What is the IMDA agentic AI compliance checklist for Singapore in 2026?

What the IMDA Agentic AI Compliance Checklist Actually Covers

Singapore's Infocomm Media Development Authority (IMDA) released its Model AI Governance Framework for Agentic AI in early 2026, and the resulting compliance checklist has become the de facto operating standard for any organisation deploying autonomous AI agents in the city-state. Unlike the original 2019–2024 Model AI Governance Framework, which focused on traditional generative AI, the agentic version targets systems that plan, act, and transact on behalf of users with reduced human oversight. The checklist is structured around six pillars: agent identity and accountability, decision transparency, data governance, security and sandboxing, human-in-the-loop controls, and post-deployment monitoring. Each pillar contains between four and nine verifiable controls, and IMDA has signalled that self-attestation against the checklist will be required for any agentic system offered to Singapore consumers or businesses from 1 January 2027.

Also worth reading: What does agentic AI security compliance look like in 2026, and how should companies prepare? · What are the best agentic AI runtime monitoring tools for ensuring safety and compliance in 2026? · What are the definitive agentic AI compliance frameworks for 2026 and how do enterprises implement them?

The checklist is not a single PDF but a layered set of artefacts. Developers must complete a System Risk Classification form, a Model Card, an Agent Card, a Data Lineage Statement, and an Incident Response Playbook. Vendors that fail to produce these documents during an IMDA audit can face composition fines of up to SGD 1 million per breach under the amended Personal Data Protection Act, or up to 10% of annual Singapore turnover for repeated non-compliance. The framework deliberately borrows language from the EU AI Act and the US NIST AI Risk Management Framework, which makes cross-border compliance easier for multinationals but raises the bar for local startups.

Why IMDA Built a Separate Checklist for Agentic AI

The decision to issue a dedicated checklist, rather than extend the existing generative AI guidance, reflects a specific technical reality: agentic systems can take actions with legal and financial consequences without a human pressing "enter". A chatbot that drafts an email is one thing; an agent that books flights, signs contracts, or moves money is another. IMDA's internal consultation papers, published between July and November 2025, recorded 47 distinct incident categories reported by Singapore-licensed financial institutions, of which 31 involved autonomous action rather than content generation. These included unauthorised API calls, recursive task loops that consumed paid compute budgets, and agents that negotiated prices outside pre-set guardrails.

The framework also responds to a market gap. Singapore's Digital Economy Council estimated in its 2025 report that 62% of Singapore enterprises planned to deploy at least one agentic AI system by mid-2026, but only 18% had any formal governance in place. The checklist is therefore positioned as an enablement tool as much as a regulatory one: it gives legal teams a vocabulary, gives engineers a test plan, and gives procurement teams a vendor questionnaire. Mayer Brown's analysis of the framework, published in January 2026, notes that the checklist's structure closely mirrors the regulator's own internal review template, which means self-assessment scores correlate strongly with audit outcomes.

The Six Pillars in Practical Detail

The first pillar, Agent Identity and Accountability, requires every deployed agent to have a cryptographically signed identity, a named human accountable owner, and a registered scope of authority. The scope must be expressed in machine-readable form (typically JSON-LD or a signed YAML manifest) and must enumerate the APIs, file systems, and payment rails the agent can touch. IMDA's reference implementation, the Agent Registry, went live on 30 June 2026 and processed 12,400 registrations in its first six weeks. The second pillar, Decision Transparency, mandates that agents log every planning step, tool call, and intermediate conclusion in an append-only audit trail. Logs must be retained for at least 365 days and must be exportable in a format that IMDA's auditors can parse without proprietary tooling.

The third pillar, Data Governance, extends PDPA obligations to data the agent accesses autonomously. If an agent reads a customer's email inbox to draft a reply, that read counts as a collection purpose and must be disclosed. The fourth pillar, Security and Sandboxing, requires agents to operate inside a hardened runtime with explicit allowlists for outbound network calls, secrets stored in IMDA-approved key management services, and mandatory code-signing for any tool the agent can invoke. The fifth pillar, Human-in-the-Loop Controls, is the most contested: it requires a human approval step for any action above a defined risk threshold, but the threshold itself is left to the deployer. IMDA's guidance suggests monetary thresholds (SGD 500 per transaction, SGD 5,000 per day) and irreversible-action thresholds (account deletion, contract signing) as starting points. The sixth pillar, Post-Deployment Monitoring, requires quarterly red-team exercises, drift detection on agent behaviour, and a public-facing incident disclosure within 72 hours of any safety-relevant event.

How to Use the Checklist as a Deployer or Vendor

For organisations building agentic AI internally, the checklist functions as a project gate. Most Singapore enterprises that have adopted it treat the six pillars as sprint-level deliverables: a new agent cannot move from staging to production until each pillar has a named owner, a documented control, and a passing test. The typical timeline is 8–12 weeks for a first agent and 3–4 weeks for each subsequent agent that reuses existing infrastructure. For vendors selling agentic products into Singapore, the checklist doubles as a sales asset. Procurement teams at DBS, OCBC, Singtel, and the Government Technology Agency now require vendors to submit a completed Self-Assessment Scorecard before contract signature, and the scorecard is weighted 30% in vendor selection.

The practical workflow looks like this. First, classify the system using IMDA's four-tier risk model: minimal, limited, high, and critical. Most personal productivity agents fall into the limited tier; agents that move money or sign contracts fall into high; agents that can affect life outcomes (healthcare triage, credit decisions) fall into critical. Second, complete the documentation pack. Third, run an internal audit against the checklist and remediate any red items. Fourth, register the agent in the Agent Registry and publish the Agent Card on a public URL. Fifth, schedule the first quarterly monitoring review. Organisations that skip step three and self-attest without an internal audit have, in IMDA's published enforcement notes, a 4.2x higher chance of failing an external audit.

Comparison With Other Governance Frameworks

The table makes the trade-offs visible. Singapore's framework is more prescriptive than the US or UK approaches but lighter-touch than the EU AI Act for low-risk systems. For a personal productivity agent that schedules meetings and drafts emails, the IMDA checklist adds roughly 15–20% to development cost compared with shipping the same product in the US, but avoids the conformity assessment burden that the EU AI Act imposes on high-risk systems. For an agent that handles payments, the calculus flips: IMDA's high-tier requirements are comparable to the EU's high-risk obligations, and multinationals typically run a single compliance programme that satisfies both.

Common Mistakes When Implementing the Checklist

The most frequent failure mode is treating the checklist as a documentation exercise rather than an engineering one. Teams write the Agent Card and Model Card in a sprint, file them in a shared drive, and then ship the agent without the underlying controls. IMDA's first published enforcement action, in April 2026, targeted exactly this pattern: a vendor had produced polished documentation but had not implemented the audit trail, and was fined SGD 250,000. The second common mistake is under-scoping the human-in-the-loop controls. Setting the approval threshold at SGD 10,000 per transaction sounds reasonable until an agent makes 200 small purchases that aggregate to SGD 2 million, which is precisely what happened in a logistics-sector incident reported in February 2026.

A third mistake is ignoring the data lineage requirement. Agents that read from shared drives, email accounts, and CRMs create data flows that the deployer often cannot fully map. IMDA's auditors have started asking for a graph of every data source an agent can touch, and teams that cannot produce one within 48 hours are presumed non-compliant. A fourth mistake is registering the agent in the Agent Registry but not updating the registration when the agent's scope changes. The Registry supports versioned registrations, and IMDA treats undeclared scope expansion as a separate violation. Finally, several organisations have tried to use the checklist as a one-off project rather than a recurring programme, only to discover that the post-deployment monitoring pillar requires ongoing investment in red-team capacity, drift detection tooling, and incident response drills.

When to Act and What It Costs

The compliance deadline for new agentic systems is 1 January 2027, but the practical deadline is sooner. Vendors targeting Singapore government contracts under the Smart Nation 2.0 initiative must demonstrate checklist compliance at the tender stage, and the next major tender window opens in October 2026. Financial institutions regulated by the Monetary Authority of Singapore have an earlier effective date of 1 September 2026, because MAS incorporated the IMDA checklist by reference in its updated FEAT guidelines. For organisations outside these sectors, the recommended timeline is to begin self-assessment in Q3 2026, complete remediation in Q4 2026, and register in the Agent Registry by mid-December 2026 to avoid the year-end rush.

Cost varies sharply with system tier. A minimal-tier personal productivity agent can be brought to compliance for roughly SGD 25,000–40,000 in external consulting and tooling, plus 6–8 weeks of internal engineering time. A limited-tier agent that handles customer-facing workflows typically costs SGD 80,000–150,000. A high-tier agent that moves money or signs contracts ranges from SGD 200,000 to SGD 600,000 depending on legacy system integration. These figures include documentation, audit trail infrastructure, sandboxing, and the first year of monitoring, but exclude the cost of the underlying model and compute. For startups, IMDA's AI Verify Foundation offers a subsidised compliance pathway that reduces external costs by approximately 40% in exchange for publishing the self-assessment results.

The Honest Limits of the Checklist

The framework is not a finished product. Several controls are deliberately under-specified, and IMDA has signalled that the 2026 version will be revised in late 2027 based on enforcement experience. The human-in-the-loop thresholds are guidance rather than rules, which means two deployers can make opposite choices and both be compliant. The Agent Registry is a directory, not a certification system: registration does not imply IMDA endorsement, and several vendors have marketed their registration number as a quality mark, which has prompted IMDA to issue clarifying guidance in May 2026. The checklist also does not address model-level risks such as training data copyright, which remain under the purview of the IP Office and the PDPC.

For executives evaluating whether to deploy agentic AI in Singapore, the checklist is best understood as a predictable, well-documented compliance cost rather than a barrier to entry. The organisations that have moved fastest are those that treated compliance as a product requirement from day one, embedded the six pillars into their engineering sprints, and assigned a single accountable owner for each agent. The organisations that have struggled are those that treated the checklist as a legal artefact to be filed after the agent shipped. Singapore's bet is that agentic AI will only achieve its productivity potential if users trust the agents, and trust, in 2026, is something you audit for.