Safe AI Control for Singapore Buildings: Human-in-the-Loop Guardrails for Predictive HVAC and Facility Operations

A practical framework for piloting AI-assisted building control with safety limits, human approval and reliable fallback procedures.

AI is moving from dashboards and predictive alerts towards recommending, and in some cases initiating, actions in real building systems. For Singapore facility managers, warehouse operators, building owners and SMEs, this creates an opportunity to improve comfort, energy management and response times.

It also creates a practical engineering question: how much control should an AI system have over HVAC, pumps, chillers and other building equipment?

A safer starting point is not full autonomy. It is human-in-the-loop control, where AI supports decisions while defined limits, approval steps, fallback modes and audit records remain in place.

Why AI control needs stronger safeguards

A conventional automation rule may be straightforward to test: if a temperature rises above a defined threshold, start a fan or adjust a setpoint. AI-driven control can be more adaptive, but it may also respond to incomplete data, unusual conditions or patterns that have not been tested thoroughly.

The consequences of a poor decision can extend beyond an inaccurate dashboard. An unsuitable command could affect indoor conditions, equipment operation, production activities, energy consumption or workplace safety. The software-engineering challenges are greater when an AI system operates in a physical, safety-relevant environment, as discussed in recent research on AI-driven building operation.

This does not mean AI should be avoided. It means the system should be designed so that useful recommendations can be adopted without allowing an uncertain model to bypass engineering controls.

Start with a clearly defined operating scope

Before selecting a model or platform, define what the AI is allowed to observe, recommend and control.

• Observe: operating data such as temperatures, pressures, flow rates, equipment status, alarms and schedules.
• Recommend: proposed setpoint changes, equipment sequencing, inspection priorities or maintenance actions.
• Control with approval: changes that require a designated facility operator to review and approve.
• Control automatically within limits: low-risk actions with clear boundaries, tested responses and a reliable override.

For an initial pilot, recommendation mode or approval-based control is usually easier to validate than unrestricted autonomous operation. A facility team can compare the AI recommendation with actual site conditions and record whether the action was accepted, modified or rejected.

1. Check sensor quality before trusting predictions

AI control is only as dependable as the data reaching it. A faulty temperature sensor, delayed equipment status or missing meter value can lead to an inappropriate recommendation.

Sensor readiness should include:

• Calibration and condition checks for important sensors.
• Time synchronisation across the building management system, meters and connected devices.
• Clear identification of missing, stale, duplicated or implausible readings.
• Data-quality thresholds that prevent control recommendations when input confidence is too low.
• Defined ownership for investigating recurring data problems.

For example, if a critical zone sensor has stopped updating, the AI should not continue adjusting a chiller or air-handling unit as though the reading were valid. The safer response may be to raise an alarm, hold the last approved state or revert to a known schedule.

2. Add anomaly detection before closed-loop control

Anomaly detection should not be limited to identifying equipment faults. It should also check whether the AI’s inputs, outputs and operating assumptions remain reasonable.

Useful checks may include:

• Sensor readings that change too quickly or conflict with related measurements.
• Commands that fall outside the normal operating pattern of a system.
• Repeated recommendations that do not improve the intended condition.
• Unexpected combinations, such as a command to increase cooling while a relevant system is unavailable.
• Control actions that would create excessive equipment cycling or conflict with an active maintenance activity.

An anomaly should produce a clear response, not just a red indicator. The response might be to pause AI control, notify an operator, request verification or switch to a predefined fallback sequence.

3. Define hard operating limits

AI recommendations should operate inside boundaries established by competent engineering and facility teams. These boundaries should be explicit and enforceable by the control architecture, rather than left to the model to interpret.

Depending on the system, limits may cover temperature ranges, pressure, flow, equipment loading, minimum run times, maximum start-stop frequency, valve or damper positions, and permitted setpoint changes over a defined period.

There should also be a distinction between comfort or efficiency limits and safety-critical constraints. A comfort optimisation can be evaluated differently from a command that could affect equipment protection, access conditions or a workplace activity.

Where a command exceeds a configured boundary, the system should block it or require a higher level of review. Operators should be able to see why the recommendation was restricted and what action is expected next.

4. Make human approval practical

Human-in-the-loop control only works if operators have enough information and time to make a sound decision. An approval screen should show the proposed action, the reason for it, the affected equipment, relevant sensor conditions, expected duration and any known risks or conflicts.

Approval responsibilities should be assigned by role. For example, a routine HVAC setpoint adjustment may be reviewed by an authorised facilities operator, while a change that could affect a critical process or planned maintenance activity may require escalation.

Avoid approval workflows that create constant low-value prompts. Excessive alerts can lead to approval fatigue. Use thresholds, prioritisation and operating schedules so that people focus on actions that genuinely need judgement.

5. Design reliable fallback modes

Every AI-assisted control function should have a documented fallback. This is the operating state used when data quality falls, communications are interrupted, the model behaves unexpectedly or an authorised person disables AI control.

Possible fallback approaches include:

• Return to the last approved setpoint.
• Use a fixed schedule or established control sequence.
• Transfer control to the existing building management system.
• Hold the equipment in a safe state until an operator investigates.
• Require local manual operation for selected equipment.

Fallbacks should be tested during commissioning and reviewed during drills or planned maintenance. A control system that has a fallback on paper but has never been tested may not provide meaningful resilience.

6. Escalate alarms with clear response procedures

An AI pilot should define who receives each alarm, what the response time expectation is, what information must be checked and when the issue is escalated. This is particularly important for sites with shifts, outsourced operations or multiple parties responsible for different systems.

Alarm procedures should distinguish between data-quality issues, equipment faults, abnormal AI recommendations and conditions that may affect people or operations. The procedure should also state when AI control must be suspended and how normal operation is restored.

Technology should support existing workplace safety and facility procedures, not replace them. MOM’s Workplace Safety and Health technology guidance highlights applications such as electronic permit-to-work, IoT environmental monitoring, video analytics and robotics. Where connected systems influence work activities, the response process and human responsibilities still need to be clear.

7. Keep an audit trail of decisions

Audit logs help teams understand what happened and improve the system over time. At minimum, record:

• The model or control version used.
• The input conditions and data-quality status.
• The recommendation or command issued.
• Who approved, modified, rejected or overrode it.
• When the action occurred and what result followed.
• Any fallback, alarm or escalation that was triggered.

These records support troubleshooting, operational learning and governance discussions. They can also help a business decide whether a pilot is ready for a broader scope.

8. Use a staged Singapore pilot plan

A practical pilot can follow a controlled sequence:

1. Baseline: document current sequences, equipment limits, alarms, schedules and operator procedures.
2. Data validation: assess sensor coverage, data quality, connectivity and historical records.
3. Shadow mode: allow AI to produce recommendations without changing equipment settings.
4. Approval mode: let authorised operators approve selected, low-risk actions.
5. Limited automatic control: enable only defined actions inside hard limits, with immediate override and fallback.
6. Review and expand: evaluate outcomes, incidents, overrides, operator workload and data gaps before adding systems or sites.

Start with a bounded use case, such as a selected HVAC zone, pump schedule or non-critical optimisation task. Chillers, ACMV systems and other equipment can be considered as the control design matures, but scope should be based on risk, data quality and operational readiness—not just technical possibility.

Moving from dashboards to controlled operations

BCA guidance on AI for the built environment identifies facilities management, workflow automation, inspections, knowledge management and data analysis as practical areas for adoption. Its Smart FM guidance also encourages structured, data-driven transformation. For Singapore businesses, the next step is to connect these capabilities to disciplined operating procedures.

The objective is not to remove facility teams from decisions. It is to give them better information, consistent controls and a safer way to act on useful predictions.

ISS can help businesses assess engineering requirements, facility workflows, data readiness and AI automation opportunities for a controlled pilot. Contact ISS to discuss your engineering, facility management or AI automation requirements.