Fact vs model
The twin knows what the regime should be. Before raising anything, the agent checks the sensor, the valves, the batch and the plan.
AI AGENTS
Agents that know your assets.
Each market gets its own agents. They watch the telemetry, check it against the digital twin and the plant ontology, and bring the operator a finding with its evidence and sources. The operator decides.
THTurbine healthHIGHCompressors · 1 min ago
CS1 U1: compressor wash recommended. Degradation index 3.2; power at the limit in about 17 min.
AcceptPinInspect
DADispatcher advisorCRITICALPower · 5 min ago
500 kV line tripped, auto-reclose unsuccessful. Test by voltage, check protection settings and fault locators, restore via reserve lines.
AcceptPinInspect
CHCooling headroomHIGHData centre · 9 min ago
Hall B: rack-inlet limit in about 25 min at the current training ramp.
AcceptPinInspect
Four patterns, every market
What an agent does for the person on shift.
AI alarm
Not what happened, but what will happen and what to do, with options already checked against the twin.
Ask the asset
Plain-language questions answered from your tags, archives and documents. Every fact carries its source.
Shift handover
What happened, what is still open, what to watch next. Written from the shift’s own data, ready at handover.
Agents by market
The same loop, a different crew of agents.
Dispatch centres, substations and generation, from one national model.
In production on a national transmission gridOverload forecasterHIGHPowerjust now
Line 5112 reaches its limit in about 1 h 40 min. Scenario B, redistribute and reduce, is validated safe by the twin.
Sources: SCADA · state estimator · twin run
DA
Dispatcher advisor
- Watches
- Protection events, switching, limits
- Delivers
- Grouped incident, cited instruction, next steps
OF
Overload forecaster
- Watches
- Line and transformer loading
- Delivers
- Time to limit and twin-checked scenarios
LF
Load forecaster
- Watches
- Demand, weather, calendar
- Delivers
- Short-term forecast, under 5% test MAPE
GZ
Gray-zone watcher
- Watches
- State estimation beyond sensors
- Delivers
- Places at risk of supply failure
SH
Shift handover
- Watches
- The whole shift
- Delivers
- Summary of events, open items, risks
How an agent thinks
Not tags. Connected objects.
A SCADA stores thousands of values with names like NPS1.MNA2.P_OUT. People know what is behind them; a language model does not. The ontology gives every object a type and its relations: a station creates head on a section, a batch flows through it, its viscosity changes the pump’s regime.
Everything known about an object hangs on its node: live telemetry, passport and limits, procedures, repairs and incidents. So when asked why outlet pressure fell, the agent walks the graph from sensor to pump, to batch, to its properties.
Every answer comes with a link to its source: a tag, a document, a log entry.
Live feed
One stream, already grouped.
Agents post to one feed. Accept, pin or dismiss each finding; the decision and its reason go into the log. Filter by market to see each crew at work.
Turbine healthHIGHCompressors1 min ago
CS1 U1: compressor wash recommended. Degradation index 3.2; power at the limit in about 17 min.
Sources: DCS · EGT · ΔP filter
Dispatcher advisorCRITICALPower5 min ago
500 kV line tripped, auto-reclose unsuccessful. Test by voltage, check protection settings and fault locators, restore via reserve lines.
Sources: Protection events · outage-prevention instruction §4
Cooling headroomHIGHData centre9 min ago
Hall B: rack-inlet limit in about 25 min at the current training ramp.
Sources: BMS · cluster scheduler
Heating balanceMEDIUMPipelines13 min ago
Heater point 2: current gas rate leaves 3 °C margin to pour point at the next heater. Raise outlet temperature by 2 °C.
Sources: Thermal twin · batch passport
Catenary voltage watchMEDIUMRail17 min ago
Contact line Shu–Otar: 22.8 kV during acceleration of train 3411, below the 24 kV warning level.
Sources: Traction SCADA
Simulated findings. Station, line and train names are illustrative.
Autonomy, one step at a time
Start as an advisor. Earn the right to act.
- Level 1Advisor
Read-only mirror of your SCADA, DCS or BMS. Agents post findings; no command reaches the control loop.
Where pipelines and stations start - Level 2Co-pilot
The agent prepares the action, the twin validates it, the operator approves, and the runtime executes it through SCADA.
Next step on a national grid - Level 3Pre-approved actions within limits
A bounded set of reversible actions, approved in advance by your engineers, runs inside limits taken from your procedures. Every step is logged and the approval can be revoked at any time.
Earned per action, never by default
Questions
What operators ask about agents.
What does a Faceplate AI agent do?
It watches the telemetry, checks it against the digital twin and the plant ontology, and brings the operator a finding with its evidence and sources. The operator decides what happens next.
Can an agent change the process on its own?
Not by default. Agents start as advisors on a read-only mirror of your SCADA, DCS or BMS. At the next level the operator approves each prepared action. Only a bounded set of reversible actions, approved in advance by your engineers, may run inside limits, and that approval can be revoked at any time.
How can we check an agent’s answer?
Every answer comes with a link to its source: a tag, a document or a log entry. The ontology connects each object to its telemetry, limits, procedures, repairs and incidents, so the path to the answer can be followed.
Which industries are agents available for?
Power grids, compressor stations, pipelines, rail, data centres and cities. Each market has its own crew of agents working the same loop.
Does our data leave the site?
It does not have to. Faceplate runs on-premise, and the agents can use domain-adapted language models on your own servers.
Where do we start?
With one agent in advisor mode: a read-only mirror of your control system, where findings are posted to a feed and no command reaches the control loop.