Critical cloud does not mean approved cloud

Four cloud providers have acquired a new regulatory label in UK finance. The useful first response is to remove that label from the supplier scorecard.

On 13 July 2026, Amazon Web Services EMEA SARL, Google Cloud EMEA Limited, Microsoft Ireland Operations Limited and Oracle Corporation UK Limited became the first Critical Third Parties under the UK’s financial-services regime. The Bank of England, Prudential Regulation Authority and Financial Conduct Authority can now oversee the resilience of the critical services those entities provide to the sector.

That is a material change at system level. It is not a purchasing recommendation. The regulators state the distinction unusually clearly: designation is not authorisation, does not make a provider inherently more resilient and does not make it better suited to a particular firm’s needs. Regulated firms still own their outsourcing and third-party arrangements.

For a smaller bank, insurer, lender, payments firm or regulated fintech, the designation creates a better information channel around concentrated infrastructure. It does not complete the firm’s due diligence. The practical job is to connect one named cloud service to one business process, one interruption tolerance and one credible recovery or exit route.

The legal entity matters

The four designations name legal entities, not consumer brands in the abstract. They also bring the systemic services supplied to UK financial firms under oversight, not every product sold by the wider corporate group in every market.

That boundary matters when a procurement file says only “Microsoft”, “AWS” or “Google”. A contract may be with a reseller, a different group entity or a managed-service provider that assembles several underlying services. An AI application may add a model host, vector database, identity service, observability vendor and data processor on top of the designated cloud.

The useful first line in the file is therefore factual:

contracting entity / exact service / region / reseller / material subcontractors

If that line is incomplete, the CTP label cannot repair it. Oversight of one entity’s critical services does not automatically extend to every component in the deployment chain.

The UK framework is designed around system-wide risk. HM Treasury designates a provider where disruption could threaten financial stability or confidence in the financial system. The three regulators then oversee the provider’s resilience for the relevant services. That helps with risks no single small firm can investigate alone. It does not answer whether a specific workload has the right retention settings, recovery design or contractual support.

Oversight adds evidence, not approval

The CTP rules cover governance, risk management, dependency and supply-chain risk, technology and cyber resilience, change management, mapping, incident management and termination of services. They also create channels for self-assessment, scenario testing, incident notification and information sharing.

Those mechanisms should improve the evidence available to firms over time. A provider may have to test its ability to continue a material service under a severe but plausible scenario, exercise an incident playbook with representative financial firms and share information about resilience. Regulators can gather information, commission skilled-person work and enforce the rules.

But the regime was deliberately written not to replace firm-level controls. The 2024 joint policy statement says firms remain accountable for managing risks in arrangements with a designated provider. HM Treasury repeated the point when announcing the first four names.

The distinction can be kept simple:

Evidence from CTP oversight	Evidence the buying firm still needs
System-level resilience of designated critical services	Fitness of the exact service and configuration for the use case
Provider incident communication under the regime	The firm’s own escalation, customer communication and workaround
Provider mapping and scenario testing	Mapping from the service to the firm’s important business process
Provider termination and recovery capabilities	Data portability, substitute service and an executable exit decision
Regulatory scrutiny of the designated entity	Due diligence on resellers, models, integrators and other subcontractors

Source: HM Treasury’s July 2026 designations and the Bank/PRA/FCA CTP policy statement. Last verified 2026-07-23.

The left column can reduce duplicated investigation. It cannot become a tick that erases the right column.

AI makes the dependency wider

The first CTPs are cloud and technology providers, not AI-model companies as a category. Yet the designation arrives while financial firms are moving more AI workloads onto the same concentrated infrastructure.

The Bank and FCA’s 2024 survey found that 75% of responding financial firms were already using AI. One third of reported AI use cases were third-party implementations. The three most-named providers accounted for 73% of cloud providers and 44% of model providers in the survey. Nearly half of respondents said they had only a partial understanding of the AI technologies they used.

Those figures describe a dependency chain, not merely a vendor market. A hosted model can rely on one cloud for inference, another service for identity, a third-party data source and an internal process that assumes all of them are available. A provider can be resilient while the assembled workflow is fragile.

The survey also found that 55% of AI use cases had some degree of automated decision-making, although only 2% were fully autonomous. Materiality therefore depends on the task, data and decision path as much as on the model name. A summarisation assistant and an automated fraud hold may use the same infrastructure but require different evidence, fallbacks and human authority.

This is why the government’s 14 July Financial Services AI Adoption Plan points in two directions at once. It recommends accelerating CTP oversight, including assessment of key AI and cloud providers, while also proposing a voluntary third-party AI assurance scheme. Shared assurance could reduce repeated questionnaires. The plan still says each firm would remain responsible for the risks of implementing the model.

Keep a one-page dependency file

A smaller regulated firm does not need to reproduce the regulators’ system-level work. It needs a short file that makes its own dependency and response visible.

For every material cloud-backed AI service, record:

1. The service. Name the contracting entity, product, region, reseller and model version or family where relevant.
2. The business process. State what stops, degrades or becomes unsafe when the service is unavailable or wrong.
3. The data path. Record what enters the service, where it is retained, which logs exist and how access is revoked.
4. The decision boundary. Identify what the system may do, where a person intervenes and who can disable it.
5. The interruption limit. Set an honest time after which the business impact becomes intolerable.
6. The fallback. Name the manual route, degraded service or substitute provider and the person allowed to activate it.
7. The evidence feed. List the provider notices, CTP information, tests and contract reviews that keep the file current.
8. The exit. Record export format, data deletion, replacement lead time and the last date the route was tested.

This is not a new framework. It is the minimum translation between a provider’s resilience and the firm’s service obligation. Existing outsourcing, operational-resilience, security and data-protection records should supply most of it.

Test the gap the designation cannot close

Choose one material service on a designated provider and simulate a two-hour loss. Do not test whether the provider has a status page. Test the firm’s side of the boundary.

Can the team identify the affected customers and decisions? Can it stop automated actions without losing an audit trail? Does the fallback work with current data? Who is authorised to invoke the contingency? Which reseller or integrator joins the incident call? What evidence will show that service and data integrity have been restored?

Then simulate a quieter failure: the cloud is available, but the model endpoint, identity link or data connector is not. CTP oversight is aimed at critical services and system-wide resilience. Many costly operational failures will still occur inside a firm’s particular assembly of non-critical components.

The first designations are useful because concentrated providers now face direct, coordinated scrutiny that individual firms could not create. Treat the resulting evidence as an input. Keep the supplier decision, configuration risk, fallback and exit route in the firm’s own hands.