Who Answers When Singapore's AI Agents Make Mistakes? The Accountability Question

Singapore's push to put AI agents into real business workflows is running into a question that no amount of model fine-tuning can solve: when an agent acts, who answers for what it does? It's a question that sounds abstract until an automated invoice approval pays the wrong vendor, or an AI procurement bot quietly extends its own权限 beyond what anyone authorised. The difference between a chatbot and an AI agent — between suggesting and doing — is precisely the difference that makes accountability hard to assign.

The Edge Singapore's Karl Crowther explores this boundary in a piece that cuts to the heart of what Singapore's National AI Missions will face as they scale agentic AI into priority sectors. A chatbot answers a question; an agent acts on it — approving an invoice, updating a customer record, or moving funds between accounts. The analogy Crowther uses is cruise control: it holds speed better than your foot, and will hold it straight into a stopped vehicle, because judging the road was never its job. An AI agent works the same way, doing exactly what it's asked at speed, with no instinct for when something has gone wrong.

Singtel recently redrew its organisation chart to reflect this new reality, scoping managers to lead teams of both people and agents. But The Edge piece argues the chart is the easy part. The harder question is accountability — when an agent makes a costly error, the manager who deployed it is responsible, yet in most organisations that manager cannot reliably see what the agent actually did. That blind spot is where risk compounds. Left unwatched, an AI agent can drift beyond its assigned job, slip into other processes, reach into data it was never meant to touch, and act on what it finds.

Singapore's regulators have already put this front and centre. IMDA's Model AI Governance Framework for Agentic AI makes keeping humans "meaningfully accountable" one of its four pillars, warning that an agent's autonomy can blur lines of responsibility tied to fixed workflows. The piece also highlights a quieter but equally significant risk: AI sprawl. Because agents are cheap and simple to create, teams will build their own, each solving a local problem, until the business has an uncontrolled spread of autonomous tools with no central oversight. GovTech's answer — building a registry that tracks who owns each agent and what it does before use becomes widespread — offers a model for the private sector too.

Why it matters for Singapore: As Singapore positions itself as a global hub for agentic AI through the National AI Missions, the infrastructure for accountability needs to be in place before agents are scaled, not after a regulator or customer starts asking questions. The Edge piece frames the sequence clearly: orchestration layer first — a single environment where people, robots, and agents work together, with every action logged and humans staying in the loop where it matters. Firms that rush the agents and defer the oversight will spend years managing avoidable problems. Those that build it first can grow their digital workforce with confidence. The most resilient organisations will settle on a clear division of labour: AI agents think, robots execute, and people lead — and answer for the results.