AI agents in Singapore: what they do, and what it takes to run them

Agent, chatbot, or assistant?

Three words, three different things, and the cost of confusing them is real. A chatbot answers from a script or a document set; it cannot do anything. An assistant like ChatGPT or Claude helps a person do work, in a window that person has open. An agent is given a job and a set of tools and works on it - including while nobody is watching, on a schedule, or because an enquiry arrived. The interesting problems only appear at the third level: what is it allowed to do unsupervised, what does it do when it is unsure, and how do you find out afterwards what it did.

What Singapore businesses actually use them for

Four patterns account for most of what we deploy:

• A front desk on WhatsApp, Telegram, web chat or email that answers enquiries from your own documents and live systems, qualifies leads, books appointments, and escalates anything it should not handle. WhatsApp matters disproportionately here - it is where Singapore customers already are, and it is the channel most likely to be answered at 11pm by nobody.
• Operations work - inbox triage, reply drafts queued for approval, CRM hygiene, data moving between tools that were never integrated, month-end reports, the recurring admin that quietly consumes a senior person's week.
• Market and customer intelligence - competitor tracking, supplier and product research, public mentions and reviews, industry news, tender monitoring, delivered as a briefing to the channel you already read rather than a dashboard you will not open.
• Connected systems - the ones that only work because they are wired into something specific: a live inventory API answering "do you have this in stock, and what does it cost", a tender feed scored against what your firm actually wins.

Worth noting what is absent from that list: nothing here replaces a role. These systems take the repetitive layer off a team, which is why the useful comparison is capacity against headcount rather than agent against employee.

What it takes to run them properly

The build is the easy part. Five things decide whether a deployment is still working in month six:

• An approval boundary you chose deliberately. Inbound answers can be autonomous. Anything leaving the building - a message to a customer, a payment, a public post - should wait for a human until you have reason to widen it. Trust levels let you move that line per action rather than flipping one switch for everything.
• An audit trail you can produce. Not for compliance theatre - for the Tuesday when someone asks why a customer was told something. IMDA's governance framework for agentic AI is explicit that humans remain accountable for what an autonomous system does, and you cannot be accountable for what you cannot reconstruct.
• Grounding in your own documents. An agent answering from your price list, policies and product specs states your facts. An agent answering from general knowledge invents plausible ones. This is a configuration decision, not a model quality issue.
• Spending controls. Agents that run on a schedule can run up a bill. Hard caps agreed in advance, not a surprise invoice.
• Somebody who owns it. The most common failure is organisational: the champion moves on, nobody adjusts the system as the business changes, and it drifts into irrelevance. This is the main argument for a managed engagement over a handover.

The Singapore specifics

Three things are genuinely local rather than generic. PDPA applies to the personal data an agent touches, and the PDPC has published advisory guidelines on using personal data in AI systems - consent, purpose and reasonableness all still apply, and outbound marketing to Singapore numbers additionally has Do Not Call obligations. Language is rarely just English: enquiries arrive in Mandarin, Malay and a register of Singlish that no template anticipates. And funding exists but changed this year - EDGE replaced EDG, PSG and MRA for new applications from 30 September 2026. The detail is in our PDPA guide and grant funding page.

What we deploy

Olano designs and runs these systems end to end. Practically: agents on 15+ messaging channels, 450+ connector tools, 75+ built-in integrations each with a specialist subagent, inbound webhooks so any system can trigger an agent, scheduled and overnight work, a knowledge base grounded in documents you upload, and persistent memory so the system does not start from zero every conversation. Work can be delegated between specialist agents, so a front desk agent that hits a research question hands it to the agent that does research.

Each client runs in an isolated environment - no shared multi-tenant data, encrypted at rest and in transit. Systems are multi-model across 18+ providers, so you decide which model handles which task, how much it may spend, and where data is processed; bring-your-own keys is supported at no extra cost. Self-improvement runs on a cadence you set, with changes snapshotted and reversible. The platform itself is documented at olano.ai.

Where to start

One workflow. Pick the one that bleeds most - usually after-hours enquiries, or the admin consuming your most expensive person's afternoons - and deploy that. A single-workflow deployment is typically live within 1-2 business days of onboarding, quoted as a fixed proposal beforehand. Judge the quality yourself first: message the demo agent on WhatsApp or Telegram before you talk to us.