The gains of using AI in R&D are substantial—research teams screen formulations faster, search the literature more quickly, and get help writing analysis code. Yet if you ask an R&D leader whether their last product generation took less time than the one before it, the answer is usually no. That’s changing with agentic AI.
Since late 2025, agents have become reliable enough for complex work like scientific R&D. They can stay with a problem across dozens of steps and correct their own errors, use tools autonomously and reliably, keep an entire research thread in view, and work together in coordinated teams. (See The Capability Inflection Point: Agentic AI in Scientific R&D for more on what’s changed.)
Now that AI agents have become capable of carrying complex work across stages, R&D organizations that put them to work can change two main things: how long each product cycle takes, and how far ahead each new product cycle starts.
A new material or formulation passes through several stages before it reaches the plant: discovery, formulation, testing, scale-up, and the handoff to manufacturing. Each stage involves different people, different data systems, different models, and different software. Time is lost to the transitions between stages: waiting for results to be analyzed, earlier data to be tracked down, or decisions to be explained to the next team.
AI use in this setting is broad but mostly task by task. At the task level, the results can be substantial. Mitsui Chemicals built an in-house AI agent1 that reads chemical structures in research literature and organizes the compound information it finds. A review that used to take about a month now takes about a day. PPG uses AI to narrow formulation candidates2 before lab testing, which the company says significantly shortened time to market for a new automotive clearcoat.
These are meaningful improvements, and each one speeds up a single stage. These tools move the team to a new baseline, and progress holds steady there. The handoffs between stages still work the way they always have. Finding and reusing that knowledge takes real effort, so each new generation pays that cost again instead of building on what the last one learned.
When agentic capabilities are applied within an R&D process, time saved comes from the gaps between the many steps, both within stages and across.
At the start of a cycle, an agent can gather what the organization already knows: earlier experiments, their results, and the reasons behind past decisions. That step could potentially run 10x to even 100x faster than gathering it by hand. During design, experiment plans can draw on everything tried before, including what failed. During analysis, results can be processed as they arrive, so the next round can be proposed while the current one is still fresh. Scientists still decide what to run and define what good is, and the reasoning behind those decisions is recorded as the work happens. When the work moves to scale-up, the manufacturing team receives the full history along with the recipe.
The stages themselves stay the same. The waiting, searching, and re-explaining between them is what gets shorter.
The second effect takes longer to show up, and in our view it matters more. Because the agents took part in the whole cycle, what the team learned doesn't disappear when the project ends. The next generation starts with a record of what was tried, what failed and why, and what worked. In a conventional workflow, each cycle starts from roughly the same place. In an agentic workflow, each cycle starts further ahead than the last, and with more to build on, it takes less time.
Each R&D cycle starts with more of what earlier cycles learned, and each one takes less time.
There are early signs that this kind of advantage grows quickly. OpenAI tracks how deeply its enterprise customers use AI, measured in output tokens. Output tokens are not an indicator of value, but they are a good proxy for the depth of AI use within a company. In professional, scientific, and technical services, the gap between frontier firms and typical firms grew from 2.6x in January 2026 to 7.7x in June.3 The frontier firms' use began climbing at the end of 2025, around the time agents became practical for complex work.
The gap in AI use between frontier and typical firms grew from 2.6x to 7.7x in the first half of 2026. Source: OpenAI, Signals: Enterprise Data, August 2026.
For the business, shorter cycles that build on each other mean more product generations in the same period and earlier entry into markets. They also mean scientific knowledge stays with the organization when people move on, and scientists spend more of their time on new problems.
Where to start
Because each cycle builds on the one before it, organizations that start earlier gain more from each cycle that follows. Starting doesn't require changing everything at once. A practical first phase usually looks like this:
Frontier models and off-the-shelf platforms will always be part of your solution set. The larger gains, however, come from an agentic R&D system intentionally designed and built to understand your chemistry, your instruments and data, and the decisions your team has already made.
If your team is thinking through a more agentic R&D process, we welcome the conversation. Reach out to Enthought here.
Sources