Stage-Gate® has been the default product development process in life sciences tools for decades. Most teams are running close to the original version, built in an era before agile, before AI, before anyone was scoring sustainability at a gate review.
Robert Cooper, the originator of Stage-Gate (I use phase and milestones in my blog posts), recently mapped out eight advances that define what he calls the 5th generation of the model. I'm covering all eight below, with two of them (agentic AI and treating AI adoption as its own NPD project) grouped together since they share the same adoption story. From conversations at ELRIG Drug Discovery and SLAS, plus my own client meetings, here's where the evidence is solid, and where I genuinely don't know yet.
1. Agile: the adoption gap is real and measurable
Agile development breaks a project into short, iterative sprints with frequent customer feedback, instead of locking in a full plan before work starts. An agile-stage-gate hybrid keeps Cooper's stage-and-gate structure but runs agile sprints inside each stage. It's one of the most documented of Cooper's eight advances, and the data backs up what I keep hearing in client conversations: healthcare and pharma sit at roughly 8 percent agile adoption, compared to 27 percent in technology and 18 percent in financial services. Agile is a term life sciences teams have likely heard, especially applied to software projects, not something most have built into how they run hardware or reagents projects. On a hardware project, a sprint usually ends with what Cooper calls a minimum viable prototype, not a finished part, but a version solid enough to put in front of a customer or test on the bench, so the team gets real feedback every two to four weeks instead of waiting for the next scheduled milestone. Across six manufacturing case studies, including LEGO and Tetra Pak, firms running this hybrid reported roughly a 30 percent cut in time to market and a 30 percent gain in development productivity.
2. AI-powered stages: adoption lags the hype, but the hardware use cases are concrete
Cooper's framework describes AI assisting with everything from idea screening to concept testing to continuous post-launch feedback. For instrument Product Managers (PMs) specifically, that's not abstract. Chromatography vendors are already using AI to predict optimal gradient and temperature conditions from historical run data instead of relying on trial and error, and to flag baseline noise or abnormal retention times before a bad result gets reported. One chromatography and mass spectrometry software lead described a deliberately cautious rollout: keeping a person in the loop on critical steps like data acquisition for now, while moving faster on lower-risk areas like peak integration and report generation. One published study used this kind of AI-assisted approach to predict molecular structure from chromatography data with about 70 percent accuracy on compounds that had no reference standard to compare against. On the maintenance side, one predictive maintenance system for MRI machines is reported to have cut unplanned downtime by up to 60 percent and service requests by 35 percent across more than 1,500 sites, a concrete uptime number tied to one instrument category rather than an industry-wide percentage.
Even with concrete use cases like these, adoption is still early. A 2025 survey of 400 life sciences leaders found that while 96 percent believe AI agents will be essential within two years, lack of trusted data and unclear change management plans are still holding most teams back from using it today. That gap between where the technology already works and where most teams have deployed it is the one I keep running into in client conversations.
3. Agentic AI and AI-as-its-own-project: ahead of the curve, not behind it
These are the two advances with the least adoption (that I'm aware of in the life sciences tools sector) and the most talk. Agentic AI, systems that interpret a situation and act with minimal human guidance, is still mostly theoretical for physical-product development outside pharma's clinical and compliance functions, and I don't have a documented life sciences tools example of it in new product development (NPD) yet. One real exception is early concept screening, using an AI representation of the buyer built from real customer data, voice of customer interviews, sales conversations, not a generic AI simulation, to react to a new concept the way an actual buyer would before it reaches them. That's not a replacement for voice of customer work; it's a way to surface the weak ideas faster, so the limited time a PM has with real customers goes to the concepts that already cleared a first bar. In practice, this can mean introducing a set of synthetic buyer archetypes, built from secondary market reports, dozens of voice-of-customer transcripts, published literature, and other public data, as early as the first gate, so PM and R&D teams can pressure-test a concept or run a persona panel debate before committing real resources to it, cutting down on rework later in the process. Cooper also treats AI adoption itself as its own NPD project, run through a formal go/kill process rather than just used inside an existing stage. That one, I haven't seen attempted in life sciences.
4. Parallel processing: the model already exists in life sciences
This one has the clearest proof of concept of all eight, even though it didn't originate in life sciences tools. The COVID-19 vaccine programmes ran clinical trials, regulatory review, and manufacturing scale-up at the same time instead of in sequence, compressing what would normally be a decade of development into under a year. That worked because each workstream moved forward as soon as its own risk was acceptable, not when a master schedule said it was time. Most RUO and diagnostics development still runs these workstreams in series.
5. Eco Stage-Gate: I looked, and I came up empty
Eco Stage-Gate is the advance I have the least confidence writing about. Cooper describes green scorecards at gates and lifecycle assessment built into early concept work. I searched for a life sciences tools example, a company that scores environmental impact at a gate review the way it scores technical risk, and found nothing specific to this sector. That's either because it isn't happening yet, or because it's happening quietly and hasn't been written up.
6. Tailoring process rigour to project risk
Cooper's own framework states it plainly: not every project justifies the full five-stage process. Smaller, lower-risk initiatives like product improvements and customer specials need a lighter version of the same governance, not the same process at full weight. This is the advance with the most direct upside for teams stretched thin on gatekeeper time, and the one I'll go deep on in the next blog. Cooper's own portfolio management framework makes the connection explicit: the goal of tailoring rigour to risk isn't just a faster review for any one project, it's freeing up gatekeeper time across the whole portfolio so the projects that genuinely carry open risk get the scrutiny they need, instead of every project competing equally for the same fixed pool of review time.
7. The risk of staying put
There's a reasonable objection to all of this: stage-gate has worked for forty years, so why mess with a process that isn't broken. The honest answer is that tight adherence to the traditional five-stage model carries its own risk: not the risk of using an unproven new method, but the risk of running every project through the same expensive, slow process regardless of how much uncertainty it actually carries. That cost stays invisible until you compare it to a competitor who isn't paying it.
Most life sciences tools companies are running some version of generation one or two stage-gate, with agile and AI sitting at the edges of the conversation rather than inside the process. The advance most worth your attention first is the sixth one: knowing when a project doesn't need the full process at all. That's specific enough to act on this quarter, and it's where I'm headed in the next blog. Two questions are still open for me: if you've used an agentic system in NPD, run an "adopt this AI tool" initiative through formal gates, or scored a gate review on environmental impact, please share with me what it looked like.
Frequently asked questions
I have limited gatekeeper time. Which of the eight advances should I start with?
Start with tailoring process rigour to project risk. It is the one advance you can act on this quarter without new tools or a culture change, and it addresses the problem you already feel: every project competing for the same fixed pool of review time. Decide which projects genuinely carry open risk, then route the product improvements and customer specials through a lighter version of the same governance. That frees gatekeeper attention for the launches where a wrong call is expensive to reverse.
How does an agile sprint work on a reagents or instrument project?
A sprint on a physical product ends in what Cooper calls a minimum viable prototype. That is a version solid enough to put in front of a customer or test on the bench, well short of a finished part. You get real feedback every two to four weeks, well before the next scheduled milestone. You keep the stage and gate structure and run the sprints inside each stage, so governance stays intact while the team learns faster.
Can an AI buyer representation stand in for voice of customer at an early gate?
Use it for direction and keep humans for the decision. A buyer representation built from real customer data, voice of customer interviews and sales conversations can react to a concept the way an actual buyer would, which lets you surface the weak ideas before they reach a real customer. The value is sharper triage, so the limited time a product manager gets with customers goes to the concepts that already cleared a first bar. It earns its place at gate one when it is grounded in specific buyer evidence and current enough to challenge an assumption, and it loses that place the moment the output reads like generic market language.
