AI deployment is also a data-handling decision
A team might want to use AI to summarise customer emails, search internal documents, classify applications or draft responses. Those use cases can involve personal data, confidential records or information supplied by customers and employees. The question is not only whether the model can perform the task; it is whether the organisation has a clear basis and controls for processing the information.
Start by writing down the purpose
Describe the business task in plain language. Identify why the information is needed, who will use the output and what decision or action it may support. Avoid collecting additional fields simply because they might be useful for a future AI feature.
Map the data and the parties involved
Identify where data originates, where it is sent, which service providers process it, who can access the results and whether information is retained. A business using a third-party GenAI service should understand the provider’s contractual terms, security information and data-use settings.
Review consent and notifications where relevant
Singapore’s PDPC issued Advisory Guidelines on Use of Personal Data in Generative AI on 20 July 2026. Legal commentary on the final guidance notes that, where consent is required for personal data to be used for large-scale model training or fine-tuning, a broad notice about “product improvement” may not be sufficient; an AI-specific notification may be needed. The exact requirements depend on the data, purpose and applicable PDPA basis. This is an area for the organisation’s privacy lead or legal adviser to assess, not a generic statement to paste into a privacy policy.
Separate model training from ordinary system use
Using an AI feature to process a prompt at runtime is not automatically the same as using personal data to train or fine-tune a general model. The system architecture, provider terms and data flows matter. The project team should understand whether data is stored, retained, used for training, used for monitoring or shared with other service providers.
Use data minimisation and access controls
Only provide the information required for the task. Restrict access by role, avoid putting unnecessary identifiers into prompts, use approved repositories and define how sensitive records are handled. Where feasible, consider masking or de-identifying data before it is sent to a model.
Keep human accountability visible
IMDA’s Model AI Governance Framework for Agentic AI emphasises risk assessment, limits on agent powers, meaningful human checkpoints, lifecycle controls and transparency for users. For a business workflow, that means defining which actions an AI system can take, which actions require approval and how users can question or correct an output.
Test the system before it reaches live data
Use sample or appropriately sanitised data during development. Test inaccurate inputs, prompt injection, incomplete records, unauthorised access, unexpected outputs, failed integrations and escalation paths. A pilot should have a limited scope and a named owner.
Set retention, monitoring and incident processes
Decide what logs are needed, how long they are kept, who reviews them and how errors or suspected data incidents are escalated. Document the system’s dependencies and the process for changing or switching providers.
Ask these questions before go-live
What information does the system receive? Why is each field needed? Which provider processes it? Is it used to train a model? Who can access outputs? What actions can an agent take? When is human approval mandatory? How are requests for correction handled? Who owns the incident response?
This is a governance exercise, not only a technical build
A responsible AI deployment brings business owners, system administrators, privacy or compliance personnel and technology providers into the same conversation. A clear design reduces the chance that teams discover data or accountability issues after a workflow is already in use.
Related service: AI development


Comments
No comments yet. Questions about the guide are welcome.