'Sometimes it's more expensive than having humans': Ecolab gets real about AI's limits in the physical world

The neat demo version of AI breaks down fast once it leaves a chatbot and starts touching buildings, factories and kitchens. That was the blunt message from executives at Honeywell and Ecolab at Fortune’s AIQ Summit in New York on Thursday. They weren’t down on AI. They were down on the idea that the same model that writes an email can safely run a plant, a building or a water system.

Honeywell Technologies CTO Suresh Venkatarayalu put the problem in accuracy terms. Industrial customers, he said, want 99.9999% reliability. Frontier models may be sitting closer to 85%. That gap is the whole story when the job is mission critical and safety critical. A building can’t shrug off mistakes the way a consumer app can.

Ecolab’s AJ Wijesinghe pointed to a different trap: cost. Put the “best model that’s out there” on high-volume work, he said, and the token costs can go “off the roof.” Sometimes, he said, that makes AI more expensive than having humans do the work. Ecolab then tuned the models and cut token costs by about 70% to 80%. The company uses both frontier and open-source systems, including Anthropic’s Claude and OpenAI’s models. Sometimes, as he put it, you just do not need the fastest car.

Neither executive sounded eager to sell a fully autonomous future. Venkatarayalu described a move toward a semi-autonomous world, with people still in the loop and trust built over time. Wijesinghe called Ecolab’s approach “human in the lead” and said agent tech is not mature enough to run at scale on its own. Honeywell is pushing a “see, think, act, and learn” framework that starts by mapping assets like HVAC, fire control, and security systems and tying them together over BACnet.

The physical world also moves slower than software. Venkatarayalu said over-the-air upgrades will come to commercial buildings, but not without an operator watching. He compared that with a Tesla that can sit for 1.5 hours for an update; a building, he said, cannot be taken offline like that. He also said Honeywell is fine-tuning open-source models because customers want sovereignty over their data and models, but warned that unsupported models without guardrails would be dangerous. Ecolab, meanwhile, is using sensors in dishwashers, pest traps and water systems to cut service visits and predict maintenance, while also handling water for chip production, power and cooling in AI infrastructure.

There was also a sharper business point underneath all this: the AI value is not in isolated tricks. Wijesinghe said single-use AI adds little at the enterprise level, and even vertical AI only goes so far. The bigger wins come from horizontal AI that works backward from an outcome across sales, finance and supply chain. Ecolab expects $325 million in annual run-rate savings by 2027, with “significant” amounts already in hand. The hard part, he said, is getting data, process readiness and cost discipline to line up. If one is heavier than the others, the value never shows up.