Honeywell and Ecolab Say AI in the Physical World Has Hard Limits

Ecolab cut its AI token costs by 70% to 80% after discovering that running "the best model that's out there" on high-volume work was, in the words of Chief AI Officer AJ Wijesinghe, "sometimes more expensive than having humans."

Wijesinghe and Honeywell Chief Technology Officer Suresh Venkatarayalu laid out the practical limits of deploying AI in buildings, factories and kitchens at Fortune's AIQ Summit at the New York Stock Exchange on Thursday. Both executives stayed broadly positive about the technology. Both were blunt about what it cannot yet do.

Honeywell's core problem is accuracy. The industrial giant's customers "really demand 99.9999%," Venkatarayalu said, while frontier models "could be at 85%." That gap is why Honeywell is fine-tuning open-source models rather than relying solely on off-the-shelf frontier systems.

Ecolab's problem was economics. When the water-management company put top-tier models on high-volume tasks, "the tokenomics go off the roof," Wijesinghe said. The company then optimized its models, using both frontier and open-source options, including Anthropic's Claude and OpenAI's models. "Sometimes you don't have to have the fastest car," he said.

No fully autonomous future

Neither executive described a future of full autonomy. Venkatarayalu said the realistic path is toward "a semi-autonomous world" where operators build trust in the systems over time, and he pushed back on the framing that automation eliminates jobs. "Autonomy is also not about removing people," he said.

Wijesinghe described a "human in the lead" model at Ecolab, saying agent technology "is not matured enough" to run at scale without oversight.

Physical equipment also resists the rapid-update rhythm of consumer software. Venkatarayalu said over-the-air upgrades are coming to commercial buildings, but with an operator in the loop, because buildings and industrial sites are "mission critical and safety critical." A Tesla can sit offline for 1.5 hours during an update, he noted, but "you cannot afford to have a building shut down for one and a half hours."

The Nvidia partnership and the edge question

Moderator Emily Forlini, senior AI reporter at Fortune, raised Honeywell's recently announced partnership with Nvidia, disclosed at its investor day. Venkatarayalu said Honeywell hand-picks open-source models and works with Nvidia's Nemotron team, partly because customers want sovereignty over their data and models. Deploying an unsupported open-source model without guardrails, he said, would be "dangerous."

Honeywell runs a "see, think, act, and learn" framework. It starts by cataloguing a building's assets — HVAC, fire control, security and access control — and connecting them over the BACnet protocol. AI agents then learn how those systems relate to one another and run them. Venkatarayalu said the company is also exploring how to teach AI chemistry, including catalyst formulation.

Ecolab places sensors in dishwashers, pest traps and water systems to cut service visits and predict maintenance. Wijesinghe said the company also manages water for chip production, plus power and cooling for AI infrastructure itself.

What separates results from stalls

Asked by an audience member from a systems integrator what separates AI work that creates value from work that stalls, Wijesinghe drew a clear hierarchy. Individual AI adds little measurable value at the enterprise level, he said. Vertical AI within a single function adds some. The most value comes from "horizontal" AI that works backward from a defined outcome across functions such as sales, finance and supply chain.

AI "should not be just another technology," he said, and must be tied to a company-wide transformation. Ecolab expects $325 million in annual run-rate savings by 2027, and Wijesinghe said "significant" amounts are already in hand. He framed the requirements as a balance: data foundation, process readiness and cost discipline matter equally. "If one is heavier than the other, then you don't get the value," he said.

Venkatarayalu said the choice of where AI runs — at the edge, on a customer's premises or in the cloud — depends on latency, data-transaction cost and sovereignty. The demand signal is already visible: customers who got 7% energy savings from existing controls now ask whether AI can push that to 30% or 40% more.

That question, scaled across Honeywell's installed base and Ecolab's $325 million savings target, defines the next phase of industrial AI — one measured in efficiency gains rather than model releases.