Fluke Reliability Puts Large Language Models to the Test

Last year, I participated in a roundtable discussion on artificial intelligence at Fluke Reliability’s Thought Leadership Day. They invited me to play foil at the lead-up to their Xcelerate 2025 conference because I thought large language models (LLMs) such as ChatGPT and Claude and OpenAI and all the rest had nothing to offer manufacturing.

Parsing and summarizing text; content suggestions or from-whole-cloth authorship; and iterating on visual ideas: These are all general tools powered by LLMs that have nothing intrinsically to do with the plant floor.

I stand by my cynicism if we’re talking about LLMs as standalone tools. As components in larger software packages, however, I finally see a use case for LLMs that absolutely could benefit a lot of manufacturers.

Embedding AI in Maintenance Tools

Last week, at Xcelerate 2026, Fluke general manager Jay Hack presented an update to the company’s computerized maintenance management system (CMMS), eMaint. The new update, eMaint AI, adds four primary tools/functions.

The Ask tool allows technicians and operators to speak questions aloud, and eMaint AI, using natural language processing (NLP), parses the question and provides an answer.

The Build function parses maintenance manuals to create standard operating procedures (SOP) and shells of preventative maintenance work orders to give maintenance technicians head starts on the paperwork. Hack says the Build tool is between 80% and 90% correct. The data still requires validation by humans.

The Speak tool allows technicians to generate work orders through NLP. Imagine a technician replacing a battery, and they need to log the serial numbers of the new and old battery. The technician using Speak literally speaks the serial numbers aloud, and the software fills the data into the appropriate fields on the work order form.

My questions about the Speak function mostly related to noise on the shop floor and whether any device would pick up spoken words clearly enough to populate a form. “We have some iterations to do,” says Hack.

Finally, the Learn function provides more or less a translation machine for any maintenance manuals uploaded to the software. Results very much need validation. In one instance, the English phrase “machine was jamming” was translated by eMaint AI into Italian as “machine marmalade.”

The Learn function serves only “for basic contextual” use, says Hack. Technicians still have to approve the work procedure recommended by eMaint AI, which should catch translation errors.

“I think most established companies are probably doing similar things to what we are. I think more startups are probably pushing the boundaries a little bit more. But as a system of record, we have more to lose if we push ‘marmalade’ into all the fields that aren't supposed to be there, so we've been careful,” says Hack.

AI support in CMMS software itself is nothing groundbreaking. I can’t speak to any of those systems. But as the presentation went on, I had this nagging feeling I was seeing something that put LLMs in the “useful” column for manufacturers. Speeding up work orders and maintenance jobs for increased throughput and better OEE would provide long-term value…if eMaint AI functions as advertised.

When AI Models Cannot Fail

Vineet Thuvara, chief product officer at Fluke, says there is zero room for error when Fluke incorporates AI into its maintenance tools. It can’t afford for eMaint AI to “hallucinate,” or make up false data in response to queries.

“As we think about generative AI, the public space, in the consumer space, it's kind of a free-for-all. But as you move those models into industrial, especially in places like reliability where there is no room for error, when it's a voltage, it’s a voltage,” says Thuvara.

Fluke had a golden opportunity to improve the model that runs eMaint AI. Models need data on which to train. If the general eMaint AI model could train on all the collected data generated by all the pilot tests, the model might improve faster, which benefits everyone.

But the manufacturers running the pilots put the kaibosh on that idea. Everyone wants models trained solely on their own data, no one else’s. Manufacturers don’t want their data used to train models that could serve competitors.

Thuvara thinks that’s fair.

“I remember when I was at Amazon … I was on the Alexa team [working on] Ring. The idea was, can neighbors share a little bit of their internet bandwidth so you can track your entire neighborhood? For example, a pet is lost. Your pet crosses somebody else’s door. You can find the cat. It took us eight years to get there, for customers to say ‘Okay, I see the value here,’” says Thuvara.

That was just to get people to share front door camera feed data for the greater good. Imagine convincing manufacturers to share operational data. A better model provides a better result, but right now everyone wants to stand alone.

Time For Manufacturers to Prove Out AI

Fluke officially announced eMaint AI after pilot testing at manufacturers like SFK, which suggests the pilots went well. Once customers start talking, we’ll get a definitive answer, but I’m cautiously optimistic I’ve seen a legitimate use case for LLMs in manufacturing. I’m happy to be wrong about technology’s efficacy if it means we’re seeing new tools that genuinely help manufacturers succeed.

Fluke plans to roll out the eMaint AI beta update to eligible customers one at a time beginning in Q2.

Disclosure: I attended Xcelerate 2026 at the invitation of Fluke Reliability, which provided flight and accommodations.