In March 2026, Anthropic had a 10-hour cascading outage. Authentication failed first. Then the API. Then specific model versions started throwing errors in sequence, Claude Opus first, then Haiku, in what engineers call a “whack-a-mole” pattern: fix one thing, expose the next failure underneath it.
For organisations that had built their operations around Claude Code and the Anthropic API, this was not an inconvenience, but a complete shutdown. No coding pipeline, no customer support automation, no content production, no agentic workflows. Nothing. 😬
This was not a one-off. In June 2025, ChatGPT went offline for over 15 hours when API traffic hit 4.2 million requests per minute, 37% above prior records, while three of OpenAI’s five EU data centres were under maintenance at the same time.
In November 2025, a Cloudflare configuration bug took down ChatGPT, Claude, Perplexity, and Gemini simultaneously, along with roughly 20% of global web traffic. In October 2025, an AWS DNS failure in a single region cascaded through Claude, Amazon Q, and dozens of other services in minutes. Jeeeesus! It’s getting rough. 🫣
This is happening more often, not less. Demand is outpacing infrastructure, and if your business runs on tools you don’t control, you are one incident away from a full stop.
The Case of Full AI Teams
The problem with Full AI Teams
The idea of a Full AI Team, where autonomous agents handle coding, customer support, content, analysis, and research, looks compelling on a spreadsheet. The data holds up, to a point. Organisations integrating AI into core workflows report saving 40 to 60 minutes per worker per day. Developers using AI assistance complete tasks 55% faster. Marketing teams ship campaigns quicker. 👌
The problem is the assumption underneath all of it: that the AI infrastructure will simply be there.
When you build a Full AI Team on top of a single provider’s APIs and cloud-hosted models, you are not building a team, but renting a brain. And the landlord can pull it at any time, through an outage, a pricing change, a model deprecation, or a terms of service update that breaks your integration overnight.
Klarna made this bet hard. Between 2022 and 2024, the company cut roughly 700 customer service positions, taking headcount from around 5,500 to 3,500, and publicly celebrated AI handling two-thirds of its customer service chats. By May 2025, CEO Sebastian Siemiatkowski admitted publicly they had gone too far.
Complaints had risen.
Responses were being described as generic and repetitive.
Klarna started rehiring, and customer service costs climbed back to €50 million per quarter, higher than before the cuts, despite the AI savings. The full automation bet reversed itself within 18 months.
Gartner projects this pattern will repeat across industries. By 2027, half of the organisations that planned to significantly cut customer service headcount due to AI will drop those plans. By 2028, none of the Fortune 500 will have fully eliminated human customer service. Full replacement with no fallback is operationally unsustainable. Klarna found out the expensive way.
Four risks that compound when you go all-in
Provider outages are structural, not occasional.
At 99.3% uptime, which is roughly what major LLM providers deliver, your AI infrastructure is unavailable for about 87 hours per year. That sounds manageable until those hours land during a client deadline or a product launch.
The Cloudflare and AWS incidents in late 2025 showed that failure modes go deeper than individual providers. The shared infrastructure underneath the AI ecosystem creates correlated failures that take down multiple providers at once, which is a category of risk most business continuity plans don’t account for.
Pricing and deprecation can break your economics with very little notice. OpenAI’s GPT-4.5 Preview launched at $75 per million input tokens in February 2025 and was deprecated four and a half months later. Businesses that had built their cost models around it had to start over. Flagship model pricing dropped 83 to 90% over the same period, which benefits buyers but makes 12-month cost forecasting essentially guesswork. Building a business model on LLM API pricing without fallbacks is a bet dressed up as a budget.
Vendor lock-in also works differently here than with conventional software. Switching a standard SaaS tool is painful but predictable. With LLMs, you’re locked in through prompts tuned to one model’s behaviour, embeddings that require full reprocessing if you change models, fine-tuned weights that live on the provider’s platform and don’t travel, and the fact that models produce meaningfully different outputs with identical prompts. Migration costs for AI infrastructure average $315,000 per project. Building portability in from the start costs a fraction of that. ✌️
Skill atrophy is the one that takes the longest to notice. When AI handles everything, the institutional knowledge of how to do that work without AI quietly disappears. ADP Research and Stanford tracked a 20% drop in employment for 22 to 25-year-old developers from late 2022 to mid-2025, with entry-level job postings down 35% since January 2023. MIT research found heavy AI users produced less original work and retained less over time, even when they thought the tool was helping.
When a provider goes down, you cannot fall back on human capability if you’ve already let it erode. Both exits are blocked at the same time.
What a resilient architecture actually looks like
Resilience here does not mean less AI. It means building in layers so that when something breaks, and it will break, operations slow down rather than stop.
An AI gateway is the most important piece.
This sits between your application logic and your LLM providers, normalising inputs and outputs across multiple models. When your primary provider returns an error, the gateway reroutes to a fallback automatically, without touching your code.
The main tools are LiteLLM, open-source and self-hosted, routing to over 2,000 models across 100+ providers; Portkey, a managed option handling 10 billion-plus requests monthly; and OpenRouter, with 500+ models and automatic fallback routing. Gateways also handle semantic caching, storing responses to similar queries so high-frequency requests are served from cache even during provider outages. For any production AI workload, this layer is not optional. 🫡
Prompt portability is the next thing most teams get wrong.
Your prompts, system logic, and architectural decisions should not live inside any provider’s platform.
Store them in versioned, provider-agnostic repositories that can be deployed to a different model through configuration changes, not rewrites. Langfuse, PromptLayer, and Maxim AI handle versioning, environment management, and deployment pipelines for prompt logic. The discipline is writing prompts with template variables separated from model-specific parameters, so they run against Claude, GPT, Gemini, or a local model without modification. What would otherwise be a 48-hour migration crisis becomes a configuration change.
The Model Context Protocol changes the data portability picture significantly.
MCP, released by Anthropic in late 2024 and adopted by Microsoft and Google in 2025, separates context management from prompt logic. Instead of hardcoding integrations between your AI and your data sources, your CRM, your database, your project management tools, you build MCP servers for those resources. Any MCP-compliant model can access them without further integration work. If your primary provider goes down, you route to a different model and your connected data comes with it.
Running a warm local fallback rounds this out.
The performance gap between open-source and proprietary models narrowed from 15 to 20 quality index points in 2024 to roughly 9 by late 2025, with parity expected in 2026. Llama 3.3 70B, Qwen3-235B, and Mistral Small 37B handle classification, summarisation, customer support routing, and code linting at a level that covers the majority of most organisations’ actual AI workload.
The break-even point for self-hosting versus API costs sits at around 2 million tokens per day, with H100 GPU rentals available from $1.49 to $3.50 per hour through providers like Lambda Labs and CoreWeave.
The governance problem nobody is actually solving
Most organisations building Full AI Teams are treating governance as a future problem. It is not.
When autonomous agents operate at machine speed, traditional compliance frameworks break down.
You cannot audit what you cannot trace. IBM’s 2025 report found 13% of organisations had experienced breaches of AI models or applications, and 97% of those lacked proper AI access controls. Shadow AI, employees using personal accounts or unapproved tools because the official infrastructure is too slow or too fragile, costs organisations $670,000 more per breach than standard incidents.
The practical answer is an AI control plane: a centralised layer that monitors agent actions, enforces policy in real time, and keeps end-to-end logs of intent, identity, and outcome for every agentic step. The EU AI Act’s prohibited practices provisions became enforceable in February 2025, with fines reaching €35 million or 7% of global turnover. Full high-risk AI system requirements take effect in August 2026.
For businesses that have deployed agents using service accounts that were never designed for AI workloads, which is most businesses, this means migrating to task-scoped, short-lived credentials.
An agent should only hold the permissions needed for the specific task it is running, for the duration of that task. Broader permissions are a liability across security, compliance, and operational risk if the agent misbehaves or gets manipulated through prompt injection.
The legal picture makes this more urgent. Most AI vendor contracts cap liability at the monthly subscription fee. Berkley Insurance introduced an absolute AI exclusion in its policies. Many professional liability policies do not cover AI-generated errors because coverage is limited to services provided by natural persons. If your Full AI Team causes client damage during an outage or a hallucination event, the legal exposure sits with you, not the provider.
What to actually build
The organisations that hold up through 2026 and beyond are not the ones that went deepest on a single provider, but the ones that built modular, observable architectures from the start.
- Deploy an LLM gateway with circuit breakers and automatic failover.
- Keep active accounts and tested prompts for at least two providers.
- Run a local or private cloud instance of an open-source model for your most critical, highest-frequency pipelines.
- Use MCP for all data and tool integrations.
- Store prompts in versioned, provider-agnostic repositories.
- Structure your business logic to talk to a generic interface rather than a specific vendor’s API.
- Define a recovery time objective for your AI-dependent workflows and actually test it, quarterly.
- Know which tasks a lower-capability fallback model can handle and which genuinely need frontier capability.
- Know where the human fallback is for the processes that cannot degrade gracefully.
- Implement task-scoped agent identities.
- Log all agentic workflows end-to-end.
- Map your AI use cases against EU AI Act risk categories before August 2026.
None of this is complicated to plan.
Most organisations just haven’t done it because nothing has broken badly enough yet.
AI integration has real productivity value, and the direction of travel is not changing. But treating provider APIs as permanent, stable infrastructure is a bet most organisations have not consciously made. Building a decoupled, multi-provider architecture costs a fraction of what a full operational shutdown costs. More to the point, it costs a fraction of building your competitive advantage on a foundation someone else controls.
The brain of your business should not be rented. It should be owned, or at a minimum, insured.
Leave a Reply