Ken Sinclair's building autonomy roadmap starts with the sensor labels
Ken Sinclair's five-step roadmap for autonomous buildings keeps existing HVAC hardware and starts by tagging every sensor point in an open ontology. His checklist starts by mapping points into one semantic graph and ends with a one-subsystem pilot where a person approves setpoint changes.
This story was produced through MarketScale. See how Building Management teams put it to work with Customer Stories & Case Studies.
Key takeaways
Sinclair’s step one for autonomy is to map and tag existing automation points so the data is usable. Siemens says its AI-Assisted Data Onboarding automates tagging and structuring for existing buildings.
Commissioning and fault-detection programs can achieve about 20% savings in some cases, according to a sponsored Facility Executive piece on campuses.
Free workspace
Turn your Building Management expertise into content.
Record interviews, organize footage, and write with AI on a free trial of the MarketScale platform for qualifying companies. No demo required, no credit card.
B1_FL2_VAV14_ZNT. Strings like that one, cryptic point names locked inside a proprietary building automation system, are where Ken Sinclair says most attempts at building autonomy stall. In a November 13 piece for AutomatedBuildings.com, Sinclair lays out a five-step path to autonomous operation. Functional HVAC, lighting and electrical equipment stay where they are. On top go three layers: semantic context, intelligence at the edge, and supervisory control that only acts within set limits.
For a facilities director with a mixed portfolio and a finite capital budget, the useful part is the order of the steps. Sinclair's sequence makes the building's data understandable before any software is allowed to touch a setpoint.
Why the labels come before the AI
Step one in Sinclair's roadmap is an independent data layer. Middleware such as Tridium Niagara, EdgeX Foundry or an MQTT broker translates legacy BACnet, Modbus and LonWorks traffic into modern APIs. Then every sensor and actuator gets tagged using an open ontology: ASHRAE 223P, Brick Schema or Project Haystack. The point is cause and effect. An AI agent can't reason about Conference Room B until something has told it that Sensor_104 is the room's zone temperature sensor and that AHU-2 serves it.
Commercial platforms describe the same requirement in their own terms. PassiveLogic, which launched what it calls Level 3 Autonomy on July 8, says its physics-based World Model records what each piece of equipment is, why it exists, how it behaves and how actions produce outcomes. Siemens, on its autonomous buildings page, lists an AI-Assisted Data Onboarding service that it says automates tagging, structuring and scaling across sites, reducing onboarding effort and accelerating time-to-value for existing buildings. A dedicated onboarding service suggests that tagging is where much of the labor sits.
Semantic tagging attaches machine-readable meaning to a point: what it measures, which space it serves and which equipment feeds that space. Brick, Haystack and ASHRAE 223P are open ways to record those relationships, so the resulting map isn't tied to one controls vendor's naming conventions.
That matters most for owners whose buildings run on several generations of controls from different vendors. Each system names points its own way. Without a common map, every new analytics or optimization tool starts with the same expensive translation job, building by building.
Where the guardrails live
Sinclair's second step keeps the intelligence on site, since autonomy that leans entirely on a cloud connection can be undone by latency or a dropped network. For that local layer he points to two kinds of hardware: industrial edge controllers, and open-architecture edge nodes like ARM or RISC-V gateways. Either can run local optimization loops and lightweight machine-learning models, known as TinyML, right in the building.
The hard limits go in at the controller level: minimum airflow, freeze protection, pressure limits. Inside those limits, the optimizer is free to do what it likes. The edge controller's job is to block any command that would damage equipment or violate a safety code.
With the limits set at the controller, the review question becomes whether those limits are right, and a controls engineer can check that point by point without having to judge a model's internals.
PassiveLogic frames the same concern as traceability. Its July release says most of today's building automation operates at Level 0: hand-coded logic that acts like a black box, where actions happen but the intent behind them stays hidden. The company presents Level 3, powered by NVIDIA, as an AI "white box" that is dynamic and trainable, yet deterministic and traceable, and says its platform co-manages building systems without constant tuning or manual intervention.
A platform that promises generalized autonomy could tempt an owner to skip Sinclair's slower sequence. PassiveLogic CEO Thomas Kiessling said in the release that the company has "moved autonomy from theory into deployment in live environments, something the industry has anticipated for years." Yet the company's own description of the problem leads back to step one. It says buildings require coordination across thousands of interconnected assets, multiple physical domains and time horizons from seconds to decades, and its World Model starts by describing what each piece of equipment is. That suggests the labeling work moves inside the platform rather than disappearing.
Siemens describes the human role in terms close to Sinclair's. Autonomy, its page says, manages routine adjustments, surfaces only meaningful exceptions and keeps people in control of intent, policies and trade-offs. Put next to the controller-level limits, that sketches a division of labor: engineers set the boundaries and the goals, and software handles the adjustments in between.
What the early numbers measure
Steps three and four are where the comfort and energy results are supposed to show up. Sinclair's premise for step three is blunt: traditional automation reacts to conditions after they drift out of range, and autonomy should anticipate them. So fixed schedules give way to prediction. Calendar data from Google Workspace or Microsoft 365, badge reads and anonymous occupancy sensors forecast room use hours ahead. Model predictive control then combines a thermal model with weather forecasts, time-of-use tariffs and grid carbon intensity, so the building can pre-cool during off-peak hours or coast through a demand event.
Step four is continuous automated fault detection and diagnostics, or AFDD, which Sinclair frames as the building watching its own mechanical health instead of waiting for tenant complaints. It catches problems like a leaking chilled-water valve, a stuck economizer damper or a drifting CO₂ sensor before they turn into energy waste or equipment failure. He pairs it with self-healing sequences. When a non-critical sensor or component fails, the system switches to a backup sensor, adjusts adjacent zone dampers to compensate and generates a diagnostic work order with root-cause context.
That last piece could matter most to a stretched maintenance team. A work order that arrives with a probable cause attached starts the repair where a complaint about a hot office would otherwise start with someone walking the floor.
Reference points from early deployments
Siemens (NWT Fastighet pilot reference, 2026); Facility Executive (Ryan Gaylord, Altura)
Siemens cites all three figures to its own pilot reference for Comfort AI, formerly Cloe, at NWT Fastighet in Karlstad, Sweden. The useful thing is the unit: two of the three results are compliance, meaning how much of the time a zone stays inside a defined band. That only works if someone has written the band down first.
The 20% figure comes from Ryan Gaylord of Altura, writing in a sponsored Facility Executive piece on university campuses, and it isn't an autonomy number. It's what he says retro-commissioning, monitoring-based commissioning and ongoing commissioning programs can achieve when fault detection and analytics cover every connected asset. His worked example: a university spending $30 million a year on utilities would free about $6 million a year at that rate, enough, he writes, to partially fund a recurring deferred maintenance program or hire a large team of full-time researchers.
His case for why the number is reachable is about timing. Conventional commissioning, he writes, is a snapshot: a technician checking a box on a clipboard on a random Tuesday afternoon, which can't catch a sequence error that only shows up at 2 a.m. on a Sunday. Because nearly every building drifts from its intended performance over time, connected commissioning replaces the snapshot with continuous analytics, automated functional testing and FDD.
That is the same machinery as Sinclair's step four. An owner already running a serious monitoring-based commissioning program likely holds a working piece of the autonomy stack, along with the point data it depends on.
Writing autonomy into the contract
Sinclair's fifth step is commercial, and he calls it Autonomy-as-a-Service with digital SLAs. His reasoning is that technical readiness only succeeds when it lines up with how the building is operated and financed. He argues against attempting full-building autonomy at once and suggests targeted modules first: central chiller plant optimization, automated demand response through OpenADR, or dynamic zone ventilation.
He also wants service-level agreements among landlords, tenants and service providers written in machine-readable terms, holding 21°C to 23°C and under 800 ppm CO₂, for example, so the software has explicit targets it can be audited against.
Owning the data underneath matters just as much. Gaylord describes campuses locked into proprietary controls, where adding third-party analytics or new hardware becomes too expensive and too complex, and owners are stuck with single-vendor pricing. Decoupling the software and hardware layers, he argues, opens the door to competitive procurement across campuses. He points to Dartmouth and the University of California, Irvine, which partner with master systems integrators to transition standalone building automation systems into open-protocol, enterprise control platforms.
His suggested path is phased: apply an enterprise automation and data standard to new construction and major renovations, then move existing buildings onto it as their equipment reaches end of life. Without enterprise specifications for automation, metering and data integration, he warns, each new building operates as an isolated technology island. His checklist for facility leaders adds a concrete requirement: connected commissioning, including virtual testing of control sequences, on upcoming capital projects. For owners whose first autonomy pilot is years away, specifications for upcoming projects could be the first place that standard goes.
Gaylord also ties the efficiency work to capital planning. Many university decarbonization roadmaps depend on capital-intensive projects such as central plant electrification, and he argues that cutting thermal and electrical loads first lets institutions downsize the capacity, complexity and capital expenditure those later projects need. That turns Sinclair's sequencing into a budget argument as well as a technical one: the data and commissioning work comes first and can shrink the bill for the plant project that follows.
Before signing for an autonomy module, ask which ontology the vendor's point mapping uses (Brick, Haystack or ASHRAE 223P) and whether the tagged model can be exported. If it lives only in a proprietary format, the most expensive part of the project stays with that vendor.
The first pilot and the Level 4 test
Sinclair's starting checklist is short enough to go straight into a capital request:
- Audit and normalize: map existing BACnet and Modbus points into one semantic graph using Brick, Haystack or 223P.
- Secure the edge: move to encrypted local communication such as BACnet/SC and TLS-encrypted MQTT, and isolate OT networks from IT.
- Connect external feeds: bring in real-time utility pricing, weather data and room-booking schedules.
- Pilot closed-loop control: turn on supervisory setpoint optimization for one air handler or chiller loop, with a person approving changes before anything runs unattended.
Gaylord's list opens in nearly the same place. His first step is to establish a baseline by connecting existing BAS data to an analytics platform, benchmarking energy use and finding the largest opportunities. One writer comes at the problem from building autonomy and the other from campus commissioning, and both put access to the data ahead of any new control logic.
The next milestone to watch is on PassiveLogic's calendar. The company says it is targeting Level 4 Autonomy later this year, adding learning at the edge and longer prediction horizons. Any such release would still need a building whose points someone has already labeled.
Sources
Featured companies
Your experts belong here
Every story in MarketScale Building Management starts with a company putting its facilities engineers, energy managers, and service technicians on the record. Buyers are already reading this topic. The only question is whose experts they find.
Owners and facilities teams pick on trust, and your engineers turn that trust into inbound conversations.
About the author
The MarketScale Newsroom reports on the companies, technologies, and trends shaping 16 B2B industries. It turns primary sources and expert commentary into clear, useful coverage for the people doing the work.