Governance in practice

AI governance also starts inside the organization

Watermarks, internal rules and human oversight: a practical approach with precise scope for the Loi 25 and CAI references.

Illustration of interlocking gears connecting governance, checks, analysis and the team.

Quick answer

Part of AI governance is built inside the organization: in the rules a team applies, the checks it keeps and the decisions someone can still review. Automated controls can support that work. Their presence alone does not demonstrate that an application is reliable or compliant.

My view: we can start with the AI uses already in place. We need to know what each control checks, what it misses and who intervenes when it fails.

A watermark provides a provenance signal

Google DeepMind’s SynthID embeds watermarks in content generated with supported tools. Detection can help identify that provenance. It does not establish factual accuracy or legal compliance.

DeepMind’s documentation also identifies an important limitation for text: extensive rewriting or translation can reduce detection confidence. We therefore need to understand the conditions of the check before interpreting its result. A missing detected watermark, on its own, does not establish that a human produced the content.

The organization owns the decision rule

A provider can add filters, alerts or oversight features. It may not know the commitments you have made to clients, the data your team may use or the level of risk acceptable in your process.

An internal rule needs to be specific enough to apply. Which documents may an assistant access? Which outputs require review before being sent? When must the work stop for a person to decide? A general instruction to “use AI responsibly” leaves those decisions open.

The human auditor also examines the control

The human role extends to designing the rule and checking the mechanism that applies it. An alert that works today may miss a new case. People need to be able to examine errors, adjust the criteria and remove an automation that no longer serves its purpose.

ISO/IEC 42001 provides an AI management system framework covering responsibilities, risk management, monitoring and improvement. Certification is voluntary and is performed by independent certification bodies; ISO does not certify organizations. That framework does not turn a watermark or filter into a certificate of compliance.

In Québec, separate the law from a recommendation

Section 12.1 of the Act respecting the protection of personal information in the private sector addresses an enterprise using personal information to make a decision based exclusively on its automated processing. It provides for notification, information on request and an opportunity to submit observations to an employee able to review the decision. It does not apply indiscriminately to every use of an assistant.

In its June 2026 report, the CAI recommends adding an algorithmic impact analysis to a privacy impact assessment involving an AI system intended to make automated decisions about people. This is recommendation 45, a proposed change to the legislative framework. The passage does not establish a new blanket requirement for every AI tool used in human resources.

A control record for one real use

I suggest starting with one recurring use and documenting four elements. This is a working aid, rather than a complete compliance assessment.

ElementWhat to specify
Use and ownerThe tool, task, permitted data and person accountable for the result.
Testable ruleWhat is accepted, what is rejected and what requires review before use.
Human interventionThe condition that stops work, the person called and what they can correct.
Useful recordThe test date, input, checked result and correction made.

Then test an ordinary case and one that should stop the process. If nobody can explain why the control accepted the second case, there is more work to do before entrusting it with further decisions.

Three questions to take into the week

  1. Can you name the AI tools the team uses, the tasks they support and the owner of each use?
  2. Can you show how an important output is checked, including the limits of the control?
  3. If a use makes a fully automated decision from personal information, have you had its framework examined and provided for someone who can review that decision?

The answer can begin with one small, well-documented process. You can continue with lessons from AF9000+ auditing and everyday AI use and Loi 25, then discuss how your organization uses AI internally.

Sources and scope of this update

Sources checked on October 28, 2026. The position on internal controls and the proposed control record are the author’s analysis. The references below document the technology, management framework and cited provisions.

← Back to the articles