Honeywell Technologies Corporation ( Honeywell Technologies ) has launched a new AI platform that can operate in three modes: manual, semi-autonomous, and autonomous. The company's Chief Technology Officer has not yet enabled the third mode.
"We haven't yet enabled autonomous mode," Suresh Venkatarayalu said in an interview with Fortune. The interview took place before the release of Forge Cognition on Wednesday. The platform is built using the technology of Nvidia and Dell, and it overlays AI agents and digital twins onto the systems responsible for building heating, cooling, security, and power supply. Honeywell stated in its announcement that these agents can "cooperate autonomously among existing systems." However, Venkatarayalu mentioned that at least for now, there will still be people involved in the process. "Maybe we will start operating in this mode in the next two to three years," he said. After that, whether to switch will be decided by the customers.
"For me, without a control system, I wouldn't deploy these proxies," he said.
The gap between what the software is capable of doing and what Honeywell is willing to allow it to do is precisely where this bet lies. Forge Cognition is Honeywell’s first major product after its reorganization. Over the past year, the company has divested itself of its advanced materials and aerospace businesses and now mainly sells automation products for buildings, factories, and process plants. This also represents an early test that all companies that introduce AI agents into their operations must face: how much control are you willing to hand over to the software that operates the equipment people depend on? Venkatarayalu calls this platform “the brain behind the building.” For now, at least, this “brain” is responsible for providing recommendations, while operators are responsible for making decisions.
Nvidia 90-day Test
This project began with a challenge on CES in January of this year. According to Venkatarayalu, after the management of Nvidia listened to Honeywell's presentation on autonomous buildings, they posed an obvious follow-up question: “Who are your lighthouse customers?” In other words, which real building can prove its effectiveness?
“We don’t believe in just doing POC,” said Venkatarayalu, where POC refers to proof of concept. Honeywell’s CEO, Vimal Kapur, then turned the question around and asked Nvidia, “Why can’t you become my lighthouse customer?”
Nvidia accepted this challenge. Within 30 days, Venkatarayalu went to Santa Clara to present a proposal to the operations and real estate teams of Nvidia. Subsequently, the two companies set a 90-day goal: to demonstrate meaningful results on the approximately 750,000 square foot headquarters building Voyager in Nvidia before Honeywell's investor day in June.
Venkatarayalu said that Nvidia's operations and real estate manager, Scott Ekman, wanted more than just a demonstration; he also had his own business objectives. The two parties agreed on five performance targets, conducted measurements at the construction site, and tracked the progress of about 900 devices on a weekly basis. By late April, Honeywell had prepared a three-minute on-site demonstration that was sufficient to present to investors.
The first objective is to trigger alarms. The design of building systems tends to err on the side of over-reporting rather than missing any potential threats, so all situations are marked. Venkatarayalu states that in the security and access control systems of Voyager, about 90% to 95% of the alarms are false alarms, and Nvidia has eight to nine people dedicated to filtering out these false alarms per shift. Honeywell combines the open-source Nemotron model from Nvidia with its own engineering expertise to distinguish between genuine alarms and noise. Honeywell claims that this approach has reduced unnecessary alarms related to the building management system by up to 90% in Voyager.
The operator can review every decision made by the proxy. The operator can open each alarm categorized by the system and mark whether they agree with it. “This is all done to keep the operator involved in the process,” says Venkatarayalu, so that the operator can gradually learn to trust it.
The second goal is energy. The agent split the utility bills for Voyager by floor, area, and peak-valley time periods, and then discovered a pattern in the building heating methods.
Venkatarayalu said that they checked how many areas had their heating systems set to a standard value, "for example, around 58 degrees throughout the year," and yet the system did not actually detect external temperatures exceeding 58 degrees. "Why should I still run the heating system?" he asked.
"We have identified an energy efficiency improvement of nearly 10%," he said. This figure represents the potential savings that Honeywell and Nvidia have identified by more intelligently understanding and optimizing operating conditions, which have not yet been truly realized.
Dialogue with Architecture
In the simplest mode, Forge Cognition is a chatbot in the field of infrastructure. “Can I talk to a building? Can I talk to a physical asset?” asks Venkatarayalu. Operators can inquire about performance and trends, and then click to enter the digital replica of Voyager. This replica is constructed by the simulation software of Omniverse using Nvidia, with details down to individual layers, single areas, or individual machines. Dell provides the servers for running it on-site.
Venkatarayalu describes this change as an evolution in system functionality. Traditional building automation can "see and act": sensors collect data, and operators respond accordingly. The new version adds "thinking" and "learning" capabilities: AI models, trained based on Honeywell's existing infrastructure and engineering documentation, interpret the information captured by the sensors, and operator feedback improves these models over time.
"For me, autonomy means automation, and more automation," he said.
The approach to learning varies in different locations as well. At the Honeywell campus in Bangalore, the goal is to save energy. At Voyager, Venkatarayalu states that the management at Nvidia told him, "For me, comfort is more important." Agents not only need to learn how the equipment operates but also what goals each operator wants to achieve.
In his view, the threshold to achieve complete autonomy is very high. Before customers make this decision, he says that the system should “always achieve an accuracy rate of 10 out of 10.”
Skill Gap
Venkatarayalu positions this product as a help for operators. “At least I don’t think it’s meant to replace people,” he says. “In an industrial environment, I need a co-pilot to assist and guide me, to help me do my job better.” In the United States, what they see is a shortage of labor; in other places, there is a shortage of skilled operators.
He said that the person in charge of operations for Nvidia also described the same issue. According to Venkatarayalu, Scott stated that the current performance of this system is comparable to that of the best building operators he trained, and it took him 10 to 15 years to build this team. As Nvidia adds more large and complex buildings, Scott informed him that the company will not be able to recruit and train such talent fast enough.
Predicting equipment failures is equally important. Venkatarayalu states that some replacement parts can take 6 to 18 months to arrive, so knowing in advance which components will fail allows operators to plan ahead.
Nvidia also hopes that its employees receive training. According to Venkatarayalu, Nvidia will only sign the service contract after Honeywell's frontline deployment engineers have trained the Nvidia operations team and chief engineers, enabling them to continue providing feedback to the model.
Nvidia Robot and Edge AI Vice President Deepu Talla stated in a release: "Physical AI is bringing new intelligence to the complex systems that operate our buildings and industrial facilities."
The new brain for old businesses
It has been four months since Honeywell completed the division of its business units. The Advanced Materials business became Solstice Advanced Materials in October 2025, while the Aerospace business began to trade independently in June 2026. The remaining company, Honeywell Technologies, focuses on automation. This division was prompted by a corporate review and pressure from aggressive investor Elliott Management.
Honeywell stated that its current business portfolio covers 350,000 sites. Forge Cognition may target this market; the company has not yet indicated how many sites will adopt this product. Honeywell's history can be traced back to the 1885 patent for a furnace temperature regulator, and providing a "brain" for building control systems is its next step.
Next step
Venkatarayalu said that by August, the Voyager project had reached a "critical point." The next step is to turn Nvidia into a paying customer. "This is no longer a pilot or POC," he said. To reach a commercial agreement, it means that Honeywell's service team needs to determine the deployment scale and be responsible for operations. Nvidia is currently selecting more buildings for the platform, which may also include two or three in the United States and one in Berlin, but he stated that nothing has been finalized yet.
Honeywell is also piloting at the Fairmont Hotel in Dubai, the NEX Mall in Singapore, as well as its own parks in Bangalore and Richardson, Texas (the Forge Cognition). They stated that customers can now begin planning and deploying these initiatives. Venkatarayalu mentioned that within the next six months, the company has selected 12 test locations across 12 different industries, including a factory operation of a client in Nigeria. Entering each new industry means retraining the models to address different issues and achieve different goals, he said.
Until these customers decide to turn it on, the autonomous mode will remain off.
