Guardrails Are Not Enough: AI Agents Also Need Alignment
Guardrails can tell an AI agent what it must not do. But that is not the same as helping it decide what it should do.
This was one of my reflections after attending the “Agents Governing Agents” session at IAPP Asia Forum 2026 in Singapore.
The session used a useful analogy from self-driving cars.
Containment is the brakes.
It establishes boundaries, guardrails, and kill switches.
Alignment is yielding.
It requires judgment, context, and an understanding of how to act appropriately in a particular situation.
That distinction matters for agentic AI.
Many governance controls are designed around prohibition:
Do not disclose certain data.
Do not exceed defined authority.
Do not take prohibited actions.
Those controls are essential. But autonomous systems will also encounter situations where several actions are technically permitted, while only some are consistent with the organization’s policies, risk appetite, or ethical expectations — such as an AI agent choosing how to respond to a vulnerable customer, where multiple options are allowed but only some are appropriate.
That is where the harder governance problem begins.
The session described governance logic as drawing from multiple sources, including privacy and AI laws, compliance requirements, contractual obligations, ethics policies, and cultural norms.
My takeaway: the challenge is not only to build guardrails around AI. It is to translate organizational judgment into decision logic that AI systems can apply consistently.
For legal, compliance, risk, and business teams, that raises a different set of questions:
What values should guide the system when the rules do not provide a single obvious answer?
Which decisions can be automated?
Which situations require escalation?
And who has authority to define the judgment the system is expected to follow?
As AI becomes more autonomous, governance cannot rely only on identifying prohibited actions after the fact. Organizations also need to define how acceptable decisions are made within the boundaries they establish.
Guardrails constrain behavior. Alignment gives direction.
Effective AI governance will increasingly require both.
Inspired by the “Agents Governing Agents” session at IAPP Asia Forum 2026. Views expressed are my own.
