AI development for Niagara Tridium and building automation is Software Depo’s most defensible specialty: we combine production AI-agent engineering with real operational-technology experience. The result is AI added around controlled Niagara workflows — reading data, drafting reports, triaging alarms, answering engineering questions — never given unrestricted control over a building.
The AI is ours. The Niagara engineering is delivered by our sister practice SoftwarePile, the SoftwareMile Niagara Framework development team — so every integration is grounded in genuine station and module expertise, not AI guessing at building systems.
Safety First, Always Read-Only to Start
Buildings are operational systems where a wrong command has physical consequences. Every AI-for-Niagara engagement starts read-only: the agent can query, summarize, and recommend, but any consequential action requires scoped permissions, validation, audit logging, and human approval. We add write capabilities only deliberately, one approved action at a time.
Niagara Knowledge Agent
Answers from your engineering standards, sequences of operation, customer SOPs, Niagara manuals, module documentation, commissioning records, service tickets, and site-specific notes — with citations, and respecting who is allowed to see what. It turns a filing cabinet of tribal knowledge into something a technician can query in plain language.
Niagara Alarm Triage Agent
Receives or queries an alarm, gathers related points, histories, and equipment context, compares the condition against SOPs and known faults, prepares a probable-cause summary, recommends checks, and drafts or enriches a service ticket — then requires human confirmation before any consequential action. It shortens the path from alarm to informed response.
Niagara Station Audit Agent
Automated checks across a station or fleet: missing histories, incomplete alarm routing, naming-standard violations, untagged equipment, disabled or stale points, license headroom, down devices, communication faults, and configuration inconsistencies — producing a prioritized, technician-ready report.
Niagara Engineering Copilot
Assists developers and engineers with BQL creation, Baja API examples, module documentation, code analysis, station-report interpretation, migration checklists, test-case generation, and error-log explanation — accelerating engineering work without replacing engineering review.
Niagara MCP Server
A secure MCP server that exposes only approved capabilities to AI — read station metadata, query histories, retrieve alarms, run approved reports, read module inventories, search documentation, create a draft ticket. It starts read-only; write actions are added only through scoped permissions, validation, audit logs, and approval steps. For independent security review of that boundary, our sister practice BulletproofSoft assesses AI and MCP deployments.
Niagara Workbench MCP — In Development
We are currently building a deeper MCP layer that gives a desktop AI assistant governed control of the Niagara Workbench itself — creating and editing wiresheets, managing running stations, and carrying out engineering tasks that today require someone driving the tool by hand. It runs through the same gates as everything else on this page: individually scoped tools, allowlisted actions, audit logging, and human approval for anything consequential.
Status: in development. We are describing this as a reference architecture and an active build, not something you can buy today. It is genuinely more capable than the Niagara MCP work we have published publicly — and we would rather tell you it is unfinished than imply otherwise. When it is production-ready, this page will say so plainly.
AI Autopilot with Human Escalation
Autopilot does not mean unattended control. It means the agent operates autonomously within a defined scope — the routine, allowlisted, low-consequence decisions — and hands anything beyond that scope to a person rather than guessing. When a decision exceeds what it is permitted to make alone, it escalates: a WhatsApp, SMS, or phone message to the responsible manager, carrying the context needed to decide quickly.
That escalation path is the feature, not a fallback. An agent that knows the boundary of its own authority — and reaches a human when it hits that boundary — is the difference between automation you can leave running and automation you have to supervise. Every decision, escalation, and action lands in the audit trail either way.
We have delivered an AI autopilot for a Niagara station at a major North Carolina electric utility, built to this pattern.
Station Status Widgets and Dashboards
Alongside the agent work we build Niagara station status widgets and dashboards — surfacing live station health, device and communication state, and the conditions worth acting on, in a form operators and technicians can read at a glance. These are delivered, working components; they are also what makes the AI layer above them useful, because an agent reasoning over a station is only as good as the visibility into it.
Automated Reporting
Daily alarm summaries, fleet-health reports, station-compliance reports, energy anomalies, software and module inventories, executive summaries, and technician-ready diagnostic packages — generated on schedule from live station data.
The Niagara Foundation This Runs On
None of these agents work on messy data. AI on a building whose points are unmapped, mislabeled, or inconsistently named produces confident nonsense — so the AI layer is only ever as good as the integration beneath it. That foundation is SoftwarePile’s Niagara engineering, and their team writes about exactly the groundwork this depends on:
- AI for Building Automation: What’s Real and What’s Hype — the same read-only, human-in-control philosophy, from the building-systems side
- Project Haystack and Semantic Tagging — why building data needs structure before any analytics or AI can use it
- BACnet vs Modbus — the protocol integration that gets clean data into the system in the first place
Tell us about your Niagara environment and what you want AI to help with — your Niagara version, station count, and whether you need read-only insight or approved actions. We scope from a safe starting point.
Related services
- Private and On-Premise AI Deployment — Building-systems data usually stays on-premise: private and on-premise AI.