The Magic That Stops Working
The AI tooling your business now depends on is failing every 1.8 days on average. Almost no one has built for it.
There is a sequence of incidents from the first three weeks of June 2026 that ought to be in every operations director's risk register, and almost certainly is not.
On 1 June, Microsoft Copilot suffered a routing failure that sent traffic to unhealthy infrastructure across the US morning. On 2 June, Anthropic's Claude experienced a major global service disruption affecting the API, the Claude Code CLI, and every model from Haiku to Opus (Thoughtworks, 2026). On 11 June, Microsoft Copilot Chat and the Microsoft 365 Office portal went down for a significant subset of users after a misconfigured authentication update broke token generation (Windows News, 2026a). On 15 June, Copilot suffered its third significant outage of the year, lasting over eight hours and peaking at more than 4,500 user reports per hour on Downdetector (Windows Forum, 2026). And between 5 and 16 June, Anthropic Claude recorded ten significant service disruptions in twelve days, a mean time between failures of roughly one day (Tech Times, 2026).
This is not an unrepresentative window. It is the new baseline.
Ookla, the company that runs Downdetector, published an analysis on 10 June 2026 of 471 days of US incident data covering ChatGPT, Claude, Gemini, Microsoft Copilot, AWS, and Microsoft Azure. The study, authored by Ookla industry analyst Luke Kehoe, recorded 3.72 million user-reported incidents across the period. High-signal disruption days, defined as days where reports exceeded ten times a platform's median daily volume, rose from six in the first quarter of 2025 to fifty-one in the first quarter of 2026 (Mobile World Live, 2026; Broadband Breakfast, 2026). That is a 750% year-on-year increase.
Anthropic's Claude alone accounted for 39 of those 51 disruption days. The platform recorded almost zero Downdetector reports in early 2025, then moved into a sustained reporting baseline from mid-July as enterprise adoption accelerated. By March 2026, Claude was generating 192,773 user-reported incidents in a single month, almost three times the February total (InfoTech Lead, 2026).
The conventional reading of these numbers is that AI vendors are struggling with demand. That is true. It is also not the part of the story that matters most to the businesses that have adopted their products.
What changed
In 2023 and 2024, AI tooling was a novelty layered on top of work that would happen with or without it. A timed-out ChatGPT prompt was a small inconvenience. A Copilot that occasionally lost context cost a few seconds.
In 2026, that relationship has reversed. According to a February 2026 SemiAnalysis report, Claude Code now accounts for roughly 4% of all public GitHub commits, processing more than 135,000 commits per day (Tech Times, 2026). When that service goes down, it is not a chatbot that becomes unavailable. It is 4% of the global software development workforce that pauses, including the CI/CD pipelines that have embedded Claude Code as an autonomous component.
Microsoft 365 Copilot now sits at the centre of millions of enterprise productivity workflows. When the assistant fails, the failure does not stay contained inside the chat panel. The 11 June outage cascaded into broader Office portal access because Copilot is no longer a standalone service but a mesh of AI models, Graph API calls, semantic indexing, and user-facing experiences that all must align (Windows News, 2026a).
Ookla's June report quantifies the operational reality. Sixty-eight percent of surveyed enterprises now run at least one AI-dependent business process. Forty-one percent run three or more. Agentic AI systems specifically, the autonomous workflow tools that have absorbed most of the enterprise AI budget in the past year, experience 210% more incidents than basic generative AI tools (Windows News, 2026b).
The pattern across every credible source converges on a single conclusion. AI is no longer experimental. It is infrastructure. And it is failing at rates that would be unacceptable in any other category of business-critical system.
Why this is not a vendor problem
The temptation, reading the data, is to assume the fault lies with the providers. Anthropic, Microsoft, OpenAI, and Google between them ought to be capable of running reliable services. They are, by most measures, the best-resourced technology companies in the world.
The temptation should be resisted. The vendor reliability is not the central issue.
The central issue is that businesses adopted AI tooling at a pace that assumed infrastructure-grade reliability, without any of the architectural discipline that businesses apply to other infrastructure. When a mid-market business deploys Microsoft Exchange, it understands the service-level agreement, plans for outages, maintains backup procedures, and has documented degraded-mode workflows for when email is unavailable. When the same business deploys Microsoft 365 Copilot, almost none of those disciplines are applied. The assistant is treated as a feature, not as infrastructure, and the lack of architectural preparation becomes painfully visible the first time the service is down for eight hours during a Monday morning.
Thoughtworks, writing after the 2 June Claude outage, identified the structural cause precisely. "Hardcoding a specific provider's API endpoint into your application was an acceptable availability strategy" in the early days of AI adoption, but "in 2026, it's a single point of failure that's a very real threat to business continuity" (Thoughtworks, 2026). The infrastructure was deployed with the architectural patterns of an experiment, and the experiment grew into infrastructure faster than the architecture caught up.
A Fortune 500 financial firm CIO, quoted at Microsoft Build in May 2026, made the same point in practical terms. "We can't put AI into our traders' workflows if there's a chance it'll blink out during a market-moving event" (Windows News, 2026a). The firm had delayed broader Copilot deployment specifically because the reliability story did not match the workflow integration story being sold alongside it.
This is the gap that needs closing. Not the vendors' uptime, which will improve as capacity catches up with demand, but the architectural discipline inside the businesses that depend on them.
What good actually looks like
The interventions that close the infrastructure lag gap are well-understood by practitioners and almost entirely absent from mid-market AI deployments.
Map AI dependencies onto critical business processes. Most organisations do not have this map. They know which licences they have purchased. They do not know which business processes will stop working if a specific AI service is unavailable for four hours. The first job is to produce the map.
Build documented degraded-mode procedures for each critical process. If Copilot goes down for eight hours on a Monday morning, what does the marketing team do? What does the customer support team do? What does the engineering team do? In most organisations, the answer is "panic and improvise." In organisations that have done this work properly, the answer is a documented fallback procedure that staff can execute without further escalation.
Adopt multi-provider abstractions for genuinely critical AI workloads. The simplest version is an API gateway that can route the same request to Claude, GPT, or Gemini depending on availability. The more sophisticated version uses semantic monitoring to detect quality degradation before users report it. Neither is cheap to implement. Both are cheaper than the alternative when the primary provider goes dark during a market event.
Treat AI services as infrastructure for procurement purposes. Insurance underwriters have begun to exclude business interruption losses caused by AI service outages unless the policyholder can demonstrate reasonable preparedness (Windows News, 2026b). Procurement teams are now asking suppliers, in RFPs, how many high-signal disruption days their AI platforms experienced in the previous quarter. The market is moving toward treating AI as infrastructure regardless of whether the buyer's internal processes have caught up.
Run failure drills. The same way mature IT departments run disaster recovery exercises against their email and identity systems, they should now be running scheduled exercises against their AI dependencies. Most organisations have never done this. The ones that have, find the gaps in advance rather than during an actual outage.
None of this is technically difficult. None of it requires buying more software. It requires applying the same architectural seriousness to AI that the business already applies to every other piece of infrastructure it depends on. The reason most organisations have not done it is the same reason most organisations have not done the unglamorous fundamental work this column has previously covered. The discipline is operational, not technological, and the work produces nothing demo-able until it is finished.
Where Neurotic comes in
For most businesses, the gap between current AI dependency and current AI resilience cannot be closed by buying another platform or hiring another engineer. It is bridged by an independent technical team doing the unglamorous work of mapping the dependencies, documenting the degraded-mode procedures, designing the failover architecture, and running the first failure drill.
Neurotic's technology audit and cybersecurity audit services are built specifically for this work. We are independent of any AI platform or cloud vendor, which means the recommendations are driven by the actual risk profile of your business rather than by what someone is trying to sell you. If your organisation has deployed AI tooling at scale in the past eighteen months and has not formally mapped the dependencies, documented the fallbacks, or tested a real failure, the exposure is almost certainly larger than the operations team realises, and a proper independent assessment is overdue.
The next eight-hour outage is coming. The question is whether your business is ready for it.
Talk to us for an audit today → [email protected]
References
Broadband Breakfast (2026) Ookla: AI Outages Are Becoming a Business Reliability Problem. Available at: https://broadbandbreakfast.com/ookla-ai-outages-are-becoming-a-business-reliability-problem/ [Accessed 22 June 2026].
InfoTech Lead (2026) AI Service Outages Surge as ChatGPT, Claude and Copilot Face Rising Reliability Challenges in Enterprise Workflows. Available at: https://infotechlead.com/artificial-intelligence/ai-service-outages-surge-as-chatgpt-claude-and-copilot-face-rising-reliability-challenges-in-enterprise-workflows-96470 [Accessed 22 June 2026].
Mobile World Live (2026) Ookla finds AI platform outages surge as adoption grows. Available at: https://www.mobileworldlive.com/ai-cloud/ookla-finds-ai-platform-outages-surge-as-adoption-grows/ [Accessed 22 June 2026].
Tech Times (2026) Claude Outage: Tenth Disruption in 12 Days Exposes Anthropic Infrastructure Strain, 16 June 2026. Available at: https://www.techtimes.com/articles/318514/20260616/claude-outage-tenth-disruption-12-days-exposes-anthropic-infrastructure-strain.htm [Accessed 22 June 2026].
Thoughtworks (2026) Claude outage, June 2026: Reckoning with AI's increasing status as infrastructure. Available at: https://www.thoughtworks.com/insights/blog/generative-ai/claude-outage-june-2026 [Accessed 22 June 2026].
Windows Forum (2026) Microsoft Copilot Outage June 11, 2026: Productivity Layer Failure Explained. Available at: https://windowsforum.com/threads/microsoft-copilot-outage-june-11-2026-productivity-layer-failure-explained.425458/ [Accessed 22 June 2026].
Windows News (2026a) Microsoft 365 Copilot Outage Highlights: AI Needs Infrastructure-Grade Reliability. Available at: https://windowsnews.ai/article/microsoft-365-copilot-outage-highlights-ai-needs-infrastructure-grade-reliability.425494 [Accessed 22 June 2026].
Windows News (2026b) Ookla Warns AI Reliability Now a Business-Critical Risk After 3.72M Outage Reports in 16 Months. Available at: https://windowsnews.ai/article/ookla-warns-ai-reliability-now-a-business-critical-risk-after-372m-outage-reports-in-16-months.425034 [Accessed 22 June 2026].
