TGV guide: the Artificial intelligence group

What the group covers

The Artificial intelligence group of the TGV grid (version of April 24, 2026) checks that a technological product or service (PST) that includes AI is documented, tested and governed, and that it keeps the user in control. The grid calls such a product a SARIA, from the French “solution ayant recours à l'intelligence artificielle”, an AI-enabled solution: as soon as an AI component is present, IA1 requires it to be clearly stated in the documentation. The group holds 69 of the grid's 382 criteria (IA1 to IA69), in 9 categories. Several apply only in specific cases, such as generative AI or autonomous agents. The criteria name no type of product: vendors of medical scribes, transcription tools or generative AI are concerned as soon as their solution includes an AI component.

The official grid exists only in French: the criteria are translated and paraphrased here, and the advice and suggested evidence are Sentrix's interpretation.

The categories one by one

The reference column mostly cites RS BC-VPTI (59 criteria) and OREPIA (46), then ISO 25058, LEIA, ISO 23894, ISO 27563, ISO 24368 and the TGV criteria of June 18, 2024.

Use of AI systems (IA1 to IA7)

The vendor publishes on its website a privacy policy in plain and clear language on the use of AI (IA2). IA3 to IA7 govern decisions based exclusively on automated processing of personal information, which the instruction for IA3 asks the vendor to list, citing section 12.1 of the Act respecting the protection of personal information in the private sector and section 65.2 of the Access Act (Loi sur l'accès). The SARIA informs the person concerned (IA4), provides on request the personal information used, the reasons and factors behind the decision and the right to have the information corrected (IA5), lets the person submit observations to a staff member (IA6) and allows human review of the decision (IA7).

Development process (IA8 to IA12)

The vendor presents its development process so the acquirer can assess potential, performance, risks and training and governance needs. Testing must identify differences in behaviour across the conditions or social groups of the context of use (IA8) and establish fitness for the intended use (IA9), along with an evaluation strategy with its metrics, false positives and false negatives (IA10). For generative AI, add a report on the variability of results for the same situation (IA11) and a report on malicious prompt injection tests (IA12).

Documentation (IA13 to IA32)

This is the largest category, with 20 criteria. The documentation describes the intended use, limits and warnings, bias measures and non-recommended uses (IA13 to IA16), the AI types (IA18), the architecture (IA24) and, where relevant, the control of autonomous agents (IA30). It lists the training, validation and test datasets, ideally distinct, and documents anonymization and synthetic data (IA17, IA27, IA31, IA32). The vendor changes neither the AI types nor the models without the acquirer's explicit approval (IA20) and presents an evolution plan as rigorous as the initial verification (IA21 to IA23). It also provides a report on previous deployments (IA19), user training (IA25), a plain-language document on risk mitigation (IA28) and the SARIA's environmental footprint (IA26, IA29).

Responsible use (IA33 to IA38)

The vendor agrees to work with responsible AI researchers and a committee of potential users and service users (IA33). It keeps a risk register or contributes to one kept by an independent body (IA34), and reports any new risk to the acquirer, requiring the same of the primary developer in the case of secondary development (IA35). It does not use datasets whose integrity may have been compromised in the 5 years before their use (IA36), monitors such events (IA37) and holds the rights to its training data, especially personal information (IA38).

Communication and sharing (IA39 to IA46)

Personal information exchanged between components is protected according to Quebec law (IA39). IA40 forbids any communication of personal information outside Quebec, including through a call to a generative AI model, without the person's informed, explicit and unequivocal consent. The acquirer and the Bureau de certification are informed of any communication or sale of data, and the instruction recalls that selling health or social services data is strictly prohibited unless explicitly authorized by law (IA41). The acquirer is also informed of derived products, interested third parties and analyses it did not request (IA42 to IA44), can disable one or all models (IA45) and must explicitly authorize any machine learning on usage data (IA46).

Data access (IA47 to IA51)

Users' access, decisions, reactions and comments are logged and traceable (IA47), and remain available for as long as the decisions the SARIA supports can be challenged in court (IA48). Each user sees only the data and performs only the tasks their role allows (IA49, IA50). Usage data and results are logged and accessible to the acquirer (IA51).

Monitoring and control (IA52 to IA58)

The acquirer can track and extract users' approvals, changes, comments and reactions (IA52) and grant access according to roles (IA55). Where possible, the final recommendation is built from intermediate results the user can view and extract (IA53, IA54). A model that adapts through machine learning does so according to the acquirer's approval (IA56). With generative AI, the sources appear in the results and can be reached from the interface (IA57). Users can flag an inadequate result, and the flag is logged (IA58).

Review process (IA59 to IA67)

The instruction calls for a SARIA that supports user autonomy. Screens show limits and warnings, what the grid calls an “explainability strategy” (IA59), and explain what the model does and which data it considers (IA62). The user knows they can always decline a recommendation (IA60), and the SARIA changes no human input after the user's final validation (IA61). The version is visible on screen and in each result (IA63, IA64), autonomous agents are monitored (IA65) and external calls have automatic checkpoints and a continuation strategy (IA66). Internal promotion is limited to the first 2 minutes of use, and commercial advertising is forbidden (IA67).

Artificial intelligence (IA68 and IA69)

No word, visual or sound element, movement or response time of the user triggers sponsored associations or actions (IA68), and no subliminal technique is used (IA69). These criteria extend G02 in the General group, which forbids advertising.

Evidence to prepare

• A SARIA overview (AI component, types, architecture) and the published privacy policy (IA1, IA2, IA18, IA24).
• A register of automated decisions, with the notice shown to the person and the human review procedure (IA3 to IA7).
• The test file: bias, fitness for use, metrics, false positives and false negatives, plus the variability and prompt injection reports for generative AI (IA8 to IA12).
• A dataset register: list, rights acquired, separation of stages, anonymization, synthetic data (IA17, IA27, IA31, IA32, IA38).
• The evolution plan, user guide, plain-language risk document and environmental note (IA21 to IA23, IA25, IA26, IA28, IA29).
• Written commitments, ideally in the contract: no model change without approval, risk register, data integrity, no learning on usage data (IA20, IA34 to IA37, IA46).
• A map of personal information flows, with each call to an external model and where it is processed (IA39, IA40).
• A demonstration of the features: disabling models, logs, flags, displayed sources, version, ability to decline (IA45, IA47, IA57, IA58, IA60, IA63).

Points to watch

• Conditional criteria, among others: IA11, IA12 and IA57 (generative AI), IA30 and IA65 (autonomous agents), IA31 (synthetic data), IA56 (adaptive model), IA19 (prior deployment).
• IA48 gives no fixed duration: log retention depends on how long decisions can be challenged in court.
• Certification attests the conformity of a version of a PST: any model change goes through the acquirer (IA20), and the version appears on screen and in each result (IA63, IA64).

Need help with this group?

Sentrix helps PST vendors prepare their file for this group, from the dataset register to the written commitments; certification itself is the role of Santé Québec's Bureau de certification. Book a meeting or see our TGV certification support service.

Next in the guide

See the complete TGV controls guide and the TGV framework page.