How to Set Clear Company Rules for AI Use

Singapore’s courts offer companies a useful starting point for AI governance: classify uses by how much responsibility is being handed to a system, rather than relying on a blanket statement that AI is “allowed” or “human-supervised.” The courts’ guidance permits generative AI to help prepare court documents while keeping responsibility for accurate, independently verified material with the people who use it. A company can adapt that principle, but its own rules should be designed for its tasks and risks—not presented as requirements imposed by the courts. 36

Set the boundary by task, not by tool

One AI product may be suitable for a low-risk drafting task and unsuitable for a decision that affects a person. Singapore’s courts have described a traffic-light approach: green for uses such as translation and summarisation that help people understand information without deciding a case, amber where AI needs more caution, and red for uses the courts will not consider. 638

For a company, the colours can become a practical policy structure. The examples below are proposed starting points, not a universal classification: assess the actual task, information used and consequences of an error.

Green: assist with routine work

Allow approved AI tools to help with lower-risk tasks such as formatting, translation, organising information or preparing a first draft. Specify which tools are approved, what information employees may enter and whether the output needs a check before use. Make clear that an AI-generated draft is not, by itself, an authorised company decision or position.

Amber: prepare, but do not decide

For consequential work, AI might help organise evidence, summarise records or outline options. Require a named reviewer to check important claims against the original records, consider relevant context and approve the work before it affects a person.

Independent verification means returning to the underlying evidence or authoritative record—not asking a second AI system to confirm the first. Singapore’s court guide places responsibility on court users to ensure material they put before the court is independently verified, accurate, true and appropriate. 3

Red: reserve the decision for a person

A company can designate decisions involving matters such as employment, safety or access to essential services as human-reserved. In those cases, AI may assist with preparation, but should not make or automatically trigger the final decision. The responsible person should assess the case, be able to reject the system’s recommendation and explain the decision.

This is a recommended corporate safeguard, not a claim that Singapore’s courts prescribe this exact rule for businesses. The judiciary has said that AI tools may assist judges in reading and digesting material, while its current approach does not deploy AI to weigh evidence, apply law to facts or exercise discretion. 40

Make the rules public and specific

A public policy is more useful when it describes actual tasks instead of announcing a broad commitment to “responsible AI.” For every approved use, publish a short entry that answers:

• What may AI do in this task—and what is out of bounds?
• What information may employees provide to the tool?
• Who checks the output, and what must they check it against?
• Who owns the final decision?
• How can an employee or affected person raise a concern or challenge an error?

Internally, keep a proportionate record of the AI tool and task, the sources reviewed and the person responsible for approval. Give reviewers enough time and authority to disagree with an output; otherwise, human review can become a formality rather than a real safeguard.

Recheck the classification and protect staff judgment

Treat each colour assignment as a decision to revisit, not a permanent label. Test the specific tool in the specific workflow, examine errors and decide whether controls need to change or use should pause. Avoid assuming that a general claim about a model applies to every task.

Also preserve opportunities for employees to practise the work AI is helping with. Singapore’s Chief Justice has raised concerns that AI may reshape routine legal tasks that have traditionally helped junior lawyers build foundational skills. 2335 For companies, the practical lesson is broader: staff need enough experience to understand the work, spot mistakes and exercise independent judgment before they can review AI output meaningfully.

A simple standard to apply

The traffic-light framework is most useful when each colour changes what happens in practice: what AI is allowed to do, what a person must verify, who remains accountable and how a decision can be challenged. Publish those boundaries by task, review them as the tools and workflows change, and keep consequential judgment with people who have the authority and skill to exercise it.