Let’s talk

Insights

What still works when your main systems go down?

Lessons from November's Microsoft, US aviation and banking disruptions for disaster recovery, local AI and data sovereignty.

Share this insight

A perspective worth sharing.

Your LinkedIn caption and article image are ready. Edit the caption, copy it, then paste it into your LinkedIn post.

Share on LinkedIn copies your caption and opens the article preview in a new tab. Paste the caption before publishing.

Share on LinkedIn

Conceptual scene of an airport, dark primary server racks and banking terminals, with an illuminated recovery facility.
Conceptual illustration of business continuity across aviation, digital services and banking. AI-generated; not a depiction of the November incidents or a shared cause. Bridges

A single point of failure

Illustrative scenario. A distributor is starting its working day. Trucks are ready, customers are calling and the warehouse has power. Staff cannot open the approved dispatch list. The recovery instructions sit behind the same login as the unavailable system. The backup exists, but nobody can reach the console needed to restore it.

A single point of failure is a dependency whose loss stops a service. It can hide inside an otherwise redundant design. A second environment may share the first environment's identity service, network route or encryption-key access. A second supplier may rely on the same downstream provider. A recovery process may depend on one person being available to approve it. Counting copies does not tell you whether the business can use them independently.

Three disruptions, different causes

Between 12 and 16 November 2026, disruptions affected Microsoft services, US air travel and bank transactions. The banking incident was reported on Monday 16 November, after the Microsoft and aviation incidents the previous week. The available evidence does not establish a shared cause.

Microsoft's disruption began on 12 November. In his preliminary account on 16 November, Microsoft executive Charles Lamanna described a latent defect in Microsoft 365 networking gateways. Earlier Teams failover had moved traffic onto gateways also serving Exchange Online. Additional load and reduced capacity helped trigger the defect, and traffic redirection spread the failure. Some services could move off the affected gateway infrastructure, while Exchange Online could not. The final post-incident review was still pending in that account.

On 13 November, the FAA issued a nationwide ground stop. Its 15 November update described excess network connection attempts that destabilised communications and radar feeds. Isolating two suspected locations helped stabilise the network. Some radio frequencies and radar feeds remained operational, and the FAA said established procedures maintained aircraft safety. Its root-cause investigation remained open.

On 16 November, JPMorgan attributed slower transactions to a software bug in systems on its own premises, according to Reuters. It said other banks could be affected through transactions involving JPMorgan. US Bank said its own systems were not directly affected, while some customers could experience delays sending or receiving Zelle payments involving another institution. Customer complaints across several banks do not establish that each bank suffered its own infrastructure outage.

For a business buying recovery services, these accounts suggest three checks. Ask for evidence that failover can absorb the expected load, that faults can be contained and that critical work can continue when a counterparty is unavailable. The JPMorgan account also shows why moving systems on premises cannot, by itself, settle the resilience question.

Decide which promises must survive

Before choosing a second region, another cloud provider or a backup platform, decide which business services must remain possible, at what capacity and with what information.

For each critical service, name its business owner and define the maximum interruption the organisation can tolerate. Agree what reduced service is acceptable during that interval, what must stop and who can authorise the change. Include customer commitments and the backlog that will accumulate.

Then set the technical targets. The recovery time objective, or RTO, is the target time within which the supporting system must be restored. The recovery point objective, or RPO, specifies how far back its data can be recovered, reflecting the amount of recent data loss the business can tolerate. A restored application can still leave hours of work to validate and reconcile before normal operation resumes. NIST's contingency-planning guidance makes this connection between business impact, system recovery and recovery objectives. It is US federal guidance, not a GCC regulatory requirement.

Backups, disaster recovery and business continuity do different jobs. Backups preserve recoverable data. Disaster recovery restores systems and their supporting resources. Business continuity determines how people keep delivering priority services while that work happens.

Follow the dependencies into the recovery environment

Ask the technology team to trace one critical journey from employee access to a completed business action. Do the same for its proposed fallback. Mark every component the two routes share.

What the plan assumesWhere both routes may still failEvidence to request
Staff can use the recovery applicationBoth environments require the same unavailable single sign-on serviceA controlled exercise using approved emergency access, with authentication and an audit trail
The backup can be restoredThe backup console or encryption keys depend on the failed environmentA successful isolated restore, including key recovery and a usable business record
The second site is independentBoth sites share connectivity, power dependencies or a management serviceA dependency map and a test of the specific failure the second site is meant to survive
The alternative supplier can take overA shared downstream provider or insufficient spare capacity prevents deliveryDisclosed critical dependencies and a rehearsed switchover under agreed load
Someone can make the decisionThe sole approver is absent or unreachableA named deputy with tested authority and an alternative communications channel

This is a proposed management checklist, not a claim that any named incident had all these weaknesses.

Multi-region and multi-cloud arrangements can reduce particular risks. They also bring configuration, staffing and testing costs. Two platforms with the same identity dependency may still stop together. Fund the independence justified by the business impact, including the work needed to keep it usable.

Protect recovery copies against corruption or compromise as well as interruption. A replica can copy a damaging change. Keep appropriately isolated, versioned recovery data and prove that it restores. Store an accessible, protected copy of the recovery instructions, including contacts and the steps needed to recover keys. Emergency access needs controls and review. It should never become a shared password circulated during a crisis.

Design a smaller service that can keep operating

Consider an illustrative UAE distributor whose cloud ERP is unavailable. Its business objective might be to complete a limited set of already authorised deliveries while holding transactions that require fresh financial or inventory checks. This is a proposed design, not a Bridges client result.

Before an incident, the company prepares a protected local dispatch snapshot, with a timestamp, authorised orders and allocated stock. Operations agrees a strict freshness limit and the circumstances in which even those deliveries must stop. The fallback does not accept new credit decisions or treat an uncertain payment as settled.

During the disruption, a trained supervisor checks each permitted delivery against the snapshot. A deputy can approve the operating mode if the usual owner is unavailable. Staff record every action with a unique reference and tell customers which commitments remain possible. The temporary process has a known capacity, so the business can ration service honestly.

When the ERP returns, finance and operations reconcile temporary records before replaying transactions or releasing held orders. A duplicate shipment or payment can be more damaging than a deliberate delay. Our separate guide to reconciling business transactions after an outage covers that return to normal operation.

An alternative banking arrangement needs similar preparation. Treasury would need an active account, authorised signers, funding and tested payment instructions before an incident. Changing the bank name on a slide creates none of that. Where a payment's status is uncertain, verify it before sending another one through a different route.

Give local AI a bounded role

Local AI can make approved operational knowledge easier to use during a cloud or connectivity outage. A locally running model could retrieve the relevant recovery procedure, summarise an accessible incident log or draft a customer update from confirmed information. It cannot restore a bank's payment service or infer the current stock position from an old snapshot.

Microsoft's Foundry Local documentation says models and execution providers can run offline once the necessary assets have been downloaded. That capability applies to inference. It does not establish that a business application, its data access or its authentication works offline.

For the distributor, a useful implementation might answer staff questions using a locally stored, approved operations manual. Every answer should point to its source and show when that source was updated. If the answer depends on unavailable live data, the system should say so and refer the decision to a person.

Test the complete path on the actual recovery hardware. In an isolated exercise, remove the external connections the design claims to survive, then restart the machine and application. Check model files, retrieval indexes, local permissions, encryption keys, drivers, licences and any remote tools the application calls. A warm demonstration with an existing login session proves less than a cold start under the intended outage conditions.

The model should have access only to the information needed for its task. Keep approvals for consequential actions with authorised staff. Log its use, retain approved documents in a form people can read without AI, and test a simple manual procedure if the model fails. Local equipment also needs power, maintenance and a plan for the loss of its site.

Make data sovereignty usable during an outage

Data residency describes where data is held. Sovereignty also concerns the laws, organisations and access arrangements governing it. Recovery also depends on authorised people being able to retrieve and use the data when a provider or control system is unavailable.

A database hosted in-country may still need an overseas service to authenticate its users or release its encryption keys. An application using a local model may still send prompts, logs or tool requests to external services. Trace those routes before describing either arrangement as independent.

For each critical dataset, record the production and backup locations, operator access, key custody, export format and approved recovery destination. Include prompts, retrieval indexes, logs and temporary files where they contain sensitive information. Open the export in the fallback process and complete a permitted business action with it.

For GCC organisations, validate requirements for each jurisdiction, legal entity and sector. Saudi Arabia, for example, has a specific framework governing transfers of personal data outside the Kingdom. Assess the UAE and other jurisdictions separately, including relevant sector and free-zone requirements. Do not assume a GCC-wide localisation rule. Have legal and security owners approve recovery destinations and access arrangements before an emergency, and recheck the applicable rules when the design changes.

An in-country recovery site and locally hosted AI may be appropriate for a particular business. They still need separate failure analysis, tested access and maintainable operating procedures. The location of a server alone cannot guarantee either continuity or compliance.

A practical first month

Start with one business service whose interruption would hurt customers or cash flow. Use the first month to establish and rehearse a limited fallback. A complete DR programme may take longer.

Week one. The business owner and finance team estimate the operational impact of losing that service at its busiest time. Agree the acceptable interruption, minimum operating capacity and transactions that must stop. Identify the person who will make each decision and their deputy.

Week two. Trace the main and recovery paths through identity, network, data, suppliers and approvals. Ask providers for evidence about the dependencies and capacity behind the recovery commitment. Select one shared dependency to remove or one restricted operating mode to build.

Week three. Prepare the fallback data, access, instructions and communications. Assign an owner to refresh and test them. If local AI earns a place, give it a narrow task with measurable acceptance criteria and a non-AI alternative. Include the cost of staffing, licences, hardware and recurring exercises in the decision.

Week four. Run a controlled rehearsal with the chosen dependency unavailable. Measure time to the first valid business action, the volume the fallback can handle and the backlog left afterwards. Check that staff can identify stale information and uncertain transaction status. Complete the return to normal service and record what failed.

Take the rehearsal record into the next budget discussion. Show the business action that succeeded, where work stopped and how much capacity was available. Fund the next change against that evidence, with a named owner and a date to test it again.

Bridges' Disaster Recovery, Business Continuity & Resilience work connects these operating decisions with recovery design. Related capabilities include Managed IT & Cloud Operations, Data Engineering & Analytics and AI Strategy & Readiness.

Share this insight

Related perspectives & expertise