- AI Governance
- Cybersecurity
- Public Sector
AI Incidents Need an Executive Playbook
Dana Saltern8 min read

In the last several weeks, a pattern has become hard to ignore. Public-sector AI governance is moving from principles and acceptable-use policies toward incident readiness, cyber resilience, and named accountability.
For municipal, provincial, federal, and large enterprise leaders, this shift matters. AI is no longer only a tool to pilot, procure, or regulate. It is becoming part of the operating environment in which public trust, service continuity, privacy, and safety must be actively managed.
On October 1, the Government of Canada issued a Security Policy Implementation Notice on cyber security readiness in the frontier AI era. It directs federal organizations to apply baseline cyber controls, identify gaps that could be exploited by AI-enabled threat actors, and work under an “assume breach” model.
That is a clear leadership signal. The question is no longer whether AI introduces risk. The question is whether the organization knows who decides, who acts, who communicates, and who is accountable when something goes wrong.
Why this changed this month
Three recent developments point in the same direction.
First, Canada’s federal direction frames frontier AI as a cyber resilience issue, not only a digital innovation issue. The notice says AI can make weaknesses across systems, networks, cloud services, user access, configurations, and external dependencies easier and faster to find. It also says risk assumptions can become outdated in months, not years.
Second, New York City Council recently put forward a package of AI safety proposals that includes contractor reporting of AI safety incidents, public disclosure timelines, third-party validation, and a plan for responding to AI events that compromise city systems, infrastructure, operations, or public safety. The September 25 announcement from the Council is important for Canadian readers because it shows how municipal AI governance is starting to resemble emergency management, procurement oversight, privacy protection, and cyber command all at once.
Third, the United Nations launched a people-centred smart cities toolbox for the AI era on November 4. Its message is practical: cities should ask what problem an AI system is solving, who could benefit or be harmed, how it will be monitored, who is accountable, and when the city should change course or stop using it.
These are not abstract questions. They are operating questions.
The gap I see in many executive conversations
In my experience, many organizations have made progress on AI policies. They have acceptable-use rules, procurement checklists, privacy impact assessment processes, and early governance committees.
Those are useful. But they are not enough.
The gap appears when leaders ask a harder question: if an AI-enabled service produces harmful advice, exposes sensitive data, amplifies a cyberattack, corrupts a record, or disrupts a public service, what happens in the first 24 hours?
Often, the answer is split across several functions:
- IT owns the platform.
- Cybersecurity owns the incident process.
- Legal owns liability and notification.
- Privacy owns personal information risk.
- Communications owns public messaging.
- The program area owns service delivery.
- Procurement owns the vendor relationship.
- The executive team owns public trust.
That division of responsibility can work during normal operations. During an incident, it can create delay, confusion, and avoidable exposure.
AI makes this harder because the incident may not look like a traditional technology failure. It may be a wrong recommendation made at scale. It may be a chatbot that gives unsafe instructions. It may be a model connected to data it should not access. It may be a supplier tool that changes behavior after an update. It may be a cyber event accelerated by AI-enabled reconnaissance.
The leadership task is to turn distributed responsibility into a clear operating model.
What an AI incident playbook should cover
An AI incident playbook does not need to be long. It needs to be usable under pressure.
For public-sector and enterprise leaders, I would start with six decisions.
Define what counts as an AI incident.
Do not limit the definition to cyber intrusion. Include privacy exposure, harmful outputs, discriminatory impacts, loss of human oversight, unsafe automation, vendor model failure, data misuse, and service disruption.
This definition should be plain enough that frontline staff and suppliers can recognize it. If only specialists can identify the problem, escalation will be late.
Set severity levels before the incident.
A low-risk drafting error is not the same as an AI tool changing an official record or affecting eligibility for a public service. The OECD’s recent work on government transformation notes that controls should be proportionate, with stronger safeguards for tools that can change records or initiate transactions. Its Government Unstuck report also warns that trust depends on clear authority, transparent records, human accountability, and routes to contest decisions.
Severity levels help executives avoid two common errors: underreacting to a material risk, or freezing useful low-risk work because every issue feels urgent.
Name the accountable decision-makers.
AI governance committees can advise. Incidents need decision rights.
- Who can pause a tool?
- Who can order a rollback?
- Who can notify a regulator?
- Who can instruct a vendor?
- Who can approve public communication?
- Who can decide that a service must continue manually while the technology issue is contained?
These questions should not be answered for the first time during a council meeting, media call, board update, or ministerial briefing.
Connect AI governance to cyber, privacy, and business continuity.
The Government of Canada’s October direction is useful because it connects AI-era risk to asset visibility, baseline controls, containment, and recovery. Municipal and provincial organizations should take the same lesson.
AI governance should not sit apart from cyber resilience. It should be wired into incident response, records management, privacy breach procedures, emergency communications, and continuity planning.
This is especially important for smart-city environments. Sensors, cameras, traffic systems, permitting platforms, 311 tools, inspection systems, and resident-facing chatbots may be governed by different contracts and business owners. Residents experience them as one public service environment.
Build supplier obligations into contracts.
Public organizations often depend on vendors for AI models, cloud services, analytics platforms, and managed tools. If the contract is silent on incident notification, audit access, model changes, data use, subcontractors, and shutdown support, the organization may not have what it needs when risk appears.
New York City’s recent proposals are worth watching because they bring contractor reporting and validation directly into the governance conversation. Canadian governments do not need to copy another city’s model, but they should ask whether their current procurement terms would support timely action in an AI incident.
Practice the playbook.
A tabletop exercise is one of the fastest ways to expose weak points. Bring together executives, legal, privacy, IT, cybersecurity, communications, service owners, procurement, and, where appropriate, a key supplier.
Use a realistic scenario. For example: a resident-facing AI assistant gives incorrect eligibility guidance for a benefit program, and the error affects a vulnerable group before staff notice the pattern.
Then ask:
- Who detects the issue?
- Who validates the harm?
- Who pauses or limits the tool?
- Who contacts the supplier?
- Who informs affected residents?
- Who briefs council, the board, or the minister?
- Who approves the restart?
- What record is kept for audit and public accountability?
The value is not the exercise itself. The value is the set of decisions made before pressure arrives.
Privacy and public trust cannot be afterthoughts
On October 29, Canada’s federal, provincial, and territorial privacy regulators called for responsible AI use in health and health research. Their joint statement reinforces the importance of AI systems that protect privacy, are valid and reliable, operate safely, uphold human rights, and demonstrate transparency and accountability.
Although the statement focuses on health, its principles are relevant across the public sector. Many public services involve sensitive personal information, vulnerable populations, consequential decisions, and inherent power imbalances between individuals and the institutions serving them.
Public trust cannot be established simply by demonstrating that an AI system is innovative or efficient. It depends on an organization’s ability to explain how the system works, monitor its impacts, address errors and harms, and intervene or discontinue its use when necessary.
The executive role
Budget 2026 proposes major federal investment in digital transformation, including Digital Transformation Canada and a Government Transformation Fund. The budget chapter on a more efficient government describes AI and emerging technologies as part of improving service delivery, reducing duplication, and strengthening digital sovereignty.
That ambition is understandable. Public organizations are under pressure to deliver faster, better, and more accessible services with limited capacity.
But execution will determine credibility. AI adoption that is not matched with clear accountability will create avoidable risk. AI restraint that is not matched with practical pathways will leave value on the table.
This is where executive guidance matters. Leaders need to ask sharper operating questions, not just approve broader AI strategies.
I would put these questions on the next executive agenda:
- Which AI-enabled services or tools could affect residents, employees, regulated parties, or critical operations?
- Which of those tools can access sensitive data, generate official advice, alter records, or influence decisions?
- Do we have a named executive owner for each higher-risk use case?
- Are cyber, privacy, legal, procurement, communications, and service continuity connected in one incident process?
- Do our contracts require suppliers to notify us quickly and support containment?
- Have we tested the process in the last year?
These questions are simple. The answers often reveal whether strategy has reached execution.
Practical takeaway
Treat AI incident readiness as part of the operating model, not as an appendix to the AI policy. Start with your highest-risk public-facing or decision-support use cases, name decision rights, connect the playbook to cyber and privacy processes, and run one tabletop exercise before the next budget cycle.
If your organization is moving from AI principles to accountable execution, contact us.
Sources
- Direction on Government of Canada Cyber Security Readiness in the Frontier Artificial Intelligence Era: Security Policy Implementation Notice
- New York City Council Unveils Legislative Proposals to Safeguard New Yorkers from Potential Risks of Artificial Intelligence
- New UN guide keeps people in charge of AI in smart cities
- Government Unstuck: The transformation in practice
- Canada’s federal, provincial, and territorial privacy commissioners and ombuds call for the responsible use of artificial intelligence in health and health research
- Chapter 6: A more efficient government, Budget 2026
Written by Dana Saltern, founder of BusinessOne.


