Red, amber or green: Why more companies should use this AI framework and make their lines public

SINGAPORE’S courts have a simple framework for using artificial intelligence. Chief Justice Sundaresh Menon has described the uses in court as either green, amber or red, with each colour setting a different limit on work that can be handed to a machine.

Although developed for the courts, these distinctions offer corporate executives a way to decide how AI can assist their staff while keeping people responsible for the work. In a May 2026 interview marking Singapore’s legal bicentennial, Menon warned of “truth decay” – which is the blurring of fact and opinion that weakens institutional trust. Courts depend on that trust when examining evidence, hearing both parties and explaining their decisions, so AI must improve access without undermining confidence in their decisions. Green covers tools that help people understand and use the courts. In the Small Claims Tribunals, where parties appear without lawyers, AI translates documents and summarises each side’s material without deciding who wins.

Amber requires greater caution because AI may influence the arguments presented to the judge. The courts are evaluating tools to help self-represented litigants organise those arguments. Red covers AI that decides cases or predicts judicial outcomes, neither of which the courts are considering. Even if AI could predict the outcome accurately, letting it decide the case would compromise judicial integrity.

That concern becomes easier to understand when you put yourself on the wrong end of a decision. Anyone before a judge is entitled to believe somebody considered their case, understood what was at stake and weighed the evidence. Getting the answer right is not the same as doing justice to the person standing before the court. Responsibility also applies to permitted uses. Under the judiciary’s 2024 guide on the use of generative AI tools by court users, lawyers and self-represented litigants must independently check AI-assisted submissions against the sources. Asking another AI is insufficient.

In a lecture in September, Menon explained why responsibility for judicial decisions must likewise stay with people. Judicial authority rests on reasons as well as outcomes, so parties must be able to understand a decision, challenge errors and seek review. An opaque system cannot meet that obligation if its contribution cannot be “exposed, tested and defended”.

The task, not the tool

For other professions, applying this approach means looking at what people do with AI. The same summarisation tool carries different risks when reading an annual report and a patient’s history before surgery, because an omission has different consequences.

In medicine, transcribing a consultation may be green, drafting a diagnosis amber, and consent to a risky procedure, red. The clinician must judge whether the diagnosis fits this patient and discuss treatment with them. Businesses should approve specific uses. A customer service assistant might translate delivery updates safely but need review when responding to a disputed charge. The software has not changed, yet the consequences have. The service owner must decide which uses are permitted and answer for failures. An auditor faces a similar distinction when using AI to compare invoices while remaining responsible for the opinion. Preserving that ability to judge also means protecting how people learn it.

At the Asia-Pacific Legal Congress in April, Menon warned corporate counsel that AI follows a “jagged frontier” and improves unevenly. His caution about automating junior work follows from this. If AI drafts every memo, who learns to spot weak arguments when the system gets something wrong?

Drawing your own lines

Executives can begin by breaking a job into its tasks. “Use AI in customer service” is too vague when translating a message, drafting a reply, approving compensation and closing a disputed account have different consequences.

SEE ALSO

What happens when the digital option is the only one available?

Who is accountable when an AI agent decides?

Buy or build software? Firms need to look at control, reversibility and risk

Consider each task’s consequences, whether errors are visible and whether the result can be reversed. Green work has limited consequences and correctable errors. Amber can influence an outcome, so a competent person needs the evidence, time and authority to challenge it. Red involves authority the organisation should retain – especially over someone’s rights, livelihood, health or liberty. For amber type of work, the review arrangements must allow someone to intervene. A claims officer receiving 500 AI recommendations each morning with barely 20 seconds to approve each has little opportunity to examine them. A junior forbidden to disagree with the model has even less.

For each task, specify who reviews it, what evidence and time they have, and what happens when they disagree. Keep records that explain the decision and practise stopping the system, so staff can use the override when they need it. Before expanding the roll-out, test whether reviewers spot a plausible but incorrect recommendation, consult the evidence and reject it. Fix any gaps before increasing volume. Revisit the classification when the system gains new data, tools or authority, because approval for drafting a reply should not become permission to issue a refund.

The value of red

These checks help preserve the promise each profession makes. We trust a doctor to care, an auditor to give an independent opinion, and a director to exercise judgment. Delegation is familiar in all three professions, but trust suffers when nobody takes responsibility for the result.

A patient who is told that “the system” found nothing on a scan expects a qualified person to stand behind that assurance. If automation removes that person, it may also remove the reason the patient sought professional help. More professions should use this framework and make their lines public. Doing so may reveal that a task selected for automation is what customers thought they were paying a professional to perform. Explaining those limits also gives customers a basis for challenging a decision and knowing who should answer for it. The benefit could extend beyond greater access. If professionals use the time saved on routine work to give difficult cases more attention, their organisations may earn greater trust. That depends on classifying tasks carefully and giving reviewers authority to act. Knowing where the light turns red helps people give that work the attention it deserves.

The writer is the chief technology officer and deputy chief executive at GovTech Singapore