---
title: "Common-Mode Failure"
subtitle: "When one provider runs the AI and its oversight, the risk register counts four controls and an outage counts one."
author: "Ritesh Vajariya"
publication: "The Forward View"
publishedAt: 2026-11-17T15:20:20.361Z
updatedAt: 2026-11-17T15:26:29.406Z
pillar: second-order
pillarLabel: "SECOND ORDER"
tags: ["ai-governance", "concentration-risk", "operational-resilience", "financial-regulation", "vendor-risk", "board-oversight"]
readingTimeMinutes: 10
canonical: https://theforwardview.com/essays/common-mode-failure-ai-oversight-provider-concentration
---

# Common-Mode Failure

**TL;DR**

On November 12, Australia’s banking supervisor, APRA, said some individual service providers support institutions that hold more than half of regulated financial assets, and that 80 of the top 100 are based outside Australia. The same day, a global Microsoft 365 outage hit banks and courts, and took out the email and meeting channels that many AI governance programs rely on for oversight. No AI system had to fail for the oversight around it to fail. If one provider runs both the AI and its oversight, you have one layer, not four.

We do not run on Microsoft. On Thursday, November 12, Microsoft still decided how we could talk to our customers.

Our customers do run on it. When their Outlook stopped, our official channel to them stopped with it.

The work did not stop. It moved to WhatsApp threads and phone calls, with texts filling the gaps. That kept us moving. It also put us outside every process either side had agreed to, with no record anyone could rebuild later with confidence.

Nothing in our own stack would have flagged this. The dependency was not on our vendor list. It sat on theirs.

That Thursday was the clearest test I have seen of an assumption many AI governance programs make without writing it down: that the controls will still work on the day the system fails.

## A Regulator Starts Counting

In June, in my newsletter [Access Is Not Property](https://theforwardview.com/essays/access-is-not-property-frontier-model-revocation-enterprise-risk), I asked who else holds the off switch for the model your company depends on. That was a question about one vendor.

The harder question is what happens when most of an economy depends on the same few providers. Their systems then fail the same way at the same moment. Banking supervisors have had words for this for decades, from concentration risk to single points of failure.

On November 12, a banking supervisor started counting. The Australian Prudential Regulation Authority used its [System Risk Outlook](https://www.apra.gov.au/news-and-publications/apras-system-risk-outlook-november-2026) to say it is "intensifying its focus on AI-related risks and improving visibility of material service providers." These are the firms that support a financial institution's critical operations.

One finding is the line boards should read twice: "some individual providers support institutions that collectively account for more than half of APRA-regulated financial assets." Of the top 100 providers that regulated institutions reported, 80 are headquartered or principally operate outside Australia.

APRA does not say these providers are AI companies. It is describing the foundation Australian banks and insurers stand on. And it published that finding next to its new focus on AI.

That pairing matters. AI does not arrive as a separate system. It arrives on the same foundation. So APRA will formalize data collection on these providers, assess how systemically important and how substitutable each one is, and run a 2027 system stress test on AI-driven economic disruption.

The same day, the foundation was tested. A [global Microsoft 365 outage](https://www.goodmorningamerica.com/news/story/outlook-down-microsoft-outage-reported-mail-138088203) began on Thursday, November 12. Microsoft first reported mailbox problems in the Americas. Then it warned the issue "may affect Microsoft 365 service connectivity for users in any region." The outage ran across two workdays.

In New York, the Department of Investigation, after consulting the Civil Court, [told city marshals to suspend](https://www.nyc.gov/assets/doi/downloads/pdf/marshals/Evictions%2011.13.2026.pdf) every eviction scheduled for November 13. According to [Fortune](https://fortune.com/2026/11/13/microsoft-outage-courts-police-hospitals-it-monoculture/), police departments struggled to receive online crime reports, and hospitals fell back on other ways to communicate. The UK's National Health Service was among the institutions [issuing notices](https://abcnews.com/Business/microsoft-outlookoutage-impacts-governments-courts-hospitals-universities/story?id=138107029).

In Mumbai, Tata Motors Passenger Vehicles revised the link for its investor call. Its filing with the National Stock Exchange [cited](https://nsearchives.nseindia.com/corporate/TATAMOTORSSJS_13112026144421_Revisedearningcall.pdf) "the reported outage on Microsoft platforms."

Microsoft's own admin-center advisory, tracked as MO1550821, [listed Microsoft 365 Copilot](https://www.bleepingcomputer.com/news/microsoft/massive-microsoft-365-outage-disrupts-outlook-teams-copilot-worldwide/) among the services customers might be unable to reach. Exchange Online, Teams, SharePoint Online and Defender XDR were on the same list. Its service notes warned that Copilot Chat might not load, and that Copilot might be [unable to reach email, calendar and files](https://www.on-sitetechnology.com/microsoft-365-outage-november-2026/).

Microsoft said there was no indication of malicious activity. Its status updates described servers [failing automated health checks](https://controlplane.com/blog/post/microsoft-365-outage-november-2026-cause) and dropping out of service. That strained the infrastructure that remained.

By Friday, according to Fortune, Microsoft had traced the problem to an issue introduced during an earlier network capacity upgrade, which surfaced after a later network change. It said the incident was unrelated to a cyberattack or to AI systems. Service was broadly restored by Friday.

## The Outage Read

The usual response is reasonable. Cloud outages are as old as cloud computing. Microsoft found the fault and restored service within two working days. Enterprise contracts carry service level agreements and credits for exactly this. Concentration on a few providers buys scale and security that almost no institution could build alone, and a day of degraded email is a tolerable price. Read the same way, APRA's finding is a regulator doing its job and asking for better data.

I think the bargain is sound and the conclusion is wrong.

A service credit pays the customer after the fact. It did nothing to keep an eviction court running on November 13. And it does nothing for a company whose AI incident has to be escalated through the very platform that is down.

## Independent Layers

Process engineers solved this problem long ago. In [What the Plant Floor Knows About Robots](https://theforwardview.com/essays/plant-floor-process-safety-physical-ai-governance), I argued that an embodied agent on a warehouse floor needs a layer-of-protection analysis that shows independent protection layers. The same test works well off the plant floor.

The rule is simple. A safeguard counts as a protection layer only if it is independent of the event that starts the trouble and of every other layer. Two alarms wired to the same sensor count as one alarm. A shutdown valve powered from the same bus as the pump it protects counts for nothing if the bus fails. Engineers call the failure those rules exist to prevent common-mode failure. One cause disables several defenses at once.

Many AI governance programs never apply that test. A typical AI risk register lists a system, its risks, and the controls against them. A human reviewer checks flagged output. An alert notifies the system owner. An escalation path leads to legal and compliance. An incident bridge is where decisions get made. Each control is assessed on its own design. Whether the controls are independent of each other, or of the system they govern, is rarely asked.

In many enterprises, November 12 answered that question. The AI assistant ran on Microsoft. The email that would carry a reviewer's concern ran on Microsoft. So did the Teams channel for escalation and the meeting link for the incident bridge. By the engineer's standard, that stack has one protection layer, however many the register lists.

Where work carried on, it carried on the way mine did. Over WhatsApp and phone calls. That kept the business moving. It left almost nothing an auditor could rebuild afterward.

The counterexample arrived on my phone. During the outage, my bank in the UAE, Mashreq, sent customers a text message. It warned that "a global Microsoft Outlook service disruption" might delay email communications with the bank. It is a small gesture. It is also what an independent layer looks like in practice. The bank knew which channel had failed and reached customers through one that did not share the failure.

None of my US banks sent anything similar. From my side, I cannot tell whether their customer email never depended on Microsoft or nobody thought to say anything. Silence looks the same in both cases. A board should not accept that ambiguity about its own institution.

![A diagram headed "Four controls, one layer". Four controls on the risk register, a human reviewer, an alert, an escalation path and an incident bridge, all sit on one provider that runs the AI and all four controls. Counted on the risk register, four. Counted by independence, one. Illustrative: the pattern many enterprises met on November 12, when the AI, the email, the escalation channel and the incident bridge shared one provider.](https://theforwardview-assets.s3.us-east-1.amazonaws.com/essays/common-mode-failure-ai-oversight-provider-concentration/fig1-independent-layers.png)

None of this required the AI to misbehave. The governance failed at the same moment as the system because it sat on the same foundation. Microsoft says AI played no role in the outage. The lesson is about structure. An AI failure caused by the provider's infrastructure arrives at exactly the moment the oversight built on that infrastructure is unavailable.

Institutions miss this for the reason APRA is building a data collection. From inside any single bank, vendor management can look careful. Contracts get reviewed. Exit plans exist on paper. Concentration shows up only when someone counts across all the institutions at once and finds the same provider names under more than half the system's assets. No single risk committee can produce that view. A supervisor can, and in Australia one now has.

The blind spot also runs past your own vendor list. My company runs nothing on Microsoft. We still lost our formal channel to customers on November 12, because the dependency sat with them.

APRA's word, substitutability, is the board-level question hiding inside the engineering one. Ask whether a provider is reliable and you get a reassuring answer, because most are, most days. Ask how long it would take to run a critical process somewhere else and you get an honest one. In my experience, for many AI-dependent processes today, that answer is months.

The 80-of-100 offshore figure adds a second problem, and it returns to June's question. A provider outside your jurisdiction can be forced to withdraw a service by rules your regulator does not write.

There is a balance-sheet consequence too. Insurers find correlated failure the hardest loss to diversify away. A pooled book assumes its policyholders do not all fail on the same afternoon. An outage that stalls eviction courts and corporate investor calls on the same afternoon is a correlated loss.

The question for a board is narrow. Does any single provider sit underneath both an AI system and the oversight of that system?

![A waffle chart headed "Who the system stands on". Of the top 100 material service providers regulated institutions reported to APRA, 80 are headquartered or principally operate outside Australia and 20 are other. Some individual providers support institutions that hold more than half of APRA-regulated financial assets. Source: APRA System Risk Outlook, November 12, 2026. APRA does not classify these providers as AI companies.](https://theforwardview-assets.s3.us-east-1.amazonaws.com/essays/common-mode-failure-ai-oversight-provider-concentration/fig2-concentration.png)

## What This Means in Three Years

Push the timeline out to 2029 and APRA's stress test will have run. It is scheduled for 2027 and built around AI-driven disruption. Other supervisors read APRA's work. The data request it is formalizing for material service providers is the kind of instrument that tends to spread once one regulator has shown it can be done.

Regulated institutions will split along a line that is already visible. Some will show a supervisor or an insurer that their AI oversight runs on something independent of the AI, and that their critical processes have a tested path to a second provider. Others will find their real dependency map one of two ways: through the regulator's questionnaire or through the next outage.

The first group will negotiate from evidence. The second will negotiate after the fact. That position gets more expensive as supervisors move from asking to measuring.

## What a Decider Should Do

If you sit on a board, ask for one map that management has probably never drawn. Show where your material AI systems run. On the same page, show where the oversight of those systems runs, down to the channel an incident escalation would actually use. Then ask whether any provider appears on both sides. If one does, you have counted one protection layer where the register claims several. Add the providers your largest customers and suppliers run on. That is where my own exposure turned out to be. For the controls behind this map, the AI Governance Institute library, which I founded, covers [third-party model and agent dependencies](https://aigovernance.institute/implement/controls/third-party-model-and-agent-dependencies) and [board and executive assurance reporting](https://aigovernance.institute/implement/controls/board-and-executive-assurance-reporting).

If you run a PE portfolio, do at portfolio level what APRA is doing at system level. Count how many portfolio companies would lose both their AI capability and their means of managing the loss to the same outage. Correlation across a portfolio is concentration risk under another name. The November outage gives you a recent date to test the count against.

If you are a CTO or a buyer, keep one escalation channel and one copy of the incident runbook on infrastructure that shares nothing with your primary provider. Decide in advance which fallback channel is sanctioned, and how a decision made there gets recorded. Test it on a day when nothing is broken. For each critical process that now depends on an AI tool, write down the manual fallback and how long you could sustain it. Then check whether anyone has actually run it. The Institute's control on [AI incident classification and response](https://aigovernance.institute/implement/controls/ai-incident-classification-and-response) is a starting point for the runbook.

If you advise regulators or the public sector, APRA's sequence is a usable template. Start with provider data collected across institutions. End with a stress test that assumes a provider is gone. On November 13, the most visible public-sector casualties were courts and police reporting. Their continuity plans deserve the same independence test as any bank's.

## The Bottom Line

The industry has spent two years debating whether AI systems can be trusted. It has spent far less time asking whether the controls around them can survive the failure they exist to manage. On November 12 and 13, for a large number of organizations, they could not. The AI had nothing to do with it.

APRA has started counting from the supervisor's side. Every board can run the independence test from its own side this quarter. The boards that run it before their regulator asks will hold a much better hand when the questionnaire arrives.

![Pull quote card: "A board should not count a control as a protection layer until it has shown that the control survives the failure of the system it governs."](https://theforwardview-assets.s3.us-east-1.amazonaws.com/essays/common-mode-failure-ai-oversight-provider-concentration/fig3-pullquote.png)

**P.S.**  In your next AI incident review, ask if your primary cloud or model provider were down for a full working day, how would the first escalation reach the person who needs to act on it? If the answer runs through that same provider, you have found the dependency the next three years will reprice.