The practical answer
Predictive maintenance and smart HVAC projects need usable operational data, a clearly defined objective and a process for acting on results. Check assets, measurements, controls, compatibility and engineering oversight before purchasing a solution. A general chatbot is not a failure-prediction system.
Separate three different building applications
| Application | Purpose | Evidence required |
|---|---|---|
| Predictive maintenance | Estimate future condition or failure risk for a defined asset problem | Relevant history or condition measurements and validated predictions |
| Fault detection and diagnostics | Identify abnormal operation and help investigate its cause | Usable system data and a checked detection method |
| HVAC optimisation | Improve operation within agreed comfort and engineering limits | Control-system compatibility, constraints and performance measurement |
The terms overlap in marketing, but they describe different decisions. NIST's work on automated building diagnostics includes expert systems, machine learning and statistical analysis. Ask which method a proposed product uses and what it has been validated to detect.
Our facility management AI upskilling guide places these specialist applications after the foundations of data quality and verification.
Check the operational foundations
Start with a defined question, such as detecting a specified abnormal operating condition on a selected system. Confirm asset identifiers, system relationships, measurement units, sampling frequency, timestamps and missing-data treatment.
For a building-management-system connection, document what information can be read and whether the proposed solution can write control settings. Reading trend data and controlling equipment create different operating responsibilities.
Identify a named engineering owner, the people who investigate alerts, and the existing response process. Without that route to action, a stream of alerts can add work without improving reliability.
Ask vendors questions that can be checked
- Which asset types and fault conditions are supported?
- What data and history are required before the first useful result?
- Which controllers, protocols and software versions are compatible?
- How are missing measurements, sensor errors and changed operating conditions handled?
- How are alerts explained, validated and prioritised?
- Which control limits, approval steps and fallback arrangements apply?
- What implementation and recurring costs are outside the quoted licence?
Ask for evidence relevant to a comparable building or asset, not only a broad savings claim. A vendor-reported case can demonstrate that an implementation exists, but it does not establish the result your property will achieve.
Design a bounded pilot and a fair baseline
Scope: one selected HVAC system. Objective: identify a defined operational fault. Inputs: agreed trend points and the relevant work-order history. Reviewer: the responsible engineer. Output: an evidence-backed alert register, including false alarms and missed conditions where known.
Agree the baseline, trial duration, review frequency and stop conditions before the pilot starts. The appropriate duration depends on the system, operating cycles and evaluation question.
When assessing energy outcomes, account for relevant weather, occupancy, operating hours and scope changes. Record the method and assumptions so an apparent improvement is not simply the result of a different operating period.
Reliability outcomes may need a different evaluation period from energy outcomes. Do not combine them into a single unsupported return-on-investment figure.
Decide whether to scale, revise or stop
Review whether the pilot found useful conditions, whether staff could act on the alerts, and whether the evidence supports the proposed benefit. Include investigation effort, integration work and ongoing support in the commercial review.
A credible decision may be to improve the data before expanding. Missing asset IDs or poor work-order records can be addressed using the practices in our maintenance history guide.
Keep the final evaluation specific: what worked, under which conditions, and what remains unproven. That gives leaders a sound basis for procurement and helps engineering teams retain control of the operational consequences.
Practical questions
Do I need sensors for every predictive maintenance project?
The necessary measurements depend on the failure or condition being assessed. Some projects can use existing operational data; others need additional instrumentation.
Does smart HVAC guarantee lower energy bills?
No. Results depend on the building, controls, operation and evaluation method. Define and measure the intended outcome rather than assuming a vendor’s case-study result applies.
Source notes
Product capabilities were checked against the following primary sources on 3 November 2026. The workflows and worked examples are illustrative management exercises, not reported client results.
- NIST: Automated Fault Detection and Diagnostics for the Mechanical Services in Commercial Buildings — building diagnostics and analytical methods.