Waters says its AI cut manual interventions from 90 to 4 in its own 400-chromatogram pilot

The expensive part of a tedious job is rarely the job. It is how often the job calls someone away. Think of each manual intervention in a drug-testing lab as a scientist called over to hold the ladder: they put down what they were doing, help with the fix, and then try to remember where they were.

Waters Corporation says its software can make those calls rarer. In a press release on October 28, 2026, Waters announced AI-Powered Peak Integration, a new premium subscription feature for Empower, the chromatography software it says has 500,000 users who "support the release of over 80% of medicines globally." The feature is trained on each laboratory's own data and makes the routine fixes itself, "removing noise peaks, combining high molecular weight peaks and editing baselines."

Those fixes are not trivial. Waters says "Manual peak integration is the largest source of variability in chromatographic results and one of the most time-consuming steps in data review." David Leitham, who runs Waters' informatics software business, told Fortune the hand edits are "frankly not a good use of the scientist's time or expertise."

The main figure comes from a 2025 pilot with one large pharmaceutical customer, which Waters does not name. In a side-by-side review of 400 chromatograms, current practice needed 90 manual interventions. With AI-powered peak integration, the same review needed 4. Waters calls that "a 95% reduction." Its footnote for the pilot reads, in full: "Waters internal pilot study with a global pharmaceutical partner, 2025. Data on file, Waters Corporation." Separately, Fortune reports that Waters "worked closely with an undisclosed, top-ten pharmaceutical company for about one-and-a-half years," much of it spent on legal agreements and on safe ways to share the data behind the first models. Neither source says whether that company ran the 400-chromatogram pilot.

So 90 to 4 is less a speed record than a count of interruptions. Across those 400 chromatograms, a person stepped in 4 times, not 90. The 4 matter as much as the 90. Analysts still review the model's "explainable outputs," and Waters says its "version-controlled, no-drift models maintain consistent performance until the laboratory chooses to retrain and approve a new model." A person still checks the result. What changes is how often that person is pulled away from the science.

A second claim comes from a separate study. Waters says user testing "showed 50% faster review and 72% lower user-to-user variability." Its footnote for that line reads: "Waters Immerse Research Lab customer evaluation, 2026. Data on file, Waters Corporation." The release does not say how many users took part or how review time was measured, and the 50% figure is not part of the 400-chromatogram pilot.

None of this is in routine use yet. Waters is "enrolling customers now for early access beginning in December 2026, with validated, compliant deployment to follow in mid-2027." In a regulated lab, the second date is the one that counts.

Waters pitches the tool as productivity, not as a way to do without people. Brook Colangelo, its chief information officer, told Fortune the aim is to "help unlock more human potential in the scientists that are using this technology." Tina Wu, a Waters senior vice president, said customers want what AI brings "without giving up control of their data or the compliant audit trail behind it." Whether other labs, on their own data, also get down to 4 is what early access should show.