We hold an ISO 42001 AI Management System. That means our AI work is governed by the same kind of structured, auditable controls that regulated enterprises expect of their cybersecurity programs. This policy is the public-facing summary of that system.
01Our principles
Every AI system we design, build, or operate is measured against these five commitments:
- Useful before clever. We solve a real business problem or we don't ship. A working spreadsheet beats an impressive prototype that nobody uses.
- Humans accountable. Someone on your team is always named, trained, and responsible for what the AI does. No orphaned systems.
- Explainable by design. We choose techniques and interfaces that let the people using the system understand — and, where it matters, question — its outputs.
- Private by default. The minimum data, kept for the minimum time, shared with the minimum number of parties. We treat data minimization as a design constraint, not a compliance chore.
- Recoverable when it fails. We plan for the model being wrong, the provider being down, and the data being stale — before we go live.
02Scope & taxonomy
This policy applies to every AI system we design, build, operate, or advise on. For clarity, we classify AI systems into three categories, and our obligations scale with the category:
- AI-assisted — generative AI used inside our own delivery process (drafting documents, generating code) with human review before anything reaches you. Lowest risk.
- AI Assistants — systems deployed to your team that draft, analyze, or recommend, with a human-in-the-loop on consequential actions. Medium risk; explicit disclosure required.
- AI-powered applications — systems that take actions autonomously within scoped boundaries (e.g., routing, scoring, classification) without per-action human review. Highest risk; full ISO 42001 impact assessment required before deployment.
The category of your system is recorded in your Statement of Work and in the AI Registry we maintain for each engagement.
03Human-in-the-loop
We design for meaningful human oversight — not rubber-stamp approval clicks. On every system we deliver:
- A named accountable owner on your side signs off on the scope, risk category, and acceptance criteria before launch.
- Consequential actions (customer-facing communications, financial transactions, legal commitments, safety-relevant operations) require an explicit human confirmation step. We don't hide this behind an "auto-approve" toggle.
- The interface surfaces confidence signals, sources, and the ability to override — so reviewers can do more than pattern-match on the first suggestion.
- We monitor override rates post-launch. If humans are rubber-stamping 100% of outputs, we flag it as a review-quality problem, not a success metric.
04Data use & minimization
- We process the minimum personal information necessary for the task. Anonymization or aggregation is preferred wherever it produces an acceptable result.
- We do not use your data to train general-purpose AI models. We contract with providers whose terms confirm the same.
- Where fine-tuning or retrieval-augmented generation improves a system we're building for you, the resulting model weights or indexes are used only for your engagement and are deleted on termination.
- Data flows between systems are documented in an AI Data Map that is part of your deliverables.
05Fairness & bias
No AI system is neutral by default. For every system that makes or influences decisions about people — customers, employees, applicants, claimants — we:
- Conduct a fairness impact assessment during design, identifying affected groups and plausible bias vectors (including Canadian Human Rights-Act protected grounds).
- Establish measurable fairness criteria appropriate to the use case (e.g., demographic parity, equalized odds) and test against them before deployment.
- Build in opt-out pathways to human review for affected individuals, where the context allows.
- Re-test quarterly or when the underlying model or data distribution changes materially.
We will say "no" to use cases where we do not believe fairness can be assured to an acceptable standard. We would rather decline work than ship something harmful.
06Transparency & disclosure
To end users
- AI-generated or AI-assisted content published or communicated on behalf of a user (e.g., emails drafted by an assistant) is labelled as such where there is reasonable potential for confusion.
- Systems interacting directly with members of the public identify themselves as AI, unless the use case makes that unnecessary or prescribed by sector regulation.
To our clients
- Every AI system we deliver is accompanied by a Model Card documenting its purpose, scope, data sources, limitations, known failure modes, and review cadence.
- We maintain a client-facing AI Registry listing every system we've deployed for you, its risk category, and the name of the human owner on your side.
07Evaluation & testing
- Pre-deployment: task-specific evaluation harness with a labelled ground-truth dataset, red-team probes for prompt injection and data exfiltration, and user-acceptance testing with real reviewers on your team.
- Post-deployment: continuous monitoring of accuracy proxies, override rates, latency, and cost; monthly sampling of outputs for human review; quarterly drift checks against the original evaluation set.
- Changes to the underlying model (provider-side upgrades, fine-tunes, prompt revisions) are re-evaluated before being promoted to production.
08Security of AI systems
AI security inherits from our broader ISO 27001-aligned Information Security Management System and adds AI-specific controls:
- Authenticated, rate-limited, and logged access to AI endpoints.
- Input sanitization and output filtering to mitigate prompt injection.
- Secrets (API keys, customer data) never included in prompts sent to third-party providers in clear text.
- Tenant isolation: your systems run in your cloud tenant or a dedicated tenant, never in a shared multi-tenant inference environment without explicit consent.
- Standard security hygiene: vulnerability management, dependency scanning, annual penetration testing.
09Incident response
When something goes wrong — a harmful output, a data leak, a fairness failure, a provider outage — we follow a defined playbook:
- Contain — disable or degrade the system to a safe state within one hour of confirmed incident.
- Notify — inform the client's named AI owner within four hours; inform privacy authorities within legally required timeframes if personal information is involved.
- Investigate — conduct a structured root-cause analysis; document findings within 10 business days.
- Remediate — deploy a fix and add a regression test to the evaluation harness.
- Learn — share anonymized lessons across our engagements, so other clients benefit from what we've seen.
10Third-party AI providers
Our solutions are built on top of commercial AI platforms. We select providers whose public commitments align with ours — and we will not knowingly route your data to a provider that:
- Trains its general-purpose models on your inputs without your consent.
- Fails to offer reasonable data-processing terms or contractual confidentiality.
- Cannot document its own safety, evaluation, and incident-response practices.
Our current approved-provider list is available to clients under NDA. Providers are reviewed annually and on any material change to their terms. Where a provider ceases to meet our standards, we migrate affected workloads.
11Continuous improvement
This policy and the management system behind it are reviewed quarterly, audited internally once a year, and audited externally on the ISO 42001 surveillance cycle. Findings from audits, incidents, and client feedback drive updates to our controls, our templates, and this policy.
12Contact & reporting
If you believe one of our AI systems — one we've built for your organization or one operated by us — is causing harm, producing unfair outputs, leaking data, or otherwise violating this policy:
- Email: responsible-ai@aistrategists.ca
- Secure form: on request, we will provide a confidential reporting link for clients and end users.
- Mail: AI Strategists · AI Governance Office · Calgary, Alberta, Canada
We will acknowledge receipt within two business days, investigate, and report back to you with findings and next steps. Good-faith reports, including from third parties and end users, are welcome — we'd rather hear from you than learn from a regulator.