AI agents are real and useful in 2026, but only for narrow, well-defined tasks. They reliably triage support tickets, extract data from documents and draft follow-ups when the goal is clear and the data is structured. They do not yet run whole departments on their own. The safe first build is one bounded process, with a human approving irreversible actions.
What is an AI agent, precisely?
An AI agent is software that uses a language model to decide which steps to take and which tools to call in order to reach a goal, rather than following a fixed script. Anthropic draws the distinction clearly: workflows are systems where models and tools are "orchestrated through predefined code paths", while agents are systems where the model directs its own process and tool usage, maintaining control over how they accomplish tasks.
IMDA describes the same shift from a governance angle. Compared with generative AI, its framework says, AI agents "can take actions, adapt to new information, and interact with other agents and systems to complete tasks on behalf of humans". In a business context that could mean an agent that reads an incoming purchase order, checks it against inventory, drafts a sales order in the ERP and notifies the warehouse, instead of a person clicking through each screen.
| Approach | Who decides the steps | Best fit |
|---|---|---|
| Chatbot or assistant | The user, one prompt at a time | Answering questions, drafting text |
| Workflow (LLM inside fixed code) | The developer, in advance | Well-defined tasks where predictability matters |
| Agent | The model, at run time, using tools | Open-ended tasks needing flexibility and judgement |
Source: distinction between workflows and agents from Anthropic, Building effective agents.
The practical consequence is that many tasks marketed as "agentic" are better built as workflows. Anthropic's own guidance recommends "finding the simplest solution possible, and only increasing complexity when needed", noting that this "might mean not building agentic systems at all". IMDA makes the same point: "not all use cases are suitable for agents, and some may be better served by deterministic workflows".
Where AI agents deliver value today
AI agents deliver the most value on mundane, high-volume tasks that combine a conversation or document with a defined action and a clear way to check the result. IMDA notes that agents are already in use through "coding assistants, customer service agents, and automating enterprise productivity workflows". Anthropic singles out customer support and coding as the two domains where its customers found particular value, because they need both conversation and action, have clear success criteria and allow meaningful human oversight.
| Use case | What the agent does | Why it suits an agent | Human checkpoint |
|---|---|---|---|
| Customer support triage | Classifies, routes and resolves routine tickets using order and account lookups | Conversation plus tool calls; resolution is measurable | Refunds, credits and complaints escalated |
| Document processing | Extracts fields from invoices, contracts and forms into the ERP | Structured output can be validated against rules | Low-confidence or mismatched fields reviewed |
| Sales follow-up | Drafts personalised follow-ups based on CRM activity | Repetitive, with clear inputs | Messages to external parties approved before sending |
| Inventory reorder alerts | Monitors stock levels and proposes purchase orders | Rule-based thresholds with clean data | Purchase orders approved before issue |
Measure each of these against the manual baseline you have today: time per ticket or invoice, error rate and escalation rate. Vendor case-study percentages rarely transfer to a different business, so your own before-and-after numbers are the only ones that justify expansion.
The hype: fully autonomous business operations
The claim that AI agents can run entire departments autonomously is not supported by current evidence. In June 2025 Gartner warned that many vendors were engaging in "agent washing", rebranding existing assistants, robotic process automation and chatbots without substantial agentic capabilities, and estimated that only about 130 of the thousands of agentic AI vendors are real, as reported by w.media. The same report quotes Gartner analyst Anushree Verma: "Many use cases positioned as agentic today don't require agentic implementations."
The reasons are structural. Anthropic notes that agentic systems "often trade latency and cost for better task performance", and that the autonomy of agents means "higher costs, and the potential for compounding errors". Each extra step an agent takes on its own is another chance for a small mistake to carry forward. Agents also struggle with ambiguous situations, judgement calls and context that has never been written down, which is exactly where much departmental work sits.
None of this means agents are a fad. The same Gartner release, as reported, still describes agentic AI as a leap forward in capability. The gap is between "an agent can handle this task" and "an agent can own this function", and that gap is where most failed projects start.
How Singapore regulates AI agents
Singapore's main guidance for AI agents is the Model AI Governance Framework for Agentic AI from the Infocomm Media Development Authority (IMDA), a best-practice framework that IMDA updated on 20 May 2026. The framework targets organisations deploying agents, "whether by developing AI agents in-house or using third-party agentic solutions", and IMDA calls it "a living document" that will keep being updated. If an agent will handle personal data, check your PDPA obligations before the build, not after. For the wider regulatory picture, see our overview of Singapore AI regulation in 2026.
The framework sets out four areas: assess and bound the risks upfront; make humans meaningfully accountable; implement technical controls and processes; and enable end-user responsibility. Its risk factors are a useful checklist before any build.
| IMDA risk factor | Lower risk | Higher risk |
|---|---|---|
| Access to sensitive data | Public information only | Personal customer data, especially with persistent memory |
| Scope of actions | Read-only, a few predefined tools | Can write to systems, or use a computer-use tool on any interface |
| Reversibility | Scheduling a meeting | Sending emails to external parties, entering a contract or sale |
| Level of autonomy | Follows a detailed SOP | Uses its own judgement to select and execute every step |
| External systems | Internal, trusted knowledge base | Open web with untrusted data (prompt injection exposure) |
Source: factors and illustrations summarised from the IMDA Model AI Governance Framework for Agentic AI, Version 1.5.
Choosing the right level of human involvement
IMDA describes four levels of human involvement, from "agent proposes, human operates" (a person approves every action) through "agent and human collaborate" and "agent operates, human approves" to "agent operates, human observes" (actions audited after the fact). For a first deployment, the middle two levels are usually right. The framework recommends human approval before high-stakes and irreversible actions, giving examples such as "permanently deleting data, sending communications, making payments", and recommends enforcing approvals through system-level controls rather than prompt instructions alone.
A practical framework for choosing the first agent
The best first agent is a task that is repetitive, rule-based and fed by reliable data. Namtech uses this "Three R" test with clients because it screens out the use cases most likely to stall:
- Repetitive: the task is performed often enough that saving time on each instance matters.
- Rule-based: clear criteria define a correct result, so outputs can be checked.
- Reliable data: inputs are structured and consistent, with a known system of record.
Combine it with IMDA's risk factors and a simple scoring sheet emerges.
| Question | Good first candidate | Defer for now |
|---|---|---|
| How often does the task run? | Daily or more, many instances | A few times a month |
| Can a correct result be checked? | Yes, against rules or a source document | Only by expert judgement |
| Is the input data structured? | Consistent formats in one system | Scattered emails, spreadsheets and memory |
| Can mistakes be reversed? | Yes, or a human approves first | No, actions are external or binding |
| Does it need personal data? | No, or minimal with clear PDPA basis | Broad access to customer records |
Customer email classification, invoice data extraction and inventory reorder alerts usually score well for SMEs. Open-ended work such as negotiating with suppliers or handling complaints usually does not.
Implementation approach for Singapore businesses
Start with one process, deploy it narrowly, measure it against the manual baseline, then expand. In practice that means five steps:
- Baseline: record current volume, time per task and error rate before building anything.
- Bound: give the agent the fewest tools and the narrowest data access that will work. IMDA recommends the principle of least privilege and advises against granting write access to sensitive databases "unless strictly required".
- Test: run the agent in a sandbox on historical cases. Anthropic recommends "extensive testing in sandboxed environments, along with the appropriate guardrails".
- Roll out gradually: IMDA suggests rolling out to trained users first, with restricted tools, on lower-risk internal systems, alongside continuous monitoring.
- Review and expand: compare results with the baseline, audit the human approvals, and only then add a second process.
If you are unsure which process to start with, an AI Opportunity Audit maps your workflows and ranks candidates in 2–4 weeks. For businesses ready to build, Namtech Pte Ltd, a Singapore AI transformation consultancy, delivers agents within a structured AI transformation programme.
A single agent handling invoice processing well delivers more value than five agents handling various tasks poorly.