AI Glossary Singapore: AI and PDPA Terms for SME Owners

This AI glossary for Singapore SMEs defines 21 terms an owner or manager meets when buying, testing or governing an AI tool, from hallucination to data breach notification, in plain language and in the wording of IMDA, PDPC and MDDI where they have published one.

Each entry gives the definition, the official source and, where one applies, a Methodology guide; where the agencies do not define a word, it says so. Checked on 17 October 2026; general information from desk research, not legal advice. Recheck when IMDA or PDPC revise a document, especially PDPC’s Key Concepts guidelines (revised 29 April 2026).

Which AI terms should a small business learn first?

• Hallucination, which IMDA lists first among five key risks of LLM-based applications.
• Human in the loop and audit log: who checks an AI system’s actions, and what record is kept.
• System deployer, the PDPC label for an organisation using a generative AI system.
• Data residency: where a vendor processes your customers’ data.

Our suggested order, not an official ranking; see also your first AI pilot should survive a bad answer.

How do these terms fit together in practice?

• Fictional example: a ten-person renovation firm adds a website chatbot built on a model.
• The chatbot answers from the firm’s price list using retrieval-augmented generation, but may still hallucinate outside it.
• The firm sets a guardrail on prices, keeps a human in the loop for refunds and promises, and keeps an audit log.
• It publishes a chatbot information card and tells staff which AI tools are approved, to avoid shadow AI.
• Chats contain names and addresses, so they are personal data. As the system deployer, the firm asks its vendor, a data intermediary, about data residency and how it would learn of a data breach.

The scenario is illustrative; it describes no real firm.

Agentic AI

Agentic AI means software systems made up of one or more AI agents, which may work alone or together. IMDA notes there is no consensus on what defines an AI agent, but agents usually have some degree of independent planning, decision-making and action-taking over multiple steps to achieve a goal the user has set.

Source: IMDA’s Model AI Governance Framework for Agentic AI (version 1.5), section 1.1. For a plain explanation and a supplier test, read what is agentic AI.

AI agent

IMDA’s framework focuses on AI agents built on generative AI models, typically language models: software that can plan steps and use tools to act on other systems, such as writing to files and databases, controlling devices or performing transactions.

IMDA’s framework, section 1.1.1, lists the parts as a model, instructions, memory, planning, tools, protocols, controls and logging. See AI agent vs chatbot: decide what the software may do first.

Audit log

An audit log is a record of what an AI system did, decided and interacted with, kept so that its behaviour can be monitored, debugged and held to account. IMDA says ‘logging and monitoring’ and ‘audit trails’ instead, so ‘audit log’ is our shorthand.

IMDA’s framework says organisations should continuously monitor and log agent behaviour after deployment, to intervene in real time, debug incidents and audit regularly (section 2.3.3). See customer enquiries and approval rules.

Automation bias

Automation bias is the tendency to over-trust an automated system, especially when it has performed reliably in the past.

IMDA’s framework calls it a bigger concern as humans supervise more capable agents, and suggests tracking the human override rate, where a low rate may signal rubber-stamping. See checks before an agent acts for you.

Chatbot

In IMDA’s chatbot guidelines, a generative AI chatbot is an application that uses foundation models such as large language models, in whole or in part, to produce adaptive, human-like responses in natural language through an interactive text or speech interface.

IMDA’s Transparency Guidelines for Generative AI Chatbots, announced in an IMDA release of 20 July 2026, cover external-facing chatbots; tools used solely inside an organisation are outside their scope. See IMDA chatbot transparency guidelines: draft your information card.

Chatbot information card

A chatbot information card, which IMDA calls a ‘chatbot info card’, is a single consolidated reference point where users can find what a chatbot can and cannot do, how reliable and safe it is, how their data is handled and how to report issues.

IMDA likens it to a ‘medical label’. The guidelines are voluntary, but IMDA strongly encourages at least one substantive disclosure under each of those four questions, and says the card should be published by the time the chatbot is made available. See our guide to drafting an information card.

Data breach notification

A data breach notification is the report an organisation must make to the PDPC, and where required to affected individuals, when a data breach is notifiable under the PDPA. A breach is notifiable if it is likely to result in significant harm to affected individuals or involves the personal data of 500 or more individuals. Once the organisation determines that a breach is notifiable, it must notify the PDPC as soon as practicable and no later than three calendar days after that determination.

Before that, once it has credible grounds to believe a breach has occurred, the organisation must take reasonable and expeditious steps to assess whether it is notifiable. PDPC’s Key Concepts guidelines (revised 29 April 2026) say it should generally do so within 30 calendar days (paragraph 20.4); the 2021 breach guide qualifies the 30 days with ‘in general’ in Part II but states the duty ‘within 30 calendar days’ in Part III. A data intermediary must notify the organisation that engaged it, which stays responsible for assessing and notifying: PDPC’s breach-reporting page (22 April 2026) says ‘immediately upon discovering the breach’, and the Key Concepts guidelines say without undue delay once it has credible grounds to believe a breach has occurred (paragraphs 20.7 and 20.8). See PDPA data breach notification: the 3-day rule for SMEs.

Data intermediary

Under the PDPA, a data intermediary is an organisation that processes personal data on behalf of another organisation, but not an employee of that other organisation.

PDPC’s Key Concepts guidelines say the organisation has the same obligations for that data as if it processed the data itself (section 4(3)). PDPC’s generative AI guidelines treat a system provider that processes data for a deployer, for example through software as a service, as a data intermediary (paragraph 8.1). See AI vendor questions before you sign.

Data Protection Officer

A Data Protection Officer (DPO) is the person an organisation designates to be responsible for ensuring that it complies with the PDPA. Section 11(3) of the PDPA requires an organisation to designate one or more such individuals, and legal responsibility for complying stays with the organisation, not the DPO.

PDPC’s DPO page (published 1 September 2026) says every organisation in Singapore that handles personal data must appoint at least one DPO and calls appointing one ‘your mandatory first step’. The DPO’s contact details must then be made publicly available, a legal requirement under the PDPA, either by registering them for PDPC’s DPO Registry or by listing them on the organisation’s own website. See Data Protection Officer Singapore: appoint and register a DPO.

Data residency

Data residency means the places where data is stored and processed. PDPC uses the term in its generative AI guidelines without defining it, so this is our plain-language gloss.

PDPC’s generative AI guidelines say it is good practice for model providers and system providers to document and share safeguards such as data access controls and data residency and retention policies (paragraphs 7.5 and 8.3). Our suggestion: ask a vendor to state its data residency in writing. Our inference: an answer outside Singapore may engage the Transfer Limitation Obligation below. See AI vendor questions before you sign.

Guardrail

In IMDA’s framework, a guardrail is a control that monitors and constrains an agent’s behaviour before, during or after it acts, helping to detect unsafe instructions, policy violations or actions that appear inconsistent with user intent.

IMDA lists guardrails with access controls and human approvals as the controls most relevant to agents (section 1.1.1 of its framework).

Hallucination

A hallucination is incorrect or fabricated output from an AI system, whether about universal facts or about a specific source or document such as company policy.

IMDA’s Starter Kit for Testing LLM-Based Applications (version 1.0, January 2026) names the risk ‘hallucination and inaccuracy’ and lists it first among five key risks. Method: AI hallucination: checks before customers see an answer.

Human in the loop

Human in the loop means a person reviews or approves an AI system’s actions at defined checkpoints; IMDA’s framework lists human approvals as a control for agents.

IMDA advises defining checkpoints that need human approval, especially before sensitive actions such as high-stakes decisions and irreversible steps like deleting data, sending communications or making payments (section 2.2.2 of its framework). See approval rules for customer replies.

Model

In IMDA’s agentic AI framework, the model is the small, large or multimodal large language model that serves as an agent’s central reasoning and planning engine, or ‘brain’.

PDPC defines generative AI models as models that display significant generality and can competently perform a wide range of distinct tasks (paragraph 1.3 of its guidelines).

Personal data

The PDPA defines personal data as data, whether true or not, about an individual who can be identified from that data, or from that data and other information the organisation has or is likely to have access to.

PDPC’s Key Concepts guidelines (revised 29 April 2026) say the term is not narrowly construed, and that the Data Protection Provisions do not apply to business contact information. See PDPA and AI: what PDPC’s 2026 guidelines expect of SMEs.

Prompt

A prompt is the text or other input that a user, or the builder of an AI system, gives an AI model to direct what it does. IMDA and PDPC do not define the word in the documents we read, so this is our plain-language gloss.

PDPC says a deployer must safeguard personal data in new data sources such as end-user prompts, inputs and generated outputs (paragraph 9.3 of its guidelines). Our suggestion for staff rules: the AI use policy template for staff.

Prompt injection

In IMDA’s Starter Kit, prompt injection is an attack that uses adversarial instructions to override an AI application’s safety controls. It is direct when an attacker interacts with the application’s interface, and indirect when the instructions are embedded in content the application takes in, such as a web page or an uploaded file.

IMDA’s agentic AI framework warns that, because tools connect an agent to external systems, prompt or code injections can manipulate the agent to exfiltrate or otherwise manipulate the data it can access. The Starter Kit covers testing for it (section 3.6).

Retrieval-augmented generation

Retrieval-augmented generation (RAG) is a technique in which information relevant to the user’s input is retrieved from a knowledge base, such as a company’s FAQs, and supplied to the model so that its answer is based on that information.

IMDA’s Starter Kit says many app developers use it to address hallucination and inaccuracy, and names two points of failure to test: whether the right information was retrieved and whether the output was grounded in it (section 3.2.3). Method: checking an answer before a customer sees it.

Shadow AI

Shadow AI is the unsanctioned use of AI tools by employees without the approval or oversight of the IT department.

Senior Minister of State Tan Kiat How gave it as an AI risk, alongside accidental leakage of information and inappropriate output, in a speech on 15 April 2025; a ministerial speech is not a rule. See our AI use policy template for staff.

System deployer

In PDPC’s generative AI guidelines, a system deployer is an organisation that uses and/or enables the use of generative AI systems under its authority.

A deployer may build systems in-house or procure them as a service, for example as software as a service. Either way, PDPC says it bears primary responsibility for ensuring that the systems it chooses can meet its PDPA obligations (paragraph 9.1 of the guidelines, issued 20 July 2026). Our inference: the label fits an SME that buys or switches on a ready-made AI tool. See PDPA and AI: what PDPC’s 2026 guidelines expect of SMEs.

Transfer Limitation Obligation

The Transfer Limitation Obligation (PDPA section 26) says an organisation must not transfer personal data to a country or territory outside Singapore except in accordance with the requirements prescribed under the PDPA, which are meant to ensure the data receives protection comparable to that under the PDPA.

PDPC’s Key Concepts guidelines (revised 29 April 2026) say it applies where an organisation relinquishes possession or direct control of the data, for example by transferring it to a data intermediary or group company, and that the organisation must ensure the overseas recipient protects it to a comparable standard (paragraphs 19.1 and 19.2). PDPC’s page for the guidelines says the 29 April 2026 revision updated its section on conditions for transfer overseas.