AI vendor data processing: a Singapore due-diligence file

AI vendor data processing needs a factual record before employees upload customer files, staff information or confidential business material. A polished demonstration does not answer where data goes, how long it remains, whether it trains a model or who can act on the output.

Describe the exact use case

Name the task, user group, input, output and business decision. Distinguish drafting, retrieval, classification, prediction and autonomous action. State whether the result only assists an employee or can change a record, contact a person or trigger a transaction. The risk changes materially when a system can act rather than merely suggest.

Keep the first scope narrow. Use representative data only after its owner approves the trial. Define the correct output, review method and stop condition. The sixty-day AI trial guide can help management test value, while this due-diligence file focuses on data and accountability.

Map the data journey

List each data category entering the system, including prompts, attachments, metadata, user identity, system logs and feedback. Record where the data is collected, transmitted, stored, processed and deleted. Name the vendor, hosting provider, model provider, integration partner and any subprocessor that can receive it.

Identify personal data, client-confidential information, trade secrets and regulated records. Remove fields that the use case does not need. Use masked or synthetic examples during evaluation where possible. A no-training setting does not answer every question about transient processing, logging, abuse monitoring, backups or support access.

Apply the Singapore accountability framework

The PDPC advisory guidelines on personal data in AI systems discuss the responsibilities of organisations and service providers, including data intermediaries. They also address data quality, provenance, protection, retention and information supplied to individuals.

Determine the company’s role for each processing activity. A supplier may process personal data on behalf of the company for one function and use data for its own purposes in another. Record the lawful business purpose, notification and consent analysis, transfer arrangements, access controls and retention decision. Obtain legal advice for material or uncertain processing.

Ask contract questions that match the system

Require the contract and service description to cover permitted purpose, data categories, locations, subprocessors, security controls, incident notice, retention, deletion, export, audit information and termination. Record whether customer content is used for model training, service improvement or human review and which setting controls that use.

Check whether the supplier can change model, hosting region or subprocessors without notice. Define what happens to derived embeddings, indexes, fine-tuned artefacts and logs when the account closes. A promise to delete uploaded files may not include every derived record. Ask for evidence appropriate to the sensitivity and scale of the use case.

Test provenance and output limits

Record where the model retrieves facts and whether the user can inspect citations or source documents. For internal retrieval, verify document permissions before indexing and again when content changes. Prevent a user from receiving a passage that the source system would not allow that user to open.

Test hallucination, stale information, prompt injection and conflicting instructions using controlled cases. Decide which errors create only inconvenience and which could harm a person, client or transaction. The company should state where the output must not be used, including final legal, medical, credit, employment or compliance decisions without the appropriate qualified review.

Control agentic actions

The updated IMDA model AI governance framework for agentic AI highlights meaningful human accountability, bounded permissions, transparency and lifecycle controls for systems that can plan and act. Translate these ideas into the actual tools and accounts available to the agent.

Use least privilege, transaction limits, approved destinations and confirmation before consequential actions. Separate test and production credentials. Log tool calls and resulting changes. Provide a reliable pause and revocation route. An employee review after an irreversible action is not the same as approval before the action.

Run a security and operations review

Ask security owners to examine authentication, administrator access, encryption, tenant separation, vulnerability management, backup, monitoring and incident response. Confirm how employees report an incorrect output or suspected disclosure. Include the AI service in vendor, access and business-continuity reviews rather than treating it as an informal web tool.

Train users on the approved purpose and prohibited data. Use visible labels for experimental workflows. Review access when people change roles. Retain enough evidence to investigate an event without storing prompts indefinitely by default. Balance logging needs with privacy, confidentiality and retention obligations.

Make an adopt, redesign or stop decision

Summarise business value, affected people, data exposure, contract position, security evidence, output tests and residual risk. Name the accountable owner and review date. Adoption can be conditional on a smaller dataset, no external action, an enterprise plan or a contractual amendment. Redesign is often better than accepting a broad default deployment.

The IMDA artificial intelligence resource page links Singapore governance frameworks and testing resources. Use them as continuing references because models and supplier terms change. Recheck the file when the model, integration, data or business purpose changes.

An AI purchase becomes governable when the company can trace every data category, explain every permitted action and identify the person responsible for the outcome. If those answers remain unclear, the safe business decision is to narrow the use case before adding real data.

Maintain a change register

AI services can change models, settings, connected tools and terms faster than ordinary annual vendor reviews. Keep a short register of material changes and the date each was assessed. Trigger a fresh review when the supplier introduces training on customer content, adds a subprocessor, changes hosting location or enables a new action.

Test existing permissions after each integration change. An approved read-only pilot can become a write-enabled workflow when a connector is updated. Record the owner who may approve the added capability and the evidence required. This keeps adoption aligned with the reviewed use case instead of allowing convenience settings to expand the system’s authority quietly.