AI in statistical programming: better work or fewer jobs?

AI in statistical programming can speed up routine coding and drafting, creating room for more demanding scientific work. Turning that possibility into better jobs depends on what employers do next. They can use the extra capacity to strengthen research, or look for ways to produce the same outputs with fewer people.

That choice gave urgency to discussions at the PHUSE Single Day Event in Basel on 17 September 2026, which entimo attended. Held at the Novartis Campus, the event explored Reimagining the Clinical Data Scientist Role: Future-Proofing Skills and Partnerships. Practical demonstrations sat alongside questions about professional skills, accountability and the organisation of work.^[1]

Pharma speakers were broadly optimistic. They described AI taking on repetitive activities while statisticians and programmers spent more time on problems requiring experience, creativity and judgment. Some audience questions were harder to settle. Where would the jobs be as more work became automated? And if a person still had to check and approve the result, how much time had really been saved?

The answers are connected. Before companies can decide what to do with greater productivity, they need to understand what AI actually delivers—and what people must still contribute.

The symposium exceeded every expectation, both in content and in the quality of professional engagement.

AI can provide a faster starting point for work that still needs professional judgment. It can help draft code and analysis documents, giving specialists something to examine and improve. Whether that saves time overall depends on the quality of the draft and the effort required to bring it to completion.

There is evidence of the opportunity in software development. A study of the effects of AI on skilled work published in Management Science in February 2026 combined three randomised workplace experiments involving 4,867 developers. It estimated a 26% increase in completed tasks among those using an AI coding assistant. Results varied across the experiments. The study measured general software development, so its findings cannot establish equivalent gains in regulated pharmaceutical work or show that fewer employees are needed. ^[2]

Clinical research offers a more targeted example. Pfizer researchers evaluated an AI system across 71 statistical analysis plans. Their 2026 Clinical Trials paper reported average first-draft generation times of 1.0–3.4 minutes, compared with a conventional first-drafting benchmark of one to two days. These were initial drafts. The study did not establish that reviewed and approved plans could be completed in minutes. ^[3]

For a statistician, the attraction is easy to see. Time spent assembling an initial plan could be used to challenge an assumption or discuss an uncertain endpoint. But that opportunity begins only when the draft is useful enough to work with.

This is where the Basel discussion about human approval becomes central to the productivity question.

Human review is needed because AI can produce an output that looks convincing while missing something important to the study. That does not make assistance pointless. Its value depends on the total effort needed to reach a dependable result, including the time spent checking, correcting and validating it.

Johnson & Johnson’s Guillaume Abgrall, Principal II Statistical Programming Lead, and Tadeusz Lewandowski, Head of Data Engineering, demonstrated this approach in Basel. Their R-code generation workflow used reporting specifications and Analysis Data Model (ADaM) metadata. Human checkpoints covered specification interpretation, variable analysis, implementation planning and code review. Their presentation stated the principle plainly:

A separate example showed why those checks matter. Using simulated data, Novartis’s Tiina Kirsilä demonstrated an application that worked and passed its tests. Its laboratory charts nevertheless used inconsistent vertical scales across patients. Recognising the risk of a misleading display required knowledge that the automated checks had missed. ^[5]

A good draft can remove hours of routine production. A convincing mistake can consume those hours again. Someone must understand what an analysis needs to show and whether the output represents it properly.

The regulatory direction reinforces that responsibility. In January 2026, the US Food and Drug Administration (FDA) and European Medicines Agency (EMA) published joint principles for good AI practice in drug development. These call for assessment proportionate to risk, traceable information and evaluation of the system as a whole, including interactions between people and AI. Faster generation therefore has to be considered alongside the evidence needed to trust and use the result. ^[6]

The case for human expertise is strong. It still leaves the audience’s employment question unanswered.

AI does not dictate a single employment outcome. A pharmaceutical company could use the same team to run more studies or explore more analytical questions. It could also retain experienced reviewers while reducing recruitment. When the technology creates additional capacity, management decisions help determine who benefits from it.

Evidence from the wider labor market explains why opportunities for people entering the profession deserve attention. A Stanford working paper revised in August 2026 examined US employment among workers aged 22–25 in occupations exposed to AI. Employment stood 19% below the level it would have reached had it followed the trend among less-exposed peers. The difference arose mainly through weaker hiring. The researchers found no evidence of widespread, economy-wide displacement and cautioned that the pattern did not establish that AI caused it. ^[7]

A different setting produced a different picture. A Danish working paper dated March 2026 used records through the end of 2024. It found no differential employment or hiring changes at workplaces encouraging chatbot adoption, including among younger workers. Work was being reorganized without corresponding reductions in employment. These studies cover different places and periods, and neither measures pharmaceutical statistical programming specifically. ^[8]

For pharmaceutical employers, that uncertainty makes their own intentions more important to explain. Employees should be able to understand whether greater capacity will support more research, change the composition of teams or reduce opportunities to enter the profession.

A company considering fewer junior appointments also needs to ask where its future reviewers will come from. If AI produces the first implementation, a new programmer still needs opportunities to understand and question it. They need to learn why a seemingly minor decision can change an analysis.

The hopeful future described in Basel depends on preserving that expertise and giving it more room to contribute. Supporting it also places a responsibility on the companies building the tools.

entimo’s approach is to build AI around the people responsible for regulated work. Automation should make routine tasks easier and leave more room for analysis and creative problem-solving, while keeping human judgment and approval meaningful. That is the balance the company is pursuing as it carefully introduces AI into its entimICE FastTrack platform.

The aim is a workflow in which users can understand what AI contributed, how the result was checked and who approved it. Human review, validation and traceability give substance to that responsibility. They help teams assess quality as production becomes faster. ^[6]

This reflects the emphasis on human expertise heard at PHUSE Basel. Faster delivery should strengthen what people can achieve while preserving their responsibility for the work. Employee development therefore belongs in the implementation plan from the start.

These principles are built into entimICE FastTrack, our platform for traceable, audit-ready clinical data analysis, and are part of everything entimo offers.

Statistical programmers need strong coding, statistical understanding and the ability to assess AI-generated work. Clinical knowledge, validation and communication are equally important. For entimo, the direction suggested by PHUSE Basel is a broader role: helping define the work, establishing confidence in the result and explaining its significance to colleagues. ^[1]

That starts with understanding code well enough to read, test and debug it, including code generated by AI. Depending on the working environment, programmers will need to maintain or deepen their expertise in SAS, R or Python. Clinical data standards are part of the same foundation. A programmer must understand how a dataset was derived, why a population was selected and whether the output answers the intended question.

AI literacy builds on that knowledge. It involves giving precise instructions, recognising when an answer rests on an unsupported assumption and designing a check that could expose the problem. The Johnson & Johnson workflow made these responsibilities visible at several points in the task, rather than leaving them all to a final reviewer. ^[4]

Technical judgment also has to be communicated. A programmer who identifies an ambiguous specification needs to explain why it matters and work with statisticians or clinical colleagues to resolve it. Cytel’s Hannes Engberg Räder made a related point for trial statisticians. His examples covered AI support with protocols, code and inconsistencies between study documents; scientific judgment and leadership of discussions remained with the statistician. ^[9]

That is where the promise of greater creativity becomes practical. It might mean finding a clearer way to present a result, improving an analysis or fixing a recurring problem in how data move through a study. Kirsilä’s presentation connected the satisfaction of programming with problem-solving, ownership and building something worthwhile. Those contributions can remain valuable as the mechanics of writing code change. ^[5]

For entimo, this is why investment in AI should go hand in hand with investment in employees. Experienced programmers need time to learn new tools and broaden their responsibilities. Junior colleagues need supervised work that teaches them why an analysis is trustworthy.

The Johnson & Johnson presentation offered one practical example: reviewing generated R code had become an unexpected learning opportunity for programmers unfamiliar with the tools. ^[4]

A promise of greater creativity will feel thin if every hour saved is immediately filled with another production target. Employees should be able to see what they will have the opportunity to learn and which responsibilities they can grow into.

When AI saves a programmer a day, someone decides what fills it. Giving that person the opportunity to investigate a difficult question, improve an analysis or develop a better way of working would make the promise of greater creativity tangible.

That is the promise pharma should be asked to keep.

All links in the reference list open in a new tab.

1. PHUSE. Basel, Switzerland — Single Day Event: programme and presentation archive. Event held 17 September 2026. Accessed 8 October 2026.
2. Cui KZ, Demirer M, Jaffe S, et al. The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers. Management Science. Published online 27 February 2026. doi:10.1287/mnsc.2025.00535.
3. Landman R, Emir B, Zhang R, et al. Leveraging generative AI to transform statistical analysis plan authoring in clinical trials. Clinical Trials. 2026;23(4):383–392. doi:10.1177/17407745261422365.
4. Abgrall G, Lewandowski T. Bringing Mock Shells to Life (PDF). Presentation at the PHUSE Single Day Event, Basel, Switzerland. 17 September 2026.
5. Kirsilä T. The Best Part of Programming Was Never Typing Code (PDF). Presentation at the PHUSE Single Day Event, Basel, Switzerland. 17 September 2026.
6. US Food and Drug Administration; European Medicines Agency. Guiding Principles of Good AI Practice in Drug Development (PDF). January 2026.
7. Brynjolfsson E, Chandar B, Chen R. Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab. Working paper. Revised 12 August 2026.
8. Humlum A, Vestergaard E. Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI (PDF). Working paper. 13 March 2026.
9. Engberg Räder H. Working Smarter with AI: The Evolving Role of the Trial Statistician (PDF). Presentation at the PHUSE Single Day Event, Basel, Switzerland. 17 September 2026.