A genetics researcher in Princeton described a small, stubborn problem. When you clean single-cell data, you have to decide where to cut off the cells with too few genes. “It’s sometimes a little bit hard to eyeball exactly where you should cut off,” they said. So they had an AI coding tool find the threshold, and it did, “better than I believe the human eye could do.”
A minute later, the same researcher named their biggest worry about AI: colleagues who “blindly use” those tools and “don’t check the code or make sure that it’s actually correct.”
They trusted the tool more than their own eye, and they worried about everyone else trusting it too. That tension ran through every conversation we had that day.
A phone booth at a conference about practice
BioNJ’s fourth AI & Digital Health Day, held October 23 at Princeton University, was titled “From Pilot to Practice: AI at Work in Life Sciences.” Our CEO, Michael Ioffe, spoke on a panel. We also brought a phone booth. Attendees picked up the handset and talked with Arist’s AI interviewer about how they use AI, what worries them and what they are watching next.
Six people picked up and five talked, for between one and seven minutes. That is a small sample, and we treat it as one. But the conference title held up. All five already use AI at work.
One sales leader has each rep record a couple of minutes every week on what worked and what they will try next, then runs AI over the recordings to find who needs help. One person builds storyboards, slides and animation, and their rule of thumb is that AI “can be used for anything.” Another drafts client proposals with it. A program manager used it to rescue a data warehousing project that had fallen three months behind.
The draft is the easy part
Two people described the same pattern. The storyboard builder handed AI notes written “in a sort of disorganized way” and got back a complete script and storyboard. “It wasn’t perfect. It needed a lot of tweaking and changes,” they said. It still got them to a first draft “much faster.”
The proposal writer gave AI an outline, a slide count, case studies and an earlier deck to work from. Then they went through the result side by side, added their own ideas and passed it to the team “to get some other humans to look at it.”
Neither treated the draft as finished. Larger studies suggest that is the right instinct. BetterUp Labs and Stanford found that 40% of workers had received AI “workslop” in the past month, and each piece took close to two hours to fix. The time AI saves on a draft gets spent somewhere. It lands on whoever reviews it.
Checking is where it goes wrong
The researcher’s worry has numbers behind it. In a study of 48,000 people, the University of Melbourne and KPMG found that 66% rely on AI output without checking it.
The program manager described the cost from the other side. Some companies will try AI where it does not fit, they said, and end up “thinking that AI, quote, unquote, ‘doesn’t work,’ when actually the application was the issue.” Without a check, nobody can tell which went wrong: the tool, or the way it was used.
The decisions nobody had made
The program manager’s own story was the most telling of the day. They were brought in “to kind of solve the problem and put out the fire” on the data project, with no extra budget. One of the first things they asked AI was: “What are the most significant issues that have been left on the table and no decision has been made?”
The answer was a handful of stalled decisions, “the elephant in the room,” as they put it. Two weeks of analysis produced three resolutions. The team adopted them, along with a daily habit of not letting issues sit unaddressed.
The AI did not fix the project. It helped someone find out why it had stopped.
The conversation leaders have in private
The last conversation was the least comfortable. An attendee who works with life sciences clients said job loss comes up in “most conversations” with senior leaders. Privately, they said, the intent is “to cut costs by removing headcount.”
That is one person’s account, not a finding. It sits awkwardly next to Deloitte’s 2026 research, which found that 89% of life sciences and health care organizations have not redesigned jobs around AI. Decisions about how many people a team needs are being made before anyone has looked closely at what those people do.
When we asked what this attendee came to the event to learn, the answer was “the intersection of the technology, the data, and the people”: how organizations find the right uses and bring their people along.
Practice is everything around the tool
“From pilot to practice” usually means more licenses and more use cases. The five people we talked to were already there. What they were working out was everything around the tool: who reviews the draft, who checks the threshold, which decisions are stuck, which jobs change.
None of that shows up in a usage report. We learned it the way the program manager rescued their project, by asking the people closest to the work. A few days later at Gartner HR Symposium/Xpo in Orlando, HR leaders told us a version of the same story.
Based on five anonymous conversations at Arist’s phone booth at BioNJ AI & Digital Health Day, October 23, 2026. Not every person answered every question. These are informal observations, not a representative survey. Quotes are lightly edited to remove filler words.
