Colgate's machine-health AI paid back in 4 months at six plants

Colgate-Palmolive says machine-health AI covered its annual cost within four months at six Hill's plants. Its first try failed. What a smaller plant can copy.

Colgate-Palmolive says the machine-health AI it put on its production lines paid for its own annual cost within four months at each of six Hill's Pet Nutrition plants. The system is a set of wireless sensors read by Augury's software, which flags motors, gearboxes and pumps that are starting to fail. Warren Pruitt, who runs global engineering services at Colgate, gave the payback figure in 2021. Colgate has never published the price.

Money is flowing into startups that bolt sensors onto factory machines and promise to predict breakdowns. For a plant manager, the useful question is not who is raising what. It is whether anyone who has actually run one of these systems has put a payback in writing. Colgate has, and it also admits its first attempt failed.

What Colgate actually installed

Small wireless sensors sit on rotating equipment and measure three things: vibration, magnetic flux, which tracks how a motor draws power, and surface temperature. The readings go to the vendor's software, which compares each machine with more than 80,000 others on its platform and raises an alert when one starts to drift. Plant maintenance teams decide what to do with the alert.

Pruitt's own example, written in 2020, is a tube-making machine on a toothpaste line. The software flagged a rising temperature on its drive motor. Technicians found a fault in the motor's water cooling and fixed it in a planned stop. Colgate counts the save at 192 hours of downtime and 2.8 million tubes of toothpaste, plus 12,000 dollars for a motor it did not have to buy and 27,000 dollars in conversion costs.

The quieter change is to routine maintenance. Machines that used to come down for service about every six months as a matter of course can now run nine months or longer when the data shows they are healthy. That is fewer planned stops, not just fewer surprise ones.

One avoided breakdown on a machine with no spare paid for a lot of sensors.

The first pilot failed

Before this vendor, Colgate tried predictive maintenance with someone else, and it did not work. IndustryWeek's 2023 account says the sensors were more or less slapped onto equipment and the software was not good enough. Pruitt concedes his team did not ask the right questions at the start.

Two habits came out of it. Engineers now sit in on early calls with any new technology partner, and Colgate asks vendors for current customer references before a pilot. Both cost nothing and both apply to a plant of any size.

What changed, in numbers

The pilot ran at Colgate's oral-care plant in Morristown. Within three months it had raised two alerts that, by Colgate's count, avoided 60,000 dollars of downtime costs. The system then spread to Colgate's North American plants and to Hill's Pet Nutrition, the pet-food business, where six plants had it by 2021. By 2023 IndustryWeek listed it at four Hill's plants in the United States and plants in the Czech Republic and the Netherlands.

The spread came through a group of 50 to 60 engineers from 45 plants that meets every two weeks to share pilot results. Once a pilot worked at one site, the next sites adopted it on that evidence rather than running their own trial.

What did it cost?

Colgate has not said. The only cost figure it has given is a ratio: at six Hill's plants, savings covered the annual programme cost within four months. Our arithmetic, not Colgate's: if savings kept that pace, a year would return roughly three times the annual cost. They may not, because the first months catch the worst faults that were already building up.

What the numbers do not show

Every figure here is Colgate's own, chosen by the person who championed the system. There is no plant-wide downtime rate before and after, no count of machines covered and no false-alarm rate. The figures date from 2020 to 2023, and we found no newer number from the company. The 2020 article sits on the vendor's blog, and the payback figure reached print through a trade-press interview rather than a company filing. Neither makes it wrong, but neither is an audit. Read them as well-documented best cases from a company with 45 plants in its engineering network and a central engineering team, not as an average.

Does this apply to a 200-person plant?

Yes, on a smaller footprint. The Colgate result rests on a few machines whose failure stops a whole line for days and has no backup. Most mid-sized plants have a handful of those. A three-month pilot on five to ten such machines is the same test Colgate ran, at a size a single maintenance lead can manage.

What a smaller plant lacks is the network of 80,000 machines the vendor brings and the 45-plant community that spread the result. Buy the first, and replace the second with a written target agreed before the pilot starts.

What to do on Monday

List the machines that have no spare and have caused a stop of more than a day in the last two years. Put a price on an hour of downtime on each line, using lost output, overtime and scrap. Ask any sensor vendor for a current customer reference in your industry and for who, on your side, will read the alerts at night. Decide before the pilot who acts on an alert and how fast, because a warning nobody reads at two in the morning saves nothing. Then set a three-month pilot with a dollar figure it has to beat, and stop it if it does not.