as open points in the documentation and record every change.

In our assessment, without legal advice: with its AI features WORKSPACE.PM is an AI system, AMNAU GmbH is its provider and your organisation is usually the deployer. The models come from third parties or are run by you.

• Not a high-risk system: WORKSPACE.PM is not intended for decisions about employees or for assessing performance and behaviour, and contains no such function.
• No prohibited practices: no emotion recognition, no biometric categorisation, no social scoring.
• Art. 50(1): the assistant is labelled as AI, and running AI operations are shown as such.
• Art. 50(2): generated content is not marked in a machine-readable way; we rely on the exception in the Commission guidelines for internal business applications. If you pass AI content on externally, we recommend labelling it.
• Art. 4 AI literacy: training your users is your responsibility as the deployer; the policy template covers the key points.
• No automated decisions under Art. 22 GDPR: suggestions are adopted by a human.

Classification per feature

Because the available tools are objectively suitable for analysing behaviour or performance, we assume co-determination applies under Section 87(1) no. 6 of the German Works Constitution Act (BetrVG). A template for informing the works council under Section 80(2) BetrVG is included, and a works agreement template for the agreement itself.

For customers with a management system under ISO/IEC 42001 or ISO/IEC 27001, the documentation maps the AI features to the clauses of the standards. It shows which product property contributes to a control; a control is never fulfilled by the product alone, but in your management system. No certification of the product or of AMNAU GmbH is claimed.

ISO/IEC 42001 (AI management system)

ISO/IEC 27001, Annex A

The full mapping, including the articles of the GDPR, the EU AI Act and the German Works Constitution Act, is in Appendix H of the documentation.

The documentation is built as a working basis for your review: for audits, data protection impact assessments, security questionnaires and works council involvement. The templates are also available as editable Word files.

The answers from Appendix G of the documentation, for use in security questionnaires and supplier assessments. In the PDF, every answer references the relevant section.

Version

1.0

As of

2026-10-05

Product version

WORKSPACE.PM 2.0.10

Publisher

AMNAU GmbH, Kronenstraße 13, 30161 Hannover, Germany

Questions and error reports

info@amnau.de

We review the documentation with every release; the tool catalogue and system prompts are regenerated from the source code. At least once a year it is reviewed in full. Changes that require a renewed review on your side, such as new AI features or changed defaults, are marked as review-relevant in the change log.

In the SaaS edition a new AI feature is available to everyone with the generativeAi permission immediately after an update; in the self-hosted edition you decide when to update.

Further reading

• Security & compliance
• AI assistant Ace
• Connect AI assistants via MCP