Who answers for your AI agent's mistakes?

A Member of Parliament asked MDDI a fairly plain question on 5 August. Would the Model AI Governance Framework and AI Verify stretch to cover systems that act on their own, not just ones that answer a prompt or draft a paragraph. The written reply was short. It also confirmed something a lot of teams building agentic workflows have been quietly hoping regulators would forget to ask about.

MDDI pointed straight back to the Model Governance Framework for Agentic AI, the one IMDA launched in January, and said the same thing again: organisations deploying agentic AI need clear governance structures, someone named to hold oversight, and real human accountability in the loop. Not a policy sitting in a folder somewhere. A person who can answer for what the agent did.

01Six months, three documents, the same sentence

This is the third time in roughly six months that Singapore's regulators have said a version of this. IMDA released the Agentic AI framework in January, alongside a discussion paper asking who should actually be liable when an autonomous agent causes harm, the developer, the deployer, or whoever signed off on the deployment. In June, PDPC opened its own consultation on how the Personal Data Protection Act applies when generative AI systems handle personal data, closing that consultation on 1 July. Then, this month, MDDI stood up in Parliament and pointed back to all of it.

Three separate documents, one thread running through each of them. Autonomy doesn't loosen who's responsible for a system. It tightens the question. The more a system can do without someone checking each step, the more precisely regulators want to know who's answerable when it gets something wrong.

02Where the gap actually opens up

Most agentic AI pilots we see start small and reasonable. Someone in ops wants an agent that can triage claims, draft an approval, or route a document without a human touching every step of it. Nobody sits down early on and asks who signs off when the agent gets it wrong. Six months later the pilot has scaled past the point where anyone can quickly point to who owns it.

That's the exact gap MDDI is naming. It's not a technical problem you patch with better monitoring. It's a governance decision that has to be made before the agent ships, not after an auditor asks for it. Every autonomous action needs a name attached to it, not just a system log.

Autonomy doesn't dilute who's responsible. It concentrates the question onto one name.

This lands hardest in sectors where the data is already sensitive and the actions already carry weight. Insurance, government records, defence. "The system did it" has never held up as a defence in any of those rooms, and it isn't about to start now. It runs on the same logic as Section 116A of the Evidence Act: courts don't presume a process is authentic on its own. They presume it of a process someone certified and stood behind.

03What we build with this in mind

This is the design principle sitting underneath Trusted ActionVerse's autonomy tiers. Agents run on data that's already been certified and governed, but every level of autonomy carries a human checkpoint built into it from the start, not added after a regulator asks a question in Parliament. If an organisation is running agentic workflows on ungoverned data with no owner named per action, this month's reply is a preview of the first question an auditor, or a court, is going to ask.

The tools to build autonomous agents are already sitting on the shelf. What MDDI keeps asking, in slightly different words each time, is whether anyone's still actually driving.

None of this changes what agentic AI can do. It changes what you need to have answered before you turn one loose. A named owner per action isn't red tape bolted onto the technology. At this point, in Singapore at least, it's the technology's actual design brief.