Secure Cloud, or Sealed?

AWS, Microsoft and Google offer cloud environments with extensive security, compliance and governance controls. Legal AI products may use those platforms to process client information. A firm assessing such a product should distinguish cloud storage from AI processing, then decide which work may be processed by an external model and which should remain on premises.

1Cloud security is real and relevant

Major providers encrypt data at rest and in transit, offer customer-managed encryption keys, and now sell contractual assurances that a firm's data will not be used to train their models and can be held to zero retention: Amazon added an enforceable zero-data-retention mode for its Bedrock service in 2026, and Microsoft's Azure OpenAI markets "your data is your data." They carry the certifications a buyer looks for: SOC 2, ISO 27001, HIPAA eligibility. Legal AI tools including Harvey and Robin AI use this type of infrastructure. For many firms, a well-managed cloud service can offer stronger operational security than poorly maintained local infrastructure.

2Cloud storage and external AI processing are different decisions

Many firms use cloud services for email, practice management, CRM, file storage and backup. That does not mean the firm has made the same decision about AI. When an external model processes a prompt, confidential text enters the model provider's service boundary.

Lawyers remain responsible for confidentiality and legal professional privilege when vendors process information on the firm's behalf. The relevant question is therefore not whether cloud services are secure. It is whether a particular processing model is appropriate for the information, the retainer and the firm's professional duties.

3Contractual controls and on-premises processing

A secure cloud combines technical controls with contracts, policies, provider operations and shared responsibility. Secure Law uses a different architecture. It runs on premises, so confidential client files can be processed without being sent to a public or third-party AI model. Both approaches can be legitimate. They create different boundaries and suit different work.

4Four questions for external AI processing

Four points help a firm assess whether an external AI service is suitable for confidential client work.

First: where is text processed? Encryption at rest and in transit is important. AI processing also requires the service to make the relevant text available to the model within its protected runtime. The firm should understand where that processing occurs and which controls apply at that point.

Second: which legal regimes can reach the provider? When a third party holds or processes data, lawful demands may be directed to that provider. Notification and challenge rights depend on the law and the circumstances. The US CLOUD Act can require US-headquartered providers to produce data in their control wherever in the world it is stored. This is not a theoretical worry the vendors dispute; it is one they concede. Asked before the French Senate in June 2025 whether he could guarantee that French citizens' data would never be passed to the US government, a senior Microsoft France executive answered, under oath, "No, I cannot guarantee that," and confirmed the company would comply with a properly framed CLOUD Act request. Analysts describe the incident confirms that server location alone does not answer every jurisdictional question.

This is a legal power, not evidence of routine access. AWS reports that it has disclosed no enterprise or government content stored outside the United States since it began publishing figures in 2020, and that it challenges overbroad demands and redirects them to the customer where it can. But the providers' own transparency reporting shows the door is real: Microsoft disclosed data in response to dozens of law-enforcement demands for enterprise customer accounts in the second half of 2024 alone, a portion of which included content. The point is not that this happens routinely; it is that the power exists, the firm cannot see it operate, and the provider concedes it would comply.

Third: shared responsibility requires active governance. Under the cloud's own shared-responsibility model, the provider secures the infrastructure while the firm remains responsible for classifying data, configuring access, and managing keys, so a single misconfiguration is the firm's liability, not the provider's. "We can't see your data" is a statement about policy, not architecture: by default, some providers' abuse-monitoring can route prompts to human reviewers unless a firm applies for an exemption, and the zero-retention settings that make the assurance real may require configuration. The firm should verify settings, permissions, retention, monitoring and contractual terms rather than treating a certification as the complete assessment.

Fourth: map every organisation in the processing chain. A firm adopting a legal-AI product on a hyperscaler is rarely trusting one party. It is trusting the software vendor, and the cloud they run on, and the model provider behind them, each a separate custodian and a separate point at which a legal demand can land. And putting many firms' data in a handful of very large pools makes those pools a prize: in the 2024 Snowflake incident, attackers used stolen customer credentials to reach roughly 165 organisations' cloud environments, a single class of failure cascading across hundreds of tenants at once.

5Why firms may keep confidential AI work on premises

SLIA (Secure Law Internet Appliance), branded as Secure Law, runs private AI on premises. Confidential client files can therefore be used for research, document review and drafting without being sent to a public or third-party AI model. Access and governance remain the firm's responsibility. The material stays within an environment the firm controls.

That is also the cleanest way to meet the duties that bind law firms in Malaysia and Singapore. Malaysia's amended Personal Data Protection Act now applies a risk-based regime to cross-border transfers and places direct, penalty-backed obligations on data processors; keeping privileged matter data on premises can reduce the need for cross-border AI processing. Singapore's PDPA carries an equivalent transfer-limitation duty, and its Ministry of Law's guidance for the legal sector tells practitioners to safeguard client confidentiality, to understand how each tool processes and stores data, to remain personally accountable for the output, and to consider telling clients where a tool's data-handling may not align with their requirements. On-premises private AI can support those duties by keeping AI processing under the firm's control.

6Choose the processing model for the work

A well-governed cloud can be the right home for much of what a firm does: administrative and back-office data, public-record and legal research, marketing and CRM, document backup, and non-privileged, lower-sensitivity matters. For a firm without deep IT, it is genuinely safer than self-hosting badly. The two approaches are complementary, not a contest.