Canada’s E-23 shows model governance is now an AI operations problem

At a glance

What this is: ValidMind says Canada’s OSFI Guideline E-23 reframes bank AI oversight as continuous lifecycle governance, not a periodic documentation exercise.

Why it matters: For IAM, NHI and AI governance teams, E-23 matters because it pushes accountability, evidence and oversight into the operational systems that manage AI models, agents and their change histories.

Context

Canada’s OSFI Guideline E-23 is a model risk management requirement that now reaches far beyond static validation and periodic review. It matters because adaptive AI systems, especially generative AI and autonomous agents, do not fit the old control rhythm of document, validate, monitor, repeat.

The governance gap is operational, not conceptual: many institutions have policies, committees and inventories, but struggle to keep lineage, approvals, monitoring and remediation current across fast-changing AI estates. In identity terms, the challenge is increasingly about governing machine decision-makers and the lifecycle evidence around them, not only human approvers.

Key questions

Q: How should banks implement continuous AI governance under regulations like E-23?

A: Banks should treat AI governance as an operational workflow that spans inventory, approval, validation, monitoring, escalation and decommissioning. The objective is to keep evidence current as models, data and use cases change, rather than relying on periodic review cycles that quickly become stale in fast-moving GenAI environments.

Q: Why does dynamic AI behaviour create more governance risk than traditional software models?

A: Dynamic AI systems learn, adapt, and make decisions in ways that are harder to predict and fully explain than static applications. That creates risk because bias, data leakage, privacy exposure, and misuse can emerge during normal operation, not just during development. Continuous mapping, measurement, and control are needed to keep those risks visible and manageable.

Q: What are the signs that AI governance is failing in the enterprise?

A: Common warning signs include rapid growth in AI use without matching policy coverage, sensitive files being copied into personal accounts, and a large share of AI apps carrying high or critical risk. Another indicator is weak visibility into who is using which tools and what data they are sending. If teams cannot answer those questions, governance is not working as intended.

Q: What does continuous oversight change for boards and senior management?

A: It shifts oversight from receiving periodic summaries to owning clear escalation paths, risk tolerances and aggregate exposure reporting. Boards should expect evidence that the institution can explain model limitations, track change history and demonstrate how AI systems are governed across their full lifecycle.

Technical breakdown

Why lifecycle governance matters for adaptive AI systems

Traditional model governance assumes a relatively stable asset that can be validated at intervals. Adaptive AI systems break that assumption because prompts, retraining, third-party model updates and changing deployment contexts can alter behaviour between review cycles. Lifecycle governance therefore has to cover development, deployment, monitoring, change management and decommissioning as one control plane, not as separate assurance events. In practice, that means the evidence trail must stay current enough to explain what the system did, why it changed and who approved the change.

Practical implication: Treat AI governance as a continuous operational control, not a periodic compliance review.

What E-23 implies for AI accountability and evidence

E-23 shifts accountability upward and outward. Senior management and boards are expected to understand model limitations, governance processes, escalation paths and aggregate exposure, which means governance artefacts can no longer sit in disconnected spreadsheets or one-off validation reports. The control problem is traceability: institutions need repeatable workflows that preserve lineage, approvals, monitoring results and remediation history across many systems. This is where AI governance starts to resemble identity governance, because the question becomes who approved what, when, under which risk tolerance, and with what ongoing oversight.

Practical implication: Build audit-ready workflows that preserve approvals, monitoring evidence and escalation history across the AI lifecycle.

How third-party models and autonomous agents change the control surface

Externally sourced models and autonomous agents expand the governance perimeter. The institution is no longer only assessing its own model code or data, but also the upstream behaviour of foundation models, the quality of training and operational data, and the way agents make or influence decisions in production. That widens the scope of monitoring to include drift, performance degradation, opaque behaviour and use-case sprawl. When agents are involved, the control question becomes similar to NHI governance: what authority they have, how that authority is bounded, and how changes are tracked over time.

Practical implication: Extend governance scope to third-party AI behaviour, agent authority and change tracking, not just internal model ownership.

NHI Mgmt Group analysis

Continuous AI governance is now the baseline, not an advanced operating model. E-23 reflects a broader shift in which model oversight must track change, use and accountability throughout the full lifecycle. Institutions that still treat validation as a checkpoint are governing yesterday’s systems with yesterday’s assumptions. The practical conclusion is that governance must move into the operational pipeline, where evidence is generated continuously.

AI governance debt is becoming a material control risk. When policies, inventories and attestations exist but cannot be kept current across hundreds of models and agents, the control surface starts to outrun the oversight model. That gap is not merely administrative because it affects escalation, traceability and residual risk decisions. The practitioner takeaway is to measure whether governance artefacts are actually operational, not merely documented.

E-23 is a signal that machine decision-makers now need lifecycle control similar to privileged identity governance. As soon as AI systems influence decisions, interact with customers and evolve in production, their authority must be bounded, reviewed and revoked like any other high-risk digital capability. This is where AI governance and identity governance intersect most clearly. The field should expect stronger demands for bounded authority, auditable approvals and change provenance.

Third-party and autonomous AI force governance teams to think beyond model ownership. The decisive issue is no longer who built the model, but who controls its runtime behaviour, update path and downstream use cases. That widens the accountability chain and makes cross-functional operating models unavoidable. The practitioner conclusion is to govern the AI system as a living service with accountable owners, not as a static artifact.

Canada’s E-23 is an early indicator of where regulatory expectations are heading globally. The direction of travel is continuous oversight, enterprise accountability and evidence-driven supervision, especially where AI can alter outcomes at scale. Financial institutions outside Canada should expect similar demands to spread through banking and adjacent sectors. The practical implication is to align governance architecture now rather than retrofit it later.

From our research library:

• Only 44% of organisations have implemented any policies to manage their AI agents, despite 92% agreeing that governing AI agents is critical to enterprise security, according to the 2026 Infrastructure Identity Survey.
• 52% of respondents see AI security decision-making power shifting toward platform and infrastructure teams rather than the executive suite, according to the 2026 Infrastructure Identity Survey.
• Read next: AI Governance Guide

What this signals

AI governance debt is now a board-level operating problem. Institutions can no longer rely on static policies if inventories, approvals and monitoring evidence cannot be maintained continuously across GenAI and agentic systems. That shifts the work from policy design to workflow design, which is where most governance programmes will either scale or stall.

Lifecycle evidence is the new control surface. Once models and agents change faster than quarterly review cycles, the decisive question becomes whether change history, lineage and remediation can be reconstructed on demand. Banks that cannot answer that question will struggle to demonstrate control effectiveness when supervisors ask for evidence.

Machine decision-makers need bounded authority, not just documented intent. As AI systems influence customer outcomes and internal decisions, governance has to specify who may change them, under what risk tolerance, and with what audit trail. That is the practical bridge between AI oversight and identity governance.

For practitioners

• Standardise lifecycle evidence capture Automate the collection of inventories, approvals, validation results, monitoring outputs and remediation history so evidence remains current across the AI estate.
• Define bounded authority for AI systems Document who can deploy, change, approve, suspend and decommission AI systems, including autonomous agents that influence decisions or customer interactions.
• Centralise governance workflows Move governance reviews out of spreadsheets and email chains into repeatable workflows that preserve lineage, escalation paths and sign-off history.
• Track third-party model changes continuously Require ongoing review of externally sourced models, foundation model updates and vendor changes so drift and performance degradation are detected as operational events.
• Align board reporting to aggregate AI risk Report model exposure, escalation patterns and control exceptions at portfolio level so senior leaders can govern aggregate risk, not isolated projects.

Key takeaways

• E-23 reinforces that AI governance must operate across the full lifecycle, not just at validation checkpoints.
• The main control gap is operational traceability, where approvals, lineage and remediation evidence fall out of sync with system change.
• Banks that can govern AI as a live service, with bounded authority and auditable workflows, will be better positioned than those relying on static policy packs.

Key terms

• API Governance Debt: API governance debt is the accumulation of weak, inconsistent, or missing controls around how APIs are designed, approved, documented, secured, and retired. It creates hidden risk over time. Technically, it includes gaps in authentication, authorization, versioning, inventory, ownership, logging, and policy enforcement that make APIs harder to trust and govern.
• Lifecycle Governance: Lifecycle governance is the set of controls that cover creation, assignment, review, rotation, and retirement of identities and credentials. For NHIs, it is the difference between a temporary automation asset and a persistent access risk. Strong lifecycle governance keeps ownership and expiry tied to actual business use.
• Aggregate Model Risk Exposure: The combined risk posture created by all models and AI systems across an organisation, rather than the risk of any single use case. It is the portfolio view senior leaders need to understand concentration, drift, escalation patterns and unresolved exceptions.
• Bounded Authority: A bounded authority model gives an agent access because it belongs to a workflow or system, not because it represents a person. Its permissions should be limited to the exact operational boundary of the run, with short-lived credentials and separate policies for each environment it can touch.

Deepen your knowledge

NHI Foundation Level course, the industry's only accredited NHI security programme, covers NHI governance, agentic AI identity and machine identity security alongside the governance patterns modern security teams need. It is a fit for practitioners building accountable identity controls across AI and automation programmes.