What is agentic AI? A plain guide for Singapore SME owners

What is agentic AI? It is software that works towards a goal over several steps, deciding for itself what to do next and using tools to act, such as searching the web, creating files or updating records. Singapore’s Infocomm Media Development Authority (IMDA) puts it carefully: there is no consensus on what defines an AI agent, but agents usually have some independent planning, decision-making and action-taking over multiple steps to achieve a goal the user has set.

For a small-business owner the practical meaning is this (our plain-language framing, not IMDA’s): a chatbot or writing assistant produces words for a person to use, while an agent can go on to do something with them. That is why the label matters when a supplier says a product is “agentic”: it is a claim to test, not a specification.

This guide explains the idea and gives you a four-question test for sorting products. For safeguards, see two companion guides: permission settings for AI software and our checklist drawn from IMDA’s agentic AI framework.

What is agentic AI, in plain terms?

IMDA’s Model AI Governance Framework for Agentic AI, version 1.5, describes agentic AI systems as software made up of one or more AI agents that may work alone or together. It focuses on agents built on generative AI models, and notes that software agents are not a new concept; other types exist, such as those that use deterministic rules.

Compared with generative AI, the framework says, agents can take actions, adapt to new information and interact with other agents and systems to complete tasks on behalf of people.

Our plain-language illustration: you give it an outcome, such as “find three suppliers who can deliver this part by Friday”. It works out the steps, uses tools such as a browser or your email, and carries on.

How is it different from a chatbot, an assistant or an automation rule?

The four labels overlap in marketing, so this is our own shorthand for telling them apart:

• Chatbot: answers questions inside a conversation. It does not go off and do things.
• Generative AI assistant: drafts, summarises or rewrites when asked. It performs one step, and nothing happens until a person acts on the result.
• Automation rule: follows steps you fixed in advance, such as “send a reminder on day 7”. It acts, but it makes no decisions.
• AI agent: takes a goal, chooses its own steps and tools, and may act on systems outside the chat.

The line blurs in real products: a chatbot may read your calendar, and an “agent” may be set to answer questions only. IMDA makes a related point: it says most agentic implementations in enterprise deployments are hybrids that combine deterministic rules with some autonomy. So the labels above are practical shorthand, not IMDA categories.

What is inside an agent?

IMDA lists eight core components of a simple agent:

• Model: the language model that reasons and plans.
• Instructions: the written commands that set its role and limits.
• Memory: information it can store and look up.
• Planning and reasoning: working out the steps a task needs.
• Tools: how it acts, such as writing to files and databases or making transactions.
• Protocols: standard ways to connect to tools and to other agents.
• Controls: access limits, guardrails and human approvals.
• Logging and monitoring: a record of what it did.

Version 1.5 added the last two to the core components. Our suggestion: ask a supplier to show you both.

Four questions to test whether a product is really agentic

Ask these about any product described as an agent. They are our rule of thumb, not an IMDA test, though questions 2 to 4 loosely echo IMDA’s distinction between an agent’s action-space (its tools and permissions) and its autonomy (how far it decides how to act, including the level of human involvement). That mapping is ours.

1. Goal: do you give it an outcome rather than a single question or prompt?
2. Steps: does it work out and run the steps itself?
3. Tools: can it change something outside the chat, such as a record, an email, a file or a payment?
4. Autonomy: does it carry on without someone approving each step?

Three or four yes answers: treat it as an agent and set limits before it touches live systems. Two: look closer, because it may be automation with an AI label. Zero or one: a chatbot or assistant, which still needs checking.

Worked example: five fictional product descriptions

All five descriptions are invented for illustration.

• “Answers customers’ questions from your FAQ on WhatsApp.” No goal, no steps, no tools. Zero yes answers: a chatbot.
• “Sends a reminder email on day 7 and day 14 for every unpaid invoice.” It uses a tool and runs without approval, but the steps are fixed. Two yes answers (tools and autonomy), and the second look shows an automation rule.
• “Paste in a meeting transcript and get the action items.” One step, and a person acts on the result, so none of the four is a clear yes. An assistant.
• “Given a goal like ‘find three quotes for replacing the cold-room compressor’, it searches supplier sites, builds a comparison sheet and emails suppliers with questions after you approve each email.” Goal, steps and tools are yes, and autonomy is partial because each email waits for your approval. Three yes answers and a partial fourth: an agent, with approvals that should stay in place.
• “Runs your hiring inbox: screens CVs, shortlists, books interviews and sends rejections.” Four yes answers. An agent acting on decisions about people; the framework gives hiring decisions as an example of biased or unfair actions.

What can go wrong when software acts for you?

The framework describes five types of risk when agents take actions: erroneous actions, unauthorised actions, biased or unfair actions, data breaches and disruption to connected systems. Its example of an erroneous action is an agent fixing appointments on the wrong date.

It also describes cascading or compounding effects: a mistake in one step can propagate and amplify across later steps. Its supply-chain example is a hallucinated inventory figure that leads to reordering too much or too little stock. Our inference: a small firm should watch for this, because a short chain of steps may have no second pair of eyes. See also our guide to checking AI answers before customers see them.

IMDA’s factsheet on the framework, which carries no revision date, groups its advice into four areas: assess and bound the risks upfront, make humans meaningfully accountable, implement technical controls and processes, and enable end-user responsibility.

Within that, the framework recommends human approval at significant checkpoints, such as high-stakes or irreversible actions. It also says users should keep their foundational skills, and lists their loss as a business-continuity risk: users may no longer know how to perform critical processes manually when agents malfunction or become unavailable. Our suggestion: do not give an agent the one task nobody else in the firm still knows how to do.

Do you need agentic AI at all?

Not necessarily; this is our view, not IMDA’s. If the steps are the same every time, a fixed automation is simpler to test and cheaper to supervise. An agent earns its place when the steps differ case by case and checking its work costs less than the time saved. Our first AI pilot guide shows how to measure that honestly, and onboarding an AI intern shows how to supervise one. Other terms are in our AI and automation glossary.

What to ask a supplier before you switch an agent feature on

These are our questions, not an IMDA checklist.

• Which of the four questions does it answer yes to, shown in a demo rather than described?
• Can you see its controls (access limits, guardrails, human approvals) and its logging in the product, and export the log?
• What can it read, change, send or spend, and can each be limited separately?
• Who approves its actions, who can switch it off, and how will staff report a wrong action?

For a task-by-task permission sheet, see AI agent vs chatbot. For pre-launch safeguards, see the IMDA framework checklist. More in our AI and automation topic.

Common questions