Opinion · AI Governance & Trust
Enterprise AI Leaders Should Treat the Outage Week as a Rehearsal
AI did not cause the November outages as far as anyone can show, but the next wave may involve it. Five steps enterprise data and AI leaders can take this quarter to be ready.
By The AI Enterprise Magazine Team6 min readOpinion · Weekly briefing

Key takeaways
- Know what agents did. Build an inventory with identities and logs for every AI system touching production.
- Size change capacity. Agree with change owners on how much AI-driven change can be absorbed safely.
- Prioritize patches. Match AI-speed discovery with environment-specific triage.
- Name owners and drill. Assign an accountable owner and kill switch per agent, and rehearse an AI-suspected incident.
Every few years, a cluster of failures forces organizations to look at risks they had been treating as theoretical. The week of November 12, when outages hit Microsoft, the FAA, JPMorgan Chase, United Airlines, Apple and the NSA in quick succession, should be one of those moments for enterprise data and AI leaders. Not because AI caused it. There is no evidence that it did, and every organization involved has offered an explanation that does not involve AI. It should be one of those moments because it was a rehearsal.
The failures followed the pattern that AI is about to make far more common: a change goes into a complex system, something unexpected happens, and the damage spreads faster than people can understand it. Officials described the week as a matter of patching and upgrades. Lawmakers including Senator Mark Warner pointed to AI-driven vulnerability discovery as the force behind the rush to patch. Whatever the investigations conclude about these particular events, the conditions for the next round are already in place.
The right response is not to wait for a root cause report. It is to treat this week as a dress rehearsal and ask what would have happened if AI had been involved, and whether your organization would have been ready.
Would you know what your AI systems did during an outage?
Start with the question every incident team will face once AI is embedded in operations: did our AI do this? At the FAA, officials could answer quickly because the agency's AI planning tool, SMART, is advisory and not connected to the operational network that failed, according to TechRepublic. Even so, suspicion spread for days.
Most enterprises would struggle to answer as cleanly. The fix is unglamorous: an inventory of every AI system and agent with access to production, a distinct identity for each one, and logs that record what each did and when. Vendors are moving in this direction. Microsoft announced at Ignite on November 17 that its Agent 365 system will issue identities to agents running on Salesforce, AWS, Google, Databricks, Snowflake and other platforms, and that it can now discover local coding agents on employee machines, according to its Agent 365 blog. Whatever tooling an organization chooses, the test is simple. In the first hour of an incident, can someone produce a list of every agent action that touched the affected systems? If not, that is the first gap to close.
Can your change processes absorb AI-speed change?
Nearly every stated cause last week was a change: a July update at Microsoft, a router installation at the FAA, an internal update at Apple, a software bug at Chase, maintenance at the NSA. Meanwhile, the volume of change is rising sharply. Microsoft's monthly security releases went from 113 fixes in January to 1,407 in October and 1,093 in November, according to Tenable, with Microsoft citing its own AI scanning as one reason. Coding agents are writing more production code, and agents with operational permissions are beginning to make changes themselves.
Enterprise AI leaders should ask the owners of change and incident management a direct question: how many changes per week can we deploy safely, and how close are we to that limit? If AI programs will increase the rate of change, the change process needs more capacity, better automation or both. Progressive rollouts, automated rollback and testing environments that match production are not new ideas. They are now prerequisites for adopting AI responsibly, because every AI initiative that accelerates change also accelerates the arrival of the next latent defect.
Is your patching strategy built for ten times the volume?
The patch theory of last week's outages has not been confirmed. The Washington Sun's report that bank executives and NSA officials blamed AI-driven patching was based on anonymous sources, and no organization has confirmed it on the record. But the dilemma it describes is real. Anthropic, whose Mythos model found more than 10,000 serious vulnerabilities for its partners in about a month, has written that "the bottleneck in cybersecurity is now verifying, disclosing, and patching" what such models find, in its Project Glasswing update.
Organizations cannot patch everything immediately without breaking things, and cannot wait without leaving doors open. The way through is prioritization grounded in their own environment: knowing which vulnerabilities are actually reachable and exploitable in the systems they run, and moving fast on those while scheduling the rest through normal, tested processes. That requires accurate asset inventories, dependency maps and, increasingly, AI tools to do the triage. Data and AI leaders have a role here, because the triage problem is fundamentally a data problem.
Who is accountable when an agent breaks something?
Genpact's survey of 2,005 executives, published November 12, found that accountability for agent errors is split, with 26% holding the technology team responsible, 23% the business leader and only 20% naming a dedicated agent owner, according to Genpact's report. Only 17% embed governance across the full agent lifecycle.
That ambiguity becomes a serious problem during an outage. If an agent's change contributes to a failure, someone has to be able to pause it, explain it and own the fix. Organizations should name an accountable owner for every agent with production access, define who can shut it down, and make sure that person is part of incident response. AT&T, which runs nearly 500 agentic workflows, offers one model. It keeps humans in the loop on every new workflow until accuracy reaches the mid- to high-90% range, then monitors continuously, Let's Data Science reported. Autonomy, in that model, is earned and can be revoked.
How will you communicate when AI is suspected?
The outage week showed how quickly trust erodes. Officials gave inconsistent accounts. Homeland Security Secretary Markwayne Mullin relayed a contact's view that AI tools may have overwhelmed older systems, while also saying he was confident hackers were not involved, IBTimes UK reported. Rumors about a coming internet shutdown spread widely enough that news organizations had to debunk them.
Enterprises deploying AI in customer-facing or critical systems need a communications plan for incidents in which AI is suspected, whether or not it turns out to be involved. That plan should be built on evidence the organization can actually produce, which brings the question back to logging and inventory. A credible denial requires proof, and a credible admission requires knowing what happened.
Where should enterprise AI leaders start?
Five steps can begin this quarter. Build or validate an inventory of every AI system and agent with access to production, with identities and logs for each. Meet with the owners of change and incident management to agree on change capacity and the controls that new AI initiatives must respect. Put AI-assisted vulnerability triage on the roadmap so that patch volume is matched with prioritization capacity. Name an accountable owner and a kill switch for every production agent. And run a tabletop exercise based on last week's events, asking at each step whether the organization could have determined whether AI was involved.
None of these steps requires slowing down AI adoption. Each of them makes it more likely that faster adoption leads somewhere good. The capabilities of models and agents are advancing faster than most organizations' ability to observe and control them. Last week, that gap did not, as far as anyone can tell, cause the failures. The next time, it may, and the organizations that used this week as a rehearsal will be the ones that can tell the difference.
Frequently asked questions
What should enterprise AI leaders learn from the November 2026 outages?
That AI-speed change, patch volume and limited oversight create conditions for outages, and that organizations need to be able to determine quickly whether AI was involved.
How can companies prepare for an AI-related outage?
Inventory AI systems and agents, log their actions, name accountable owners with the authority to shut them down, and run tabletop exercises based on realistic incidents.
Who should own AI agents in production?
Each agent with production access should have a named accountable owner. Genpact found only 20% of enterprises assign a dedicated agent owner today.
Opinion pieces reflect the views of The AI Enterprise Magazine Team.
Sources
- TechRepublic: FAA Says Its New AI Didn't Ground US Flights — So What Does SMART Actually Control?
- Microsoft Community Hub: What's new in Microsoft Agent 365: Ignite 2026
- Tenable: Microsoft's November 2026 Patch Tuesday addresses 1093 CVEs
- Anthropic: Expanding Project Glasswing
- Genpact: Autonomy in Practice (PDF report)
- Let's Data Science: AT&T runs nearly 500 agentic AI workflows in production
- IBTimes UK: '15 Days Until America Goes Dark': List of 13 Major Tech, Banking, Aviation Outages Fuels Fears of US Net Shutdown
- Implicator.ai: Outages Rattle Americans as Spy Chief Blames Patching
Filed under: AI Governance & Trust, AI Strategy & ROI, Security & Fraud
Become a Sponsor






