Every R&D organization has a version of the same problem: the people who know the most are often the closest to leaving. The engineer who can look at a scan and know instantly what's wrong. The scientist who's run the same class of experiment for twenty years and can sense a bad batch before the data confirms it. When they leave the company or retire, that judgment doesn't get handed off in a knowledge-transfer meeting. It just leaves.
Agentic AI is some of the first technology that gives organizations a real way to keep it, not by replacing that judgment, but by encoding it into a system that gets more reliable every time a person corrects it. At Enthought, we’re actively working with our customers on building agentic AI systems that do exactly this, and what we've learned changes how we think R&D organizations should approach this shift.
Most people's first experience with AI was something like ChatGPT. It's impressive, but still just a resource you consult while you do the actual work. That changed around late 2024, when models like Claude began demonstrating something different. Instead of just answering, they could break a problem apart, spin up multiple agents, and work it together. That shift, from assistant to orchestrator, is what “agentic AI” means in practice, and it's what makes real knowledge capture possible in a way a chatbot never was.
Enthought's CEO, Eric Jones, PhD, notes his own turning point was when he was trying to convert a pulse-modulated signal into an RC circuit. The model didn't just answer the question; it wrote the Python tool needed to solve it. "To have a tool create a tool is the most amazing thing I've ever seen," he says.
Strip away the hype and an agent comes down to three things. A large language model, a set of tools it can call, and a loop that lets it decide what to do next. Through that loop, it reads a result, chooses the next action, and repeats, without someone directing each step. The model supplies raw reasoning; the tools let it do something real. But raw reasoning isn't organizational experience. Agents are brilliant PhD-level interns that know an incredible amount of knowledge, great reasoning capability, and zero work experience. The gap between what an agent already knows and what it needs to know is exactly where expert judgment has to go.
Closing that gap takes data the agent can reach, tools built for your actual systems rather than borrowed off the shelf, and instructions that spell out how work actually gets done at your company. That third piece is interesting, because it's the first time expert judgment has to get written down instead of staying in someone's head. A seasoned engineer's sense of what a bad image looks like, once it's encoded into an agent's instructions, doesn't retire with them.
That's where the real value sits. In practice it means building several specialized agents instead of one generalist trying to do everything: a reviewer agent that catches what the first agent missed, a formatting agent that's reliably good at exactly one job.
A Concrete Example: Semiconductor Metrology
A semiconductor fab takes SEM (Scanning Electron Microscope) images of wafer patterns at the nanometer scale. Reading them well has traditionally meant an expert at a workstation, running custom scripts, making judgment calls built up over years. Put that into an agentic workflow and it looks like this: a monitoring agent watches for new images and pulls the metadata, an analysis agent built with image-processing tools calibrated to that fab's process runs the read, and a review agent instructed with the same tolerance logic the expert used flags anomalies before a formatting agent packages the results into the standard report.
A human still reviews and approves before anything reaches the customer. The system still gets things wrong sometimes, and catching that is the point of keeping a person in the loop. What's different is that the judgment calls built into the pipeline came from an actual expert's process, documented and encoded rather than left in someone's head. Nobody had to hunt down files, remember which script to run, or reformat the output by hand. And the next time a reviewer corrects something the system got wrong, that correction becomes part of the system too.
Capturing Expert Knowledge
That correction loop is the part most organizations underrate. Every organization has people whose judgment took decades to build: what a good result looks like, what pattern means trouble, what shortcut is actually safe. When those people leave, that judgment has always left with them, and whoever inherits the role starts over. Agentic AI changes that math. Once an expert's process is built into agent instructions and tools, it stays, and it keeps improving every time a reviewer corrects it.
That's not just a tool one scientist or lab uses. It's a system that accumulates the R&D organization's knowledge, so the company gets smarter even as the people in it change.
Getting Started
The mistake most organizations make is waiting for the perfect starting point. This isn't a five-year build, it's something you can now get running in months. Pick one high-value workflow with a known expert, document their process, build the tools it needs, and get something running. It won't be perfect the first time. It'll need correcting, probably more than once. What's actually at stake isn't convenience. It's whether the expertise your best people built over a career is still there after they're gone.
This post is adapted from a keynote presentation by Eric Jones, CEO of Enthought, delivered at the Enthought’s 2026 R&D Innovation Summit.