Trust and compliance
AI security and the AI Act at Gfacility
We publish this information at this level of detail on purpose. Organisations that use Gfacility should be able to make their own assessment of AI risk and compliance, instead of having to rely on an opaque AI service. This page therefore sets out how we classify the AI functionality in our platform under the AI Act, how the technology is set up, and where each responsibility sits.
This is a classification assessment, not a conformity assessment. A conformity assessment is only required for high-risk AI systems, and our assessment concludes that the AI functionality of Gfacility is not one.
Version 1.0, August 2026
What was assessed
The intended purpose of the AI functionality is to support facility and IT service processes: logging, classifying, routing and resolving tickets, searching the knowledge base, making bookings and managing configuration items.
The functionality comes in two forms. The AI assistant is a conversational layer inside the interface that works strictly with the permissions of the signed-in user. Autonomous AI agents run pre-configured use cases, started by a deterministically configured trigger with no AI component: no AI model decides whether an agent runs.
Outcome of the assessment
| Part of the AI Act | Outcome |
|---|---|
| Prohibited practices (art. 5) | Not applicable |
| High risk (art. 6 and annex III) | Not applicable |
| Transparency obligations (art. 50) | Applicable |
| Model provider obligations (art. 53 and following) | Sits with the model provider |
| AI literacy (art. 4) | Applicable to Gfacility and to the customer |
Classification: a limited-risk AI system, to which the transparency obligations of article 50 of the AI Act apply.
Division of roles
| Role | Party |
|---|---|
| Provider of the AI system (art. 3(3)) | Gfacility |
| Deployer (art. 3(4)) | The customer |
| Provider of the general-purpose AI model | Google (Gemini) |
| Hosting party | Google Cloud |
Customers configure their own organisation instruction, skills, use cases and triggers inside the platform. That does not change the division of roles: Gfacility remains the provider and the customer remains the deployer, even when the customer substantially rewrites the wording or the working method in an instruction.
One exception: article 25(1)(c) treats a party as a provider when it changes the intended purpose in such a way that the system qualifies as high risk. That presupposes use for a materially different purpose falling under annex III, for example assessing the performance of individual employees. As long as use stays within tickets, bookings and knowledge-base questions, this provision does not apply.
Why this is not high risk
None of the eight annex III categories applies. No biometric data is processed, no creditworthiness or entitlement to benefits is assessed, and the functionality has no controlling or safety role in technical installations: it is administrative in nature.
Category 4 (employment and work) deserves separate comment, because a ticketing system assigns work to employees. That category covers the recruitment or selection of people, decisions on promotion or termination, the allocation of tasks based on individual behaviour or personal traits, and the monitoring and evaluation of performance and behaviour. The AI functionality of Gfacility does not fall under it:
- Assignment happens on content, not on person. Routing is based on the ticket type, the location and the configuration item involved, following the routing rules the customer has set up. Individual behaviour or personal traits of employees are not an input.
- No performance or behaviour evaluation. The system produces no assessments, rankings or productivity scores about employees.
- A human stays in oversight. Every AI action is visible and can be overridden by an authorised user.
Even if a function were read into an annex III category, the exception in article 6(3) applies: the functionality is purely preparatory and improving, and it replaces no human assessment.
Applications we exclude
Gfacility does not supply the AI functionality for:
- assessing, ranking or monitoring the performance or behaviour of individual employees;
- decisions about access to facilities, benefits or essential services;
- assessment of creditworthiness or insurance risk;
- biometric identification or categorisation;
- emotion recognition;
- automated decision-making with legal consequences for individuals without human involvement.
These exclusions are laid down in the master instruction: the instruction layer that only Gfacility manages and that customers cannot change. It applies to both the assistant and the agents, and it takes precedence over every layer the customer configures.
Where the AI runs
| Property | Value |
|---|---|
| AI model | Google Gemini |
| Model provider | |
| Hosting party | Google Cloud (Vertex AI) |
| Processing location | Saint-Ghislain, Belgium (europe-west1) |
We deliberately use a European endpoint and not the global Vertex AI endpoint, because the latter gives no guarantee about the processing location. All prompt content and model processing takes place inside the European Union, under the Vertex AI Data Processing Addendum and with EU standard contractual clauses. The underlying platform is certified to ISO 27001, ISO 27018 and SOC 2.
The model is a managed enterprise service. We do not host it ourselves, we do not modify it, and we do not train or fine-tune it on customer data. Prompt content and model output are not used to improve the model. The inference service is stateless: prompts are processed in real time and then deleted, no customer context is left behind between requests, and no model state is shared between customers.
How an AI request runs
Every AI interaction follows the same controlled path:
- 1
A user action or a system trigger starts an AI request inside Gfacility.
- 2
Gfacility assembles the context following the instruction hierarchy, supplemented with data retrieved through the internal API within the permissions of the user or the agent.
- 3
The fully assembled prompt travels over an encrypted, authenticated connection to the European endpoint.
- 4
The response comes back to Gfacility and is validated.
- 5
If the response contains proposed actions, those are carried out through the internal APIs of Gfacility, with the same permission, validation and audit checks as a manual action.
- 6
The complete interaction is recorded in the audit trail.
We manage this entire path ourselves: assembling the context, validating the response and carrying out actions all happen inside Gfacility. As a result, every action initiated by AI passes the same checks as a manual one.
Safeguards
Whatever the classification, we apply the safeguards below. They are drawn from the high-risk regime and are applied voluntarily as good practice.
-
No direct database access
AI never talks to the database directly, only through the internal APIs of Gfacility, which enforce permission and scope checks. If something is not accessible inside the platform, it is not accessible to AI either.
-
Permission model
The assistant has no permissions of its own and works inside the session and the permissions of the signed-in user. An agent works with its own account and the permission groups the customer administrator has granted it. In both cases authorisation is enforced in the application layer: no instruction can widen permissions or data scope.
-
Prompt injection
We design on the assumption that a language model can be manipulated. The safeguards therefore aim at limiting the consequences: text from tickets, attachments, email and user input is treated as content and not as an instruction, and the system can only call a predefined set of functions. Prompt injection cannot be ruled out entirely at model level; the residual risk stays limited to unwanted or incorrect textual output within the permissions granted. Privilege escalation and access to another organisation's data are architecturally impossible.
-
Traceability
All actions by assistants and agents are recorded in the audit trail and can be distinguished from actions by a natural person, including the user or permission group involved.
-
Monitoring
We monitor the quality and consistency of AI output over time, and we flag failed requests, unexpected output and anomalous use. No retraining or live learning takes place in production.
-
Control at the customer
AI functionality can be switched on and off per user role. New AI features are delivered switched off; the customer decides on activation. Proposed actions can be reviewed or overridden, and any agent or use case can be switched off at any moment.
-
No model training on customer data
Customer data is not used to train or fine-tune models. Only the data needed for the task at hand is offered to the model.
-
Quality monitoring
Users can indicate per response whether it was useful. Those ratings are recorded and analysed to sharpen the instructions and structurally improve the reliability of the output.
-
Security
Data is encrypted in transit and at rest. Gfacility is ISO 27001 certified and has a penetration test carried out every year.
What this means for you as the deployer
Two obligations sit with your organisation and not with Gfacility.
-
AI literacy (article 4)
Employees who work with the AI functionality must have sufficient understanding of how it works, what it can do and where its limits are. We make documentation, an administrator instruction and end-user guidance available for this.
-
Responsibility for your own configuration
You decide on the organisation instruction, skills, use cases and triggers inside your environment, and you are responsible for their content and lawfulness, including the GDPR basis for the processing your use cases set in motion. This division is laid down in the AI annex to the agreement.
We meet the transparency obligation of article 50 inside the product: the assistant is explicitly labelled as AI in the interface, and AI-generated proposals are marked and traceable in the audit trail.
Full documentation
This page is a summary. On request we make the full documentation available to customers and prospects:
-
AI Act classification assessment
The full test against article 5, annex III, article 50, article 53 and following, and article 4, including the reasoning per category and the complete overview of control measures.
-
AI security in Gfacility
The technical and organisational security measures around the AI functionality, including architecture, instruction hierarchy and monitoring.
Questions about this assessment or about our AI documentation?
Get in touchFrequently asked questions
Is the AI in Gfacility a high-risk AI system?
No. None of the eight annex III categories applies, so the functionality is a limited-risk AI system. The transparency obligations of article 50 of the AI Act do apply to it.
Who is the provider and who is the deployer under the AI Act?
Gfacility is the provider of the AI system (art. 3(3)) and the customer is the deployer (art. 3(4)). Google is the provider of the general-purpose AI model and Google Cloud is the hosting party. Configuring your own instructions, skills, use cases and triggers does not change that division.
Which AI model does Gfacility use, and where does it run?
We use Google Gemini as a managed enterprise service on Google Cloud (Vertex AI), on a European endpoint. All prompt content and model processing happens in Saint-Ghislain, Belgium (europe-west1), inside the European Union.
Is our data used to train AI models?
No. Customer data is not used to train or fine-tune models, and prompt content and model output are not used to improve the model. The inference service is stateless: prompts are processed in real time and then deleted.
Can we switch the AI functionality off?
Yes. AI functionality can be switched on and off per user role, new AI features are delivered switched off, and any agent or use case can be switched off at any moment.
What does our own organisation have to arrange?
Two things: AI literacy under article 4 for the employees who work with the functionality, and responsibility for your own configuration, including the GDPR basis for the processing your use cases set in motion. This division is laid down in the AI annex to the agreement.