The useful idea

Buy continuity for a defined business capability, then test the dependencies that both its normal and fallback paths still share.

This week exposed the layers around inference

On November 17, 2026, Anthropic reported errors editing documents through its Google Docs connector while reading remained available. A separate incident that day affected Apple sign-in while Google and email sign-in were unaffected. These are provider incident reports, not independent root-cause investigations. They establish narrower failures than a blanket statement that an AI service was unavailable.

On November 19, Oracle published engineering guidance connecting inference economics with resilience. It argues that alternative execution paths need qualification and capacity, and that recovery can include a reduced service mode. This is vendor guidance supporting Oracle’s infrastructure offering, not evidence that any particular architecture has passed a customer’s continuity requirements.

Together, these developments suggest a useful procurement question: what business capability survives when one dependency fails? Metrosion’s assessment is that counting model providers answers too little. A fallback deserves credit only for the failures it can actually withstand.

Map the shared dependencies before adding providers

For a hypothetical insurance operations assistant, the useful capability might be preparing a case summary from authorized records. Trace both proposed execution paths from the employee’s sign-in through retrieval, inference, validation, and saving the result. Two model endpoints may still depend on one gateway, identity service, document repository, or destination application. If either path needs the failed component, switching models cannot restore that capability.

Record dependencies by operation, not just vendor name. Reading an existing record, obtaining a fresh access token, changing a routing policy, and writing a decision can have different failure behavior. A healthy model endpoint does not demonstrate that a newly signed-in reviewer can retrieve the evidence or approve the output.

Ask suppliers which dependencies they disclose and which remain unknown. Do not assume different brands imply independent infrastructure or release processes. Where independence cannot be established, record that uncertainty in the continuity decision rather than multiplying advertised availability percentages as if failures were unrelated.

Keep the recovery mechanism available

AWS’s older Builders’ Library article on static stability offers an established architectural principle: prepare sufficient running resources and state so ordinary operation can continue through a dependency impairment. It distinguishes control-plane changes, such as provisioning resources, from the data plane that performs existing work. The article dates to December 5, 2019; it is background engineering guidance, not this week’s announcement.

Applied to AI, our recommendation is to examine what must change at the moment of failure. Does the fallback require an operator to sign into an unavailable console, fetch configuration from the same failed service, or provision capacity that has never been reserved? A recovery plan with those steps has additional dependencies to test.

Pre-positioning approved configuration and capacity can reduce that exposure, but it creates cost and security tradeoffs. Cached policy must have an explicit freshness limit. Loss of current authorization evidence may require stopping sensitive operations. Continuing to serve requests is not success if it silently extends revoked access or bypasses required review.

Define useful reduced service

Return to the hypothetical case-summary workflow. If the destination cannot accept writes, the approved continuity mode might let staff inspect existing evidence and prepare a clearly marked draft, while withholding any claim that the case record was updated. If current source records cannot be retrieved, even drafting may need to stop. The distinction follows the information required for the task, not a preference to keep the interface looking operational.

Set acceptance conditions for each mode: permitted task classes, maximum evidence age, allowed data locations, available reviewers, and the point at which work must queue or stop. Explain the limitation in the interface and preserve task status. A reduced mode needs its own quality evaluation; it should not inherit acceptance from the normal route.

Business owners should choose the minimum useful capability before engineers price the alternatives. Maintaining two fully provisioned environments may be justified for some operations. For others, a bounded queue and a rehearsed manual process may provide better value. Include the cost of maintaining those procedures and clearing delayed work, not only standby compute.

Test a failure, then test its neighbor

A meaningful exercise should disable more than the preferred model endpoint. In an isolated environment with synthetic records, separately remove access to the routing configuration service, token renewal, retrieval, and the write destination. For each exercise, predict which capabilities should remain and which must stop. Never introduce these faults into production without an appropriately controlled test plan.

Run the surviving route at the load it is expected to carry during disruption. Measure completed, accepted business tasks and the age of queued work, alongside latency and errors. Check that the fallback can emit its evidence and audit records; a result without the required record may be unusable even when the model answered correctly.

Finally, restore the dependency and inspect the transition back. Reconcile uncertain writes before releasing queued work. Verify that temporary restrictions end deliberately and that operators can distinguish completed tasks from drafts awaiting action. Recovery is incomplete while the business still cannot account for its work.

Make one continuity claim reviewable

This week, choose one consequential workflow and write a single claim: under a named dependency failure, a specified capability remains available within defined quality, access, and timing limits. Attach the dependency map, the test result, the remaining unknowns, and the person responsible for accepting them.

Use that evidence to decide whether the next investment should be another model, an independent recovery mechanism, reserved capacity, or a simpler reduced service. The objective is a continuity promise that operators and decision makers can understand and verify.

Sources & further reading

Primary sources informing this analysis. Recommendations and examples are Metrosion’s editorial perspective.

  1. Anthropic — Google Docs connector editing incident (November 17, 2026 Pacific; report November 18 UTC)
  2. Anthropic — Apple sign-in incident (November 17, 2026)
  3. David Murray, Oracle — The inference economy: Lower costs and build resilient enterprise AI with OCI (November 19, 2026)
  4. Mike Furr, AWS Builders’ Library — Static stability using Availability Zones (December 5, 2019)

Keep exploring

Back to the reading room