Resilience

AI resilience isn’t uptime. It’s what you can prove after the incident.

Every institution has a business continuity plan for servers, clouds, and SaaS. Almost none has one for the layer now making decisions inside its workflows. AI resilience has two halves — and most organizations are building only the first.

In 2026, a major LLM-provider outage took down customer-facing chatbots, internal help desks, and document workflows across thousands of organizations at once — companies that had mapped AWS as a critical dependency but never put their model provider on the continuity plan. The lesson is simple and already widely drawn: LLM providers are infrastructure now. Map them, add fallbacks, plan for their downtime like any other dependency.

the incident happens anyway, and now you must explain it. The database an agent deleted. The decision a model got wrong. The output that reached a customer. In that moment, uptime statistics are worthless. The questions are forensic: What exactly did the AI do? Under which instruction? Checked against which rule? Who reviewed it? Can you show me — without relying on the very system that failed?

Post-incident reconstruction from application logs is where AI investigations die: the logs are incomplete, mutable, scattered across vendors — or they lived inside the system that just failed. A record you assemble after the incident is a narrative. A record produced at the moment of each decision, signed, immutable, and stored outside the AI’s own reach, is evidence. Regulators have already written this expectation down: the EU AI Act requires high-risk AI to record events automatically across its lifetime, and operators to keep those logs producible. Resilience, in the regulator’s reading, includes the trail.

its AI can fail — models will fail — but the institution doesn’t fail with it. It switches providers without losing its evidence, because the evidence layer is independent of any model. It answers the auditor during the incident, not months after. And it re-enters production with proof of what changed. That is the difference between an organization that had an incident and one that is defined by it.

The first half keeps you running. The second half keeps you defensible. We build the second half — the forensic evidence infrastructure underneath the AI, independent of the model, alive in every decision.

― basis: EU AI Act Art. 12 (record-keeping) · Art. 19 (logs) · Art. 26 (deployer obligations) · 2026 LLM-provider outage analyses (business-continuity coverage)