EDMONTON — To ensure the water that flows to residents of Drayton Valley, Alta., is clean and drinkable, the town’s treatment plant must constantly tweak the filters and chemicals that keep out unwanted bits and bacteria. Instead of humans twirling the knobs, it’s AI at the controls.  

Generative tools powered by large language models (LLMs) have attracted most of the attention and funding in the recent AI boom, with developers promising they will boost the productivity of office workers. But another group of AI startups is finding success in grittier conditions, selling technology to make sites like chemical manufacturing facilities and water treatment plants more efficient. 

“We’re creating the software for industrial automation,” said Martha White, CEO of RL Core Technologies, which developed the system used in Drayton Valley, a central Alberta town of about 8,000. The Edmonton-based startup specializes in a branch of AI called reinforcement learning that trains models using trial and error. White said the approach works well for sectors with “process control problems,” in which equipment needs constant tuning to produce a consistent product.

Sewage, wildfires and other environmental conditions all influence what flows into water and wastewater treatment plants. RL Core’s software takes in data from the facilities’ sensors and continually adjusts the quantity of cleaning chemicals added to the mix, as well as the frequency at which the filters are washed. Plant operators set the ranges of the changes to ensure it’s all safe, and the AI is run from on-site servers to maintain control and security.  

RL Core’s technology saves clients money by reducing the amount of energy and chemicals they use, as well as helping their filters and equipment last longer, White said. The company focused first on the water and wastewater sector because plants typically can’t afford automated control systems designed for fields like energy, which are very expensive but don’t generate large enough savings to justify the cost. RL Core will eventually expand into industries with similar needs, like food manufacturing or oil and gas. “The idea is to use this data-driven approach to improve operations,” White said. 

Toronto-headquartered Basetwo similarly employs AI to help clients in the chemicals, pharmaceutical and petroleum industries to reduce the waste they produce and the power they consume. “Making a uniform batch of high-quality stuff [is] a very hard thing to do,” said COO Thamjeeth Abdul Gaffoor. The consistency of a face cream, for example, can be affected by slight differences in the raw materials, or even by whether the factory floor is hotter or colder on a given day. 

Basetwo’s software lets clients like Johnson & Johnson and L’Oréal build models for their manufacturing processes without having to write much code. Engineers at the clients’ plants then consult Basetwo’s system for recommendations on temperature, pressure and other settings. The startup is working to give its platform the ability to control the equipment directly, so it can run on autopilot with minimal human supervision. 

Abdul Gaffoor said the startup’s advantage is in the way its technology combines historical data, equipment specifications and the fundamental laws of science. AI alone “doesn’t give you sensible answers,” he said. For example, it might suggest heating up a boiler all the way in just a few minutes, causing it to explode. Physics and chemistry don’t always work the same way in the real world as they do in textbooks, either.

Industrial facilities already have lots of sensors and data. Many clients can roll out new AI software quite quickly once they’ve identified the problem they need to solve and done some initial testing, both startup founders said.

Sectors like wastewater and chemical manufacturing do tend to be slower to adopt new technology than the professional services generative AI startups are targeting. The large corporations to which Basetwo sells do a lot of diligence and can take more than six months to finalize a purchase, Abdul Gaffoor said.

Industrial automation startups must still educate potential clients about how their products differ from generative tools. “Now, when people think AI, they always think LLMs or ChatGPT,” White said; RL Core’s models require much less data and power than generative AI does, and they tend not to hallucinate. 

While ChatGPT may be drawing all the attention, industrial automation startups aren’t struggling for capital or talent. Basetwo closed its US$11.5 million Series A last November, with backers including AXA Venture Partners, Deloitte Ventures and major Japanese industrial firms Shimadzu and Chiyoda. “Industrial AI is not GenAI,” Abdul Gaffoor said, noting the two fields attract different investors. 

RL Core raised US$5 million in a seed round the following month from U.S. investors TQ Ventures and Flying Fish Ventures. It has hired engineers who specifically want to work on real-world applications for reinforcement learning. “There are so many generative AI companies out there that we’re actually unique,” White said.