Your BMS is already running. What needs to be added for AI control?

Do not give AI a task the controller already handles

Schedules, setpoint adjustments and even adaptive optimal start can be part of a conventional BAS. PNNL includes these functions in its list of re-tuning measures. ¹

First, therefore, we check whether restoring an existing function, fixing a sensor or adjusting a control sequence would be sufficient. A separate AI layer should add value, not relabel controls already in place.

Proven rules will be sufficient for one task. Another may require a model of the building's response, a forecast or the identification of more complex patterns. The method is chosen after the problem has been defined.

Turn tags into a clear structure

The name `AI_17` does not explain what is being measured. Moreover, in controls documentation, AI often means analogue input, not artificial intelligence. In the Project Haystack model, such points are linked to their function, equipment and site. ²

Analytics need answers: which unit's supply air temperature is this, which zone does it serve, which commands affect this parameter, and where is its history stored?

This context can be described through an agreed data model. Brick, for example, is designed to represent building equipment and the relationships between its physical and digital components. ³

There is no need to build a complex model of the entire site at once. For a pilot, it is enough to describe the relevant control loop consistently while retaining its links to associated systems.

A digital asset record is not a predictive model

Equipment lists, points, documents and service history provide digital context. But on their own, they do not predict how many minutes an office will take to cool down.

Prediction requires an additional relationship, calibrated and validated for the task. It may be physics-based, statistical or a combination of both. Calling any equipment catalogue a ready-made predictive twin would be inaccurate.

This is why the technical specification should distinguish between the outcome of structuring the data and the outcome of model validation.

From recommendation to command

Initially, the system may only observe: identifying situations and suggesting changes for an engineer to assess. The next step is a validated pilot in an agreed zone.

Automatic actions require limits: permitted parameters, ranges, rates of change, priorities and a return to normal control logic. It is also important to verify that a command has not merely been sent, but accepted by the equipment.

A cloud model must not be the sole provider of critical protective functions. We propose using it as an optimisation layer above functioning local controls — this is the approach to integration with existing systems described in HVAC AI on the SM Systems website. ⁴

Compare against actual baseline controls

The calculation must show the difference between the new solution and what the building can already do. It is not valid to compare AI against a deliberately poor operating strategy and attribute the entire difference to the technology.

Simulation allows control algorithms to be compared under identical conditions. BOPTEST, for example, uses virtual buildings and control points for this testing. But the result remains a model output, not a meter reading from a particular client. ⁵

For a pilot, we define the metric in advance: energy, cost, comfort, peak demand or manual interventions. We do not promise the same percentage for every system.

AI readiness is not about the number of connected sensors. It means a clear task, reliable data, a controlled means of intervention and the ability to verify that the decisions have improved.