Autonomous operations

Autonomous should mean
fewer surprises, not more.

Two kinds of autonomy here, and it matters which is which. The agents carry a condition from a raw reading through to a work order with nobody in the chain, but they never write to the plant. Schedules do write: when a unit has drifted off the schedule you set, the correcting value goes back automatically. Both run unattended; only one touches the plant, and the value it writes came from a rule a person wrote. No AI output has ever reached a setpoint, and that is a design decision rather than a gap.

The autonomous loop

The loop that runs without anyone watching.

This is the autonomous part: agents carry a condition from a raw reading to a work order in someone’s queue, overnight, with nobody in the chain. Each step is a separate stage rather than one opaque model, so a verdict can be traced to whichever step produced it.

1 · Sense

Readings validated before anything is concluded from them — drift, stale values, out-of-range and flapping caught first, because a diagnosis drawn from a dead sensor is worse than none.

2 · Confirm

A threshold crossed once is a spike. The condition has to hold across consecutive intervals, with a delay and severity that suit the asset, before it counts as an event.

3 · Diagnose

The fault is named and traced through the equipment chain, with candidate causes ranked by likelihood and the evidence for each one shown.

4 · Decide it matters

Not every confirmed fault earns a job. Rules on criticality, equipment type, site and alert code decide which ones do — so the queue does not fill with noise.

5 · Raise the work

The work order is created by the system, not by a person: asset and evidence attached, priority mapped from criticality, SLA clocks armed, and one order per event so duplicates are impossible.

6 · Hold the line

Separately from the agents, schedules are checked against what the plant is actually doing. When a unit has drifted off its schedule that is recorded as a violation and a correcting write is queued, executed and tracked — including when it fails.

7 · Stand down

If the condition clears before anyone starts, the order can close itself. Nobody has to go and tidy up after a fault that resolved on its own.

Deterministic control

The other layer: control, and who authored it.

Occupancy schedules per group, the values currently written to the plant, and an attribution log. Where the plant has drifted off its schedule that shows as a violation and the correcting write is queued against it — tracked through to success or failure. Everything here is deterministic: a schedule you wrote, or a person with the permission to write. Nothing on this screen was decided by a model.

Occupancy schedules by group, values written to the plant, and an attribution log of automatic and manual changes — Interface representative. Figures are illustrative demonstration data.

Where the line is

The AI does not touch the plant.

It diagnoses, ranks and recommends — and where a fix is a setpoint change, it puts that control next to its own diagnosis so an engineer can act in one place. The engineer still presses it. No model has authority over a setpoint, and none is planned.

On life-safety systems the line is harder still — fire panels are joined listen-only, and nothing is ever written back to them.

Autonomous — agents

Validation, confirmation, diagnosis, and a work order raised by the system when a confirmed condition earns one. No human in the chain, and no write to the plant.

Autonomous — schedules

Control that does write. A schedule is checked against the plant, and when the plant has drifted off it the correcting value is written back automatically — deterministic, because the target came from a rule a person wrote.

Never

An AI verdict driving a setpoint. Writes to points not marked writable. Changes without attribution. Any write-back at all to a certified life-safety panel.

What it connects to

Autonomy is only worth having downstream.

A schedule that runs itself saves energy quietly. The rest of the value shows up when a persistent condition becomes a diagnosed fault, that fault becomes a work order, and the saving gets held to a standard instead of asserted.

Occupancy schedules per Equipment and Space Group
Setpoints written only to points marked writable
Permission-gated: viewing and editing are separate rights
Schedule violations detected and corrected by an automatic write
Manual override with an expiry, reverting on schedule
Every change attributed, automatic or manual
Listen-only on certified life-safety systems

Start with the schedule, not the algorithm

Plant running outside occupied hours is expensive, and it is the kind of loss that accumulates quietly. A schedule is easy to write and easy to lose — someone puts a unit into hand for a callout and it stays there for months. Autonomy that pays is a schedule that is checked against the plant and corrected when it drifts, everywhere, without anyone remembering to look.

Book a demo See the whole platform