On September 10, 2026, Anthropic added a permissions policy to its Claude agents that evaluates every action the agent takes and pauses it for human approval when it is sensitive. This is not a technical footnote: it is exactly the mechanism Law 25 asks you to have in place when an automated decision touches your clients' personal information.
The Short Answer
A modern AI agent no longer just drafts text: it acts (it creates an account, sends an email, modifies a file). The right question for an executive is therefore not "is the AI good enough?", but "which actions happen on their own, which ones require a human hand, and where is that written down?" As of September 10, 2026, the Claude platform knows how to draw that line at the server level: every tool call is evaluated, then either executed, refused, or held for your approval. And since September 22, 2023, Law 25 has already obligated you to be able to answer that question in Quebec.
What Changed on September 10, 2026?
Two additions, from the official Claude platform release notes:
- An
autopermissions policy for managed agents (Claude Managed Agents): the server evaluates every tool call made by the agent (including tools connected via MCP), then either executes it, refuses it, or suspends it pending approval. Each evaluation is logged inagent.tool_useandagent.mcp_tool_useevents. - An
ant beta:sessions connectcommand that lets you connect live to an active agent session: follow what it is doing, communicate with it, and authorize or refuse pending actions.
(Source: Claude platform release notes, entry dated September 10, 2026: platform.claude.com/docs/en/release-notes/api.)
The distinction that matters: the safeguard is no longer an instruction written in a prompt that the AI is kindly asked to follow. It lives in the platform, outside the model's reach. A rule that exists only in a prompt is not a rule; it is a wish.
Why Does This Concern an Executive, Not Just the IT Team?
Because Quebec law has already settled the question.
Section 12.1 of the Act Respecting the Protection of Personal Information in the Private Sector (CQLR, c. P-39.1), in force since September 22, 2023, applies to any decision based exclusively on automated processing of personal information. Such a decision is defined as one made without any natural person having exercised meaningful control.
When that is the case, the organization must:
- inform the individual of the automated nature of the decision, no later than at the time it is communicated to them;
- upon request, indicate the information used, the reasons and the main factors that led to the decision, and their right to have it corrected;
- give them an opportunity to present their views to a staff member who is able to review the decision.
(Source: Gouvernement du Québec, "Décision fondée exclusivement sur un traitement automatisé," quebec.ca; Commission d'accès à l'information du Québec, "Principaux changements apportés par la Loi 25.")
Two very concrete consequences:
✦ A real human approval point takes you out of the "exclusively automated" category. It has to be real, though: named, traced, exercised, and not a checkbox in a policy nobody reads.
✦ Even outside that category, you must be able to explain. "The AI decided" is not an answer. A per-action evaluation log, like the one Anthropic just made native, is precisely what makes a decision explainable twelve months later.
How Do You Decide What an Agent Does on Its Own?
In our practice, the rule is not set by agent or by trust level. It is set by cost of error:
- Near-zero cost and reversible → autonomous. Reading a mailbox, sending an acknowledgement, serving a procedure that has already been validated. Requiring approval here protects no one and wears everyone out.
- Reputational cost → human review. Anything that commits the organization on substance: a solution, a deadline, an amount, published content.
- Irreversible cost → explicit approval. Creating or revoking an identity, resetting a password, wiping a device, sending something externally.
The classic trap is tightening control uniformly "to be safe." This produces two effects, both bad: zero-cost actions get blocked, so the service goes silent; and approval fatigue sets in. Past a certain volume, the tenth request gets skimmed and the twentieth approved by reflex. A safeguard nobody reads guards nothing.
Where Do You Start?
This is the logic behind our CAPE method:
- Conceive: inventory the actions the AI would take, and classify them by cost of error. That inventory is also your Law 25 compliance document.
- Activate: start with a narrow scope, with the approval point connected from day one, not added later.
- Propel: expand autonomy where measurement justifies it, action by action.
- Evolve: review the log: what was approved and never refused can become autonomous; what was refused must stay guarded.
FAQ
Does Law 25 prohibit using an AI agent in my organization? No. It prohibits nothing. It requires transparency, an explanation available upon request, and a reachable human capable of reviewing the decision.
Does this apply to SMEs, or only to large organizations? Section 12.1 applies to businesses regardless of size. An SME that makes an automated decision about a client is subject to it just like a large corporation.
If an agent sorts my emails, is that an "automated decision"? Classifying an email is generally not a decision within the meaning of section 12.1. Rejecting a file, setting a condition, or cutting off a client's access is. The line is documented case by case: that is the work of the Conceive step.
Does every AI action require a human behind it? No, and that would be counterproductive. A human is needed at the points where an error is costly or cannot be undone. Everything else should run on its own.
How do you prove that a human exercised meaningful control? Through the trail: who approved, what, when, and on what basis. That is precisely what a per-action evaluation log provides.
If you are considering letting an AI take actions in your systems, the "who approves what" inventory takes one meeting. Write to us: we start from your actual actions, not a template.
Article written on September 12, 2026.
LinkedIn Version
What actions is your AI allowed to take without asking your permission?
On September 10, Anthropic made that control native: Claude agents now evaluate every action (execute, refuse, or pause for human approval). Each evaluation is logged.
It sounds technical. It is an executive topic.
Because in Quebec, section 12.1 of Law 25 has been in force since September 22, 2023. It targets decisions made without any person having exercised meaningful control: it then requires you to inform the individual, explain the factors upon request, and offer them a human capable of reviewing the decision.
What we see in practice: the right question is not "how far can the AI go?", but "what does an error cost?"
✦ Zero cost and reversible → autonomous. ✦ Reputational cost → review. ✦ Irreversible cost → approval.
Tightening everything "out of caution" causes two problems: the service goes silent, and approval fatigue sets in. A safeguard nobody reads guards nothing.
No prohibition. No smoke and mirrors. A written boundary.
Law25 QuebecSME AIQuebec SMELeaders AIGovernance
Sources Cited
- Claude platform release notes, entry dated September 10, 2026: https://platform.claude.com/docs/en/release-notes/api (verified 2026-09-12; the former docs.anthropic.com address redirects 301 to this one)
- Gouvernement du Québec, "Décision fondée exclusivement sur un traitement automatisé": https://www.quebec.ca/gouvernement/travailler-gouvernement/normes-gouvernance-pratiques-internes/protection-des-renseignements-personnels/technologie-et-droit-a-la-protection-des-renseignements-personnels/decision-traitement-automatise
- Commission d'accès à l'information du Québec, "Principaux changements apportés par la Loi 25": https://www.cai.gouv.qc.ca/protection-renseignements-personnels/sujets-et-domaines-dinteret/principaux-changements-loi-25
- Act Respecting the Protection of Personal Information in the Private Sector, CQLR c. P-39.1, s. 12.1: https://www.legisquebec.gouv.qc.ca/fr/document/lc/P-39.1 (verbatim wording not re-read: legisquebec blocks automated reading, HTTP 403. The content of the section above is drawn from official sources 2 and 3, to be confirmed against the legislative text before publication.)