Three areas where AI is creating value in process chemistry
Process chemistry is probably not the first field that comes to mind when people think about generative artificial intelligence (AI). When these tools first became widely available, as a team we explored where, if anywhere, they could be used responsibly within process development.
What became clear quickly was that AI is best viewed as a tool for tackling one of the biggest challenges in pharmaceutical development: doing more in less time. As molecules progress through development, timelines tighten, budgets come under increased pressure, and the cost of getting things wrong compounds rapidly with scale. Used appropriately, AI can help relieve some of these pressures without compromising scientific integrity.
At Sterling, that means giving our scientists more time to solve problems, make decisions faster, and identify opportunities to improve processes by removing tasks that don’t require scientific expertise. Ultimately, helping us deliver better outcomes for our customers.
These examples reflect what we’ve learned so far at Sterling and are drawn from real projects to illustrate both the opportunities and the limitations of these tools.
Starting with the right question
Today, the question is not whether to use AI in drug development, but where it can genuinely add value.
That distinction matters in process chemistry because the risk is not simply generating a poor answer. It’s the impact of acting on poor insight in a safety-critical environment, with complex datasets and little room for error, especially at scale. Understanding where these tools are useful is therefore more important than trying to apply them everywhere.
As such, we’ve found it helpful to think about AI integration across three key domains:
- Streamlining the ordinary: work that must be done but is not where chemists can create the most value.
- Eliminating grey work: tasks that consume time and attention without contributing directly to the science.
- Augmenting critical problem solving: where AI can help expand the solution space and keep analysis moving while chemists focus on experimentation.
Across these three areas, our experience at Sterling suggests that routine AI use could save around 8-10 hours per chemist each week. Effectively allowing work that previously took five days to be completed in four.
More importantly, those gains allow scientists to spend more time solving the problems that matter most to customers.
1. Eliminating grey work
Grey work is the category that is hardest to justify as a good use of a chemist’s time. It includes tasks that are necessary to keep projects moving, but contribute little to the science itself, such as searching through email chains or reconciling spreadsheets for granules of information.
Unlike routine scientific tasks, grey work isn’t something to be streamlined so much as eliminated wherever possible. AI is particularly effective at compressing these low-value activities. While the individual time savings from any single task may be modest, they accumulate quickly.
An hour recovered from administrative work is an hour that can be spent designing experiments, analyzing results, or troubleshooting problems in the lab. Scientific attention is often the limiting resource in process development, so the real value lies in ensuring that chemists spend their time where their expertise creates the greatest impact.
2. Streamlining the ordinary
Every process chemistry program generates work that is required but doesn’t necessarily represent the central intellectual challenge of the project. Development reports, literature reviews and safety assessments all matter, but they are not where scientists create the most value. This is an area where AI has proved genuinely useful.
Literature reviews, for example, can be accelerated significantly. AI is particularly effective at organizing and summarizing published work when evaluating a new transformation or preparing a safety dossier. References still need to be verified and cross-checked for potential hallucinations, but the time spent manually searching and collating information can be reduced considerably.
Safety assessments require greater care. By feeding material data and reaction conditions into a secure environment, it is possible to generate useful hazard summaries without exposing full structural intellectual property (IP). These outputs provide a valuable starting point, however they are not substitutes for formal process safety evaluation.
Recognizing this, at Sterling we continue to rely on validated systems and independent human review for all calculations and decisions that could affect process safety, product quality, or manufacturing success.
3. Augmenting critical problem solving
This is where the strongest case for AI can be made, and where it is most important to be realistic about the limitations of these tools.
Although AI cannot solve chemistry problems independently, it can expand the solution space, surface ideas that might otherwise be missed, and help maintain momentum while chemists focus on experimental work.
One recent project illustrated this clearly.
Case study: Salt formation and process optimization in a multi-step synthesis
The challenge arose in the final step of a sixteen-step synthesis, where purification relied on large-scale column chromatography. Approximately 50% of the product was lost to impurity-containing fractions, at an estimated manufacturing cost of approximately $390,000 per kilogram (kg) of final product.
To evaluate the potential use of AI in a situation like this, we provided a firewalled-AI system with key functional group information and a curated set of crystallization papers, prompting the agent to explore alternative isolation strategies. The model identified a potentially protonatable functional group, a tertiary amine, that may be leveraged to generate a crystallizable salt. In addition, a small set of organic acids as proton donors was recommended for initial screening.
Initial small scale crystallization experiments were performed, screening these organic acids, of which tartaric and citric acid showed promising results. However, scale-up experiments produced a gelatinous semi-solid that could not be filtered effectively. At this stage, without AI, it would be typical to return to the original chromatographic purification approach to avoid any unnecessary delays to the project timeline.
Instead, the results were fed back into the model, which suggested screening moderately polar aprotic anti-solvents to break the water shell of the gelatinates semi-solid, which it suggested was likely to be a swollen crystalline hydrate. A subsequent solvent screen identified acetone as a hit, allowing the citrate salt to be isolated as a free-flowing and filterable white solid. The material was isolated in 65% yield, setting the stage for a process without the need for column chromatography.
Cost modeling suggests that the revised route has the potential to reduce manufacturing costs by $90,000 per kg, with further process development potentially delivering approximately $130,000 in savings. These figures rely on assumptions around scale-up and future optimization, but they illustrate the magnitude of the opportunities still available to capitalize on.
AI did not solve the problem, as the chemistry still required experimentation, interpretation, and expert human judgement. However, AI did contribute a fresh perspective that kept a promising line of enquiry alive beyond the point where instinct alone might have abandoned it. Unlocking this kind of value depends on more than the model itself. It requires the right infrastructure, controls, and governance to ensure AI can be used safely and responsibly.
The infrastructure behind the results
At Sterling, we recognize the importance of having the right safeguards in place. We ensure our AI tools operate within secure environments, keeping customer information safe and separate from consumer-facing platforms. Where necessary, molecules can also be described by their functional groups rather than full structures without significantly reducing the value of the interaction.
Verification is equally important, with all safety-related outputs independently reviewed.
AI adoption at Sterling is not confined to a single team. Our process chemists, analytical scientists, engineers, and other functions are all involved, allowing best practices and lessons learned to spread across the organization.
The tools will continue to evolve, and so will the way we use them. But the foundations around security, verification, and training are already in place.
Delivering better outcomes at Sterling
The opportunity now is not simply adopting AI, but integrating it in ways that strengthen scientific decision-making, improve efficiency, and support better outcomes for customers.
Our experience shows that AI delivers the greatest value when it complements scientific expertise, enabling our teams to focus more of their time on solving complex technical challenges and improving manufacturing processes before scale-up.
Used properly, AI can streamline routine work, remove administrative friction and help sustain promising lines of enquiry that might otherwise be overlooked. However, it does not replace the need for the scientific judgement that ultimately determines whether a process succeeds. That remains the responsibility of the scientists.
At Sterling, our multidisciplinary teams combine deep process chemistry expertise with capabilities spanning preclinical development through to commercial manufacture. This breadth of experience allows us to consider the full range of factors affecting quality, yield, efficiency and safety throughout the API lifecycle, and to apply the scientific rigor needed to address them with confidence.
AI is simply another tool that helps us direct more of that expertise where it matters most.
If you’d like to learn more about how Sterling can support your molecule through process development and beyond, speak to an expert or visit our Knowledge Hub for more resources.
