Pharma Leaders Weigh in On AI's Measurable Results In 2026

Pharmaceutical companies have spent years piloting AI.¹ The harder question for manufacturing and quality leaders is whether those investments are changing how plants and supply networks run. PharmTech connected with Mahmood Majeed, managing partner and global head of the digital and technology business at ZS, to discuss how AI is moving out of pilots and into day-to-day manufacturing work.

New research from ZS, a global consulting and technology firm, suggests the answer is increasingly yes, at least in some functions.¹The firm’s 2027 CDIO study surveyed pharma technology executives in the US, the UK and 4 EU markets. It found that only 38% of organizations report measurable AI outcomes across multiple business functions. Gains are arriving fastest in enterprise IT, commercial operations, and supply chain and manufacturing. In supply chain and manufacturing, the share of respondents reporting measurable outcomes rose from 29% in 2025 to 53% this year, the largest jump of any function. Drug discovery nearly doubled, from 17% to 35%, while clinical and medical functions stayed essentially flat.

The study also points to a gap between isolated successes and enterprise-wide control.¹ Only 17% of respondents say AI governance is fully operational across the enterprise. Among organizations seeing measurable results across multiple functions, 72% identify data infrastructure and platform modernization as mission-critical, compared with 46% of other organizations.

PharmTech: Supply Chain and Manufacturing Posted the Biggest Jump of Any Function, From 29% Reporting Measurable AI Outcomes in 2025 to 53% This Year. What Changed on the Plant Floor and in the Supply Network to Drive That?

Majeed: I think the biggest change is that we’re finally seeing AI move out of pilots and into day-to-day work. Companies have spent years improving the underlying data and connecting manufacturing, quality, planning and supply-chain systems, which is making it easier to use AI in actual workflows.

Scheduling, demand and inventory forecasting, process monitoring, deviation analysis and supply-risk sensing are good examples because companies make these decisions over and over again. We’re already seeing measurable results. In one independently published IDC case, a global pharmaceutical company improved its stock-out forecast accuracy by more than 20%.²

There’s also a lot of interest in more sophisticated applications like digital twins, cross-functional planning and AI agents that coordinate activities across functions, but we have to be realistic about where those are today. In a regulated manufacturing environment, the more consequential the decision, the more important the data, validation and human oversight become, with people remaining accountable for critical quality decisions.

Among Value Leaders, 71% Say AI Executes Decisions Within Defined Limits in Supply Chain and Manufacturing. What Do Those Limits Look Like in Practice?

The level of autonomy should match the risk of the decision. AI can monitor a process, classify an issue, reconcile records or prepare information for review, but decisions about release, disposition or corrective action still sit with a qualified person.

Those boundaries need to be established before the system goes live. The process owner and subject-matter expert need to define what the model is supposed to do, what inputs it can use and what actions it can take, with quality approving those boundaries. Then you also need to test that the system stays within them and escalates to a person when it should. Once it’s live, you continue monitoring it under change control.

The draft EU Annex 22 takes a similar risk-based approach. It emphasizes documented intended use, controlled test data and human oversight, and is currently strictest about models that are static and predictable. That helps explain why we’re seeing more autonomy in structured scheduling and monitoring workflows, while people stay much closer to decisions that can affect product disposition.

Value Leaders Are More Likely to Prioritize Validating AI Before Regulated Use (54% vs 40%). How Are Companies Approaching Validation for AI That Touches GxP Processes?

It starts with intended use. Before testing an AI system, you need to be clear about what it’s supposed to do, what data it will use, how its output could affect a regulated process and what happens when the system is uncertain.

Validation should depend on the level of risk. Companies need representative data and predefined acceptance criteria to show that the system reliably does what it’s intended to do. They also need controls around access, changes and ongoing performance.

FDA and EMA are generally moving in the same direction. Both are focused on risk-based validation, fit-for-purpose data, lifecycle management, traceability and making sure there’s appropriate human oversight. Some of the detailed AI-specific frameworks are still evolving, though. FDA’s detailed AI framework is still draft guidance, and the proposed EU GMP Annex 22 is also still draft. Companies have to meet existing GMP requirements today while preparing for those new expectations.

Only 17% Of Organizations Say AI Governance Is Fully Operational Across the Enterprise. What Is the Practical Risk of Running AI Under Governance That Is Still Being Defined?

The practical risk is that AI can scale faster than the accountability around it. If governance is still being worked out, every new use case brings you back to some basic questions. Who owns the decision? Who is allowed to change the model? What happens when it gets something wrong?

Those aren’t abstract questions in a GMP environment. Gaps can lead to data-integrity findings, uncontrolled changes or failures that go unnoticed as the system’s performance declines.

Traceability should be built into the system from the beginning as opposed to being added afterward. Actions should be logged, models and prompts versioned, access controlled, and approvals and escalation points built into the workflow. If an inspector asks what the system did, you should be able to show what data it used, which version was running and who signed off.

Value Leaders Are Keeping Core Data Engineering In-House (69%vs 53%) But Are More Willing to use Partner Agreements Elsewhere. What Does That Mean for How Sponsors Work With CDMOs?

I think this changes the sponsor-CDMO relationship because we’re no longer talking only about exchanging documents. Manufacturing and quality data are moving between organizations and could potentially be used by AI systems. Both sides need to be clear about who owns that data, who can access it and what it can be used for, particularly if process data could be used to train or improve an AI model.

Outsourcing manufacturing doesn’t remove the sponsor’s oversight responsibilities. The CDMO also remains responsible for the GMP activities it performs. Quality agreements therefore need to clearly define responsibilities, records, system controls and how changes are communicated.

Better digital exchange can give both sides more visibility and make investigations, planning and technology transfer easier. But the rules around the data and the accountability for decisions need to be clear.

Drug Discovery Nearly Doubled, From 17% to 35%, But Clinical and Medical Stayed Flat. What Is Holding Back AI Value in Development?

Discovery has an advantage because it’s high-volume and iterative. A wrong idea is relatively inexpensive, and you can test a lot of them. Clinical and medical work under very different constraints. The evidence bar is higher, the feedback cycle is longer and the cost of getting something wrong is much greater. AI is more likely to support a decision there, with an expert validating the output. Even strong pilots in clinical data work have needed expert sign-off.

CMC and process development have their own challenge because AI output can eventually affect process understanding, product quality and regulatory evidence. That raises the bar for data quality, validation and scientific oversight.

There are practical uses today in analyzing process data, planning experiments, monitoring processes, supporting documentation, and helping with scale-up and technology transfer. Digital twins and AI-enabled process control are promising too, although adoption is still relatively early.

Connecting development and manufacturing data will also matter. You don’t want valuable knowledge generated during development to get lost as the product moves through scale-up and into commercial manufacturing.

Value Leaders Are Nearly Twice as Likely to Be Redesigning Roles For AI-Supported Work (48% vs 27%) And Are Creating Roles to Oversee AI Systems. What Does That Look Like for Manufacturing and Quality Staff?

You can see the change clearly in the work people are doing. A quality assurance professional who used to spend hours reconciling batch records may now review a package prepared by an agent and decide whether it’s right. A process engineer may spend less time searching through logs and more time challenging an agent’s root-cause analysis.

This changes the skills people need. They must know when they can rely on the system, when they should challenge the output and when a person needs to step in.

For reskilling, I’d start with the people who will review and approve AI output. Formal training is useful, but it should be paired with hands-on use in their own processes. Some large companies are already running accreditation programs.² It’s still early to say we’ve landed on a completely new set of job titles, but responsibility for model performance and oversight is becoming part of operations and quality roles.

We’re seeing changes elsewhere in manufacturing too. Operators are being trained through AR/VR to switch lines quickly, while operational technology security is becoming part of the manufacturing and quality agenda. Global capability centers are also taking on more responsibility for AI and agentic operations. Novartis’s Hyderabad global capability center, for example, is the largest pharma global capability center in India, with approximately 8000 professionals across drug development and supply chain, AI, finance and HR.

Whatever the role looks like, the accountability doesn’t go away. Regulatory expectations for human oversight of critical GMP decisions mean a qualified person still has to own model performance and review.

Did You See Regional Differences Between US And EU Companies Answering the Survey?

We looked at the year over year comparison both with and without the additional respondents, and the findings were similar. We saw a minor regional difference in the share of EU respondents who said commercial sales and marketing was “already delivering value,” but the other categories for this question were very similar across regions.

The pressures companies face do differ by region. US pharmaceutical companies are dealing with questions around manufacturing footprint, trade and supply chains. In Europe, companies are operating within a more comprehensive AI regulatory environment and a strong policy focus on medicines resilience.

For manufacturing, though, there’s a lot of common ground in what regulators are looking for, including the principles I mentioned earlier around risk, data and human oversight. Global companies will need controls that can work across those different regulatory environments.

References

1. ZS. AI’s value in pharma: 5 signals shaping the 2027 outlook. Accessed October 22, 2026. https://www.zs.com/insights/ai-value-in-pharma-cdio-2027
2. Hoots C, Dummann B, Szymaszek M. How we’re upskilling and preparing our workforce to thrive in the age of AI. AstraZeneca. April 11, 2025. Accessed October 22, 2026. https://www.astrazeneca.com/media-centre/articles/2025/upskillingAI.html
3. Giguashvili N. IDC MarketScape: worldwide life sciences intelligent supply chain services 2025 vendor assessment. IDC; 2025. Report US51047023. Accessed October 22, 2026. https://www.zs.com/content/dam/global/master/pdfs/idc-marketscape-worldwide-life-sciences-intelligent-supply-chain-services-2025-vendor-assessment-2025-aug.pdf