# Route Nord — Responsible AI Starter Controls

An operational starter, not certification, legal advice, or a complete compliance program. Adjust to the sector and jurisdiction before deployment.

## Ownership and approved use

[ ] Name the sponsor, workflow owner, human reviewer, and incident contact.
[ ] Document the purpose, users, affected people, and unacceptable outcomes.
[ ] Agree which actions the system may take and which require approval.
[ ] Maintain an approved-tool and approved-use register.

## Data and access

[ ] Identify data categories, source rights, confidentiality, consent, and relevant obligations.
[ ] Minimize data before sending it to a model; redact where appropriate.
[ ] Verify vendor plan, retention, training-use settings, processing region, and contractual terms.
[ ] Use least-privilege credentials in an approved secret store, not browser code.
[ ] Restrict document access per user; retrieval must not bypass permissions.
[ ] Agree retention/deletion in source systems, model logs, automation logs, and backups.

## Output and action controls

[ ] Ground factual output in approved records or documents; show sources where useful.
[ ] Validate required fields and structured output before downstream use.
[ ] Treat messages and retrieved documents as untrusted data; test prompt injection.
[ ] Prevent the model from independently setting approval flags or authorizing actions.
[ ] Use stable IDs to avoid duplicate messages or duplicate system updates.
[ ] Preserve contact preferences and opt-out rules for each channel.
[ ] Provide human escalation and manual fallback.

## Pilot testing

[ ] Test expected, ambiguous, missing, duplicate, malformed, adversarial, and sensitive inputs.
[ ] Confirm no unauthorized retrieval or cross-client data exposure.
[ ] Verify retry behavior does not repeat consequential actions.
[ ] Check factual accuracy, review effort, and professional boundaries.
[ ] Resolve critical failures before go-live; record known limitations.
[ ] Obtain client approval of the exact live actions and permitted data flows.

## Ongoing operations

[ ] Log operational status without unnecessary sensitive payloads.
[ ] Define error alerts, monitoring owner, support hours, and escalation.
[ ] Review changes to prompts, models, integrations, vendor plans, and business rules.
[ ] Re-run affected tests before changed workflows go live.
[ ] Maintain a stop switch, rollback instructions, and incident communication process.
[ ] Measure time saved net of human review and rework.

## Quebec deployment check

[ ] Before deployment, assess whether an EFVP (privacy impact assessment) is required for a new, developed or overhauled information/service system involving personal information, or for communicating/processing that information outside Quebec. Involve the client’s privacy responsible from the outset and document processor agreements where applicable. This check is not a legal compliance guarantee.

Official reference: https://www.cai.gouv.qc.ca/protection-renseignements-personnels/information-entreprises-privees/responsable-protection-renseignements-personnels-entreprise

## Framework reference

NIST AI RMF 1.0 organizes risk work into Govern, Map, Measure, and Manage. These functions are continuous and cross-cutting, not a one-time checklist. The framework is voluntary and is being revised; consult the current official version when beginning an engagement.

https://www.nist.gov/itl/ai-risk-management-framework
https://airc.nist.gov/airmf-resources/airmf/5-sec-core/