Technology Transactions
SaaS, AI, and Outsourcing Agreements Across Borders: Data, Liability, Suspension, and Exit
A governing law clause does not allocate the main operational risks in a cross border technology agreement. The contract should address permitted data use, responsibility for AI supported outcomes, suspension, and exit.
The first contract risk is structural. The parties should know which document controls, which terms can change, and what happens when a sales promise conflicts with a policy incorporated by hyperlink.
The second risk is operational. Data, users, support teams, cloud infrastructure, subprocessors, and model providers may be spread across countries. The agreement must turn that architecture into clear responsibility.
Treat the agreement as a contract stack
The order form should identify the service, users, term, fees, environments, and purchased features. The statement of work should define deliverables, dependencies, acceptance, staffing, and change control. The service level schedule should state measurable availability and support commitments, exclusions, calculation methods, credits, and whether a credit is an exclusive remedy.
The legal schedules should then follow the service rather than repeat generic labels. At minimum, the stack should establish:
- a clear order of precedence;
- controlled updates to incorporated online terms;
- notice for material service, security, model, or subprocessor changes;
- a right to object or exit when a material change creates unacceptable risk; and
- consistent definitions across the privacy, security, AI, liability, suspension, and exit provisions.
A broad right to change the service at any time can undermine carefully negotiated protections. The contract should distinguish routine improvements from a material reduction in functionality, security, compatibility, or legal compliance.
Map roles and data before allocating risk
"Customer data" is usually too broad to support serious drafting. The parties should identify at least account and billing data, authorized user data, customer content, prompts, outputs, telemetry, feedback, derived data, security logs, and any data used for training or evaluation.
For each category, the agreement should state:
- Who determines the purpose and means of processing?
- What uses are permitted, required, or prohibited?
- Where will the data be stored, accessed, and supported?
- Which affiliates and subprocessors may receive it?
- What retention, return, and deletion rules apply?
- Who responds to individuals, regulators, and incidents?
Where Philippine privacy law applies, the same provider may act in different roles for different processing activities. A provider that processes data only on the customer's documented instructions may have different obligations from one that determines its own purpose or extent of processing. Contract labels do not determine the statutory role.
A controller remains accountable for personal data under its control or custody when processing is outsourced or transferred. Where a processor is used, the arrangement should contain binding terms reflecting the actual processing and applicable requirements.
Do not assume that contractual wording or consent language by itself makes an international transfer compliant. Determine the applicable basis and safeguards for the actual data, roles, purpose, and transfer.
Make AI terms operational
An AI clause should do more than authorize "AI features." It should identify the specific function, intended users, prohibited uses, and level of human review.
The parties should address:
- whether prompts, content, outputs, feedback, or telemetry may be used for training or fine tuning, and by whom;
- the source and permitted use of training or retrieval data;
- retention and deletion, including limits on removing information from an already trained model;
- evaluation, logging, model changes, regression testing, and notice;
- human review and restrictions for decisions with material or regulated effects; and
- responsibility for instructions, inputs, configuration, and reliance on outputs.
Where Philippine privacy law applies, its requirements continue to govern personal data used in developing, deploying, training, or testing an AI system. Public availability by itself does not remove personal data from those requirements.
Ownership language needs equal care. A promise that the customer "owns" output does not establish that the output is copyrightable or free from third party claims. The agreement should separately address rights in inputs, configurations, feedback, fine tuned models, output, and provider technology, as well as infringement protection and usage restrictions.
Connect security, incidents, and continuity
"Industry standard security" is not a control framework. The security schedule should identify measures appropriate to the service and data, such as access control, multifactor authentication, encryption, logging, vulnerability management, secure development, backups, restoration testing, network separation, personnel controls, and subprocessor oversight.
The contract should also set a practical incident process:
- Define the security events that require notice, including events that may not yet meet a statutory breach threshold.
- Require provider notice early enough for the customer to meet its own legal duties.
- Provide rolling facts rather than waiting for a final investigation.
- Preserve evidence, logs, and chain of custody.
- Allocate containment, restoration, forensic, notification, and regulatory cooperation duties.
- Control public statements without preventing legally required disclosure.
Applicable breach laws may use different definitions and short notification periods. The processor contract should require prompt notice, continuing updates, preservation of relevant evidence, and sufficient assistance for the controller to assess and meet each applicable obligation.
Allocate liability according to control
Responsibility should follow actual control. The provider may control platform security, subprocessor selection, restoration, training initiated by the provider, and aspects of technology provenance. The customer may control instructions, permissions, users, inputs, and prohibited uses. Shared decisions should have shared procedures.
The agreement should distinguish direct losses, third party indemnities, regulatory cooperation, service credits, and insurance. It should state the cap period and calculation base, and consider separate caps or exclusions where the risk justifies them. The appropriate cap structure depends on the transaction, risk allocation, and applicable law.
Definitions matter. Restoration, forensics, legally required notices, data recreation, replacement services, and transition support may be central operational costs. If the contract does not say how they are treated, a liability exclusion may decide the issue unexpectedly.
Make suspension proportionate
Suspension rights should distinguish security threats, unlawful use, nonpayment, emergency maintenance, and provider convenience. Each category should have an appropriate trigger.
Where circumstances permit, the clause should require notice and cure. An emergency suspension should be limited to the affected account, feature, data flow, or user where feasible. The agreement should state restoration criteria, escalation contacts, review frequency, and what evidence the provider will share.
The customer should preserve safe access to its data. Read only access or an export route may be possible even when normal processing must stop. Without continuity and retrieval safeguards, a commercial disagreement can become operational lock in.
Design exit at entry
Exit is a service, not a final sentence. The contract should specify:
- export formats, metadata, configurations, logs, embeddings, and documentation;
- timing, secure transfer, transition assistance, replacement provider cooperation, and fees;
- continued service during any agreed transition period;
- transfer or revocation of credentials, domains, keys, and integrations;
- deletion from active systems, backups, and subprocessors, subject to legal holds; and
- the effect of termination on licenses, confidentiality, security, audit, and payment.
If customer data may enter model training, ordinary account deletion may not remove it from an already trained model. The contract should address that technical limitation before use, but disclosure or agreement does not displace applicable data subject rights. NPC Advisory No. 2024-04 calls for effective mechanisms or alternative measures for objection, rectification, erasure, or blocking, and states that incorporation into a data set does not automatically make the exercise of those rights unreasonable.
Close the legal and operational loop
The governing law and dispute provisions should fit the data, service, and asset map. A governing law clause does not displace mandatory rules linked to data subjects, market activity, sector, or regulated use.
Electronic execution should preserve evidence of identity, authority, intent, the final agreed terms, and the audit trail. It does not cure a lack of authority or any formality required by applicable law.
The useful question is not whether the agreement looks complete. It is whether each party can understand the service, manage a serious incident, continue operating during a problem, and recover its data when the relationship ends.
Selected Authorities
Related Expertise