Singapore’s AI moment is not a doomsday script; it’s a governance stress test. The headlines about runaway models and existential risk are dramatic, yes — but the urgent, practical problem for local businesses is much closer to home: opaque models being deployed with little to no verification, untethered agentic systems performing actions they were never cleared to do, and vendors who hide training data behind competitive claims.
Stop panicking. Start auditing.
That sentence sounds blunt because the situation is blunt. The question that needs repeating, and answering, is simple: “Can you tell whether the AI system you are buying has been tested?” If the answer is uncertain, then buying that tool is a risk, and a manageable one — if treated like a controllable engineering problem rather than a headline.
Singapore has taken real, tangible steps. The IMDA’s AI Verify toolkit and the AI Tester Accreditation Programme are not theoretical gestures; they are mechanisms meant to force transparency and measurable checks. The Model AI Governance Framework for Agentic AI and the Monetary Authority of Singapore’s risk toolkit are the practical guardrails that businesses should be using right now. Treat them like mandatory maintenance schedules for an engine that could otherwise overheat.
Concrete controls that matter
- Demand vendor disclosure on testing regimes and failure modes. Not vague assurances; specifics.
- Insist on third‑party audits or accredited testers. Independent verification changes outcomes.
- Limit agent permissions. Every added capability means a larger blast radius for mistakes.
- Monitor outputs continuously and set human-in-the-loop checkpoints for consequential decisions.
- Treat training data as a risk surface: unknown data is unknown liability.
These controls sound bureaucratic. They are. And bureaucracy in this case buys time and prevents catastrophe. Vendors who say transparency will kill innovation are selling convenience over resilience.
Open models are not a panacea — but they help
Open-weight models are gaining traction for a reason. They allow deeper inspection, simpler troubleshooting, and—crucially—cost-effective deployment for SMEs. When models can be examined, the origin of bad outputs becomes discoverable. Proprietary black boxes? Hard to secure. Harder to trust.
“Not knowing the AI’s training data is dangerous too, as you won’t know if there are hidden risks in the data that could cause harm to users later.”
That is not speculation. It is a fact. Hidden biases, toxic sources, or mislabelled data will surface at the worst possible moment — during a live customer interaction, a regulatory audit, or a financial transaction.
Real-world lessons — an anecdote that matters
A late-night call with a small finance firm in Bugis revealed a familiar pattern: an AI agent deployed to automate document retrieval began retrieving and exposing internal credentials because of misconfigured access scopes. Panic followed. Contracts were suspended. Customers were notified. The aftermath was expensive, humiliating, and utterly avoidable.
That firm had bought convenience over control. That single misconfiguration could have been prevented by simple governance: stricter permissioning, proper testing by an accredited party, and an audit trail. The emotional toll was not just on balance sheets; it was on trust. Precious client relationships frayed in hours.
Where industry calls to slow down help — and where they don’t
Calls from major players to slow AI development are dramatic and partly useful. Slowing specific dangerous capabilities — autonomous hacking tools, self-improving systems with no oversight, biological design automation without strict controls — is sensible. But a blanket pause is a blunt instrument that benefits companies that can already absorb development slowdowns and harms those trying to build safer alternatives.
Third-party evaluations and accredited testers are the practical middle ground. Independent assessment cuts through marketing narratives. When vendors loudly claim existential risk, let accredited bodies check the claims. When a model is touted as “safe by design,” demand the evidence.
Practical checklist for Singapore SMEs
- Map every AI agent and model in use. Know what decisions they can make, and what data they can access.
- Require proof of third-party testing before procurement. No exceptions for convenience.
- Adopt the IMDA and MAS guidelines as minimum standards. Use CSA’s guidelines for technical security controls.
- Run periodic red-team exercises. Simulate abuse, misconfiguration, and data leakage scenarios.
- Restrict the number of tools an agent can call. Simpler is safer.
Final word: pace the race, don’t abandon it
Pacing is not a moral failing. It is strategy. Embedding transparency, third-party audits, and strict governance early costs time and money now but preserves reputation, customers, and operational continuity later. Singapore’s approach — verification toolkits, accreditation, and frameworks — should be embraced by SMEs as competitive advantage, not red tape.
Fear sells headlines. Governance saves businesses. Choose the latter. Demand evidence. Insist on accredited testing. Limit agentic power. Do not let the noise about far-future doomsday scenarios drown out the immediate task: governing tools already at work inside organisations, protecting customers, and ensuring that innovation proceeds without sacrificing control.