ISPE博客总结制药生产中AI从试点走向规模化实施的阶段与挑战

While the interest is understandable, it is often accompanied by a different set of conversations within organizations: Will AI replace jobs? Will experienced employees lose control of activities that have traditionally depended on their judgment? Can an AI-generated output be trusted in a GxP environment? What will regulators expect when AI becomes part of routine work? These questions create resistance even when the potential benefits of technology itself are promising.

The critical lesson from moving AI-enabled use cases from concept to pilot and toward production is that the difficult part is rarely demonstrating that AI brings considerable value when applied in pharmaceutical manufacturing workflows. The more difficult part is redesigning the work around it. The US FDA has described AI management across the medical-product lifecycle as an iterative process extending from ideation and design through data acquisition, model development, deployment, monitoring, and maintenance.2 That perspective is important: productionization is not a one-time software deployment. It is an operating-model change involving technology, process, people, governance, and continuous learning.

The approach presented here, was applied to AI-enabled knowledge and workflow use cases in pharmaceutical operations and quality. It is intentionally practical: begin with a real business problem, test the value in a controlled environment, and scale only after the workflow, human responsibilities, and performance expectations are understood.

Approach

The implementation approach was organized into three stages: assess and educate, prioritize and pilot, and scale and sustain. The objective was to avoid starting with the technology and instead start with the operational problem and the people performing the work.

Figure 1. Approach for moving an AI use case from identification through pilot and sustained implementation.

In the first stage, “Assess and Educate,” the organization identifies pain points and potential use cases, determines whether AI is an appropriate technology for the problem, builds basic AI literacy, and assesses organizational readiness. Design-thinking discussions are useful because they move the conversation away from “Where can we use AI?” toward “Where are we repeatedly losing time, quality, consistency, or capacity?”

The second stage, “Prioritize and Pilot,” converts candidate ideas into a small number of testable use cases. A simple feasibility-versus-value or return-on-investment matrix can help avoid selecting use cases only because they are technically interesting. A minimum viable product is then implemented in a controlled environment, with the business workflow, data inputs, expected outputs, human review points, and success measures defined up front. This emphasis on a clearly defined context of use is consistent with emerging regulatory thinking. US FDA’s January 2025 draft guidance proposes a risk-based credibility framework for AI models based on their specific context of use,3 while the US FDA-European Medicines Agency Guiding Principles of Good AI Practice in Drug Development (2026) emphasize human-centric design, risk-based application, clear context of use, data governance, performance assessment, and life-cycle management.4

The third stage, “Scale and Sustain,” is where many pilots either become useful operational capabilities or stall. Scaling requires implementation planning, change management, monitoring, optimization, continuous learning, and transfer of knowledge to internal teams. At this stage, the question is no longer whether the AI can perform a task; it is whether the organization can reliably operate the redesigned process.

Challenges

1. AI implementation is process reengineering, not just software deployment

A solution may work extremely well for one or two enthusiastic users and still fail when introduced to a broader group. The reason is that AI often changes who performs which part of the work. Consider an investigation workflow. Traditionally, an investigator may gather information, search procedures and historical events, identify possible causes, draft the investigation, and revise it after quality review. In an AI-enabled workflow, the technology may organize information, identify gaps, retrieve relevant knowledge, structure lines of inquiry, or create a first draft. Human effort then shifts toward ensuring the correct data were provided, challenging assumptions, assessing technical relevance, resolving ambiguity, and approving the final conclusion. That is a different process, not merely a faster version of the old one. Resistance frequently reflects this redistribution of work and accountability rather than resistance to technology itself.

2. Managing the expectation that AI should do everything

Consumer AI has created an expectation that a user should provide a short instruction and immediately receive a complete answer. In pharmaceutical operations, useful outputs depend on the completeness, accuracy, and context of the inputs. AI cannot identify a credible root cause when essential event information was never captured, and it cannot generate an organization-specific procedure when the relevant process requirements and source documents are missing. Users therefore need to understand what the application can do, what it cannot do, what information must be supplied, and how to challenge unsupported conclusions. Upskilling is not an optional training activity after deployment; it is part of the control strategy.

3. User expectations now move faster than traditional software cycles

AI changes expectations for software evolution. During a pilot, users may quickly identify that a question should be asked earlier, that an output should follow a different workflow, or that the tool should compare information in a different way. When those changes are incorporated quickly, users see the solution evolving around real work and engagement increases. When every change enters a long development backlog, enthusiasm drops. The challenge in a regulated environment is to combine this faster feedback cycle with appropriate testing, documentation, change control, and risk assessment.

4. Regulatory uncertainty can produce either excessive caution or excessive confidence

Organizations may delay useful experimentation because they are unsure how regulators will view AI, while others may assume that human review can be minimized because the technology appears capable. Recent regulatory signals point toward a more balanced position centered on accountability. In an April 2026 warning letter, the US FDA cited inappropriate use of AI in pharmaceutical manufacturing and stated that AI-generated documents used for cGMP activities must be reviewed and cleared by an authorized human representative of the quality unit.5 The United Kingdom’s Medicines and Healthcare products Regulatory Agency (MHRA) has similarly stated that its concern is not whether AI is used in inspection responses, but whether the information is accurate, verifiable, technically reviewed, appropriately approved, and supported by evidence.6 These examples reinforce a practical principle: AI can support regulated work, but it does not transfer accountability away from the organization.

Learnings

1. Use-case selection should be bottom-up as well as top-down

AI initiatives are frequently sponsored by senior leadership, which is necessary to secure attention and resources. However, leadership often sees the symptom while people on the shop floor see the mechanism creating it. Consider an illustrative example: management sees slow batch release and concludes that additional review capacity is required. Reviewers may reveal that a substantial portion of their effort is spent repeatedly locating the same requirements across procedures, checking entries against approved parameters, and resolving avoidable documentation inconsistencies. An AI-enabled review assistant that organizes information and directs attention to exceptions may therefore create more value than a broad “AI for batch release” program. The lesson is to define the business outcome from the top, but discover the specific use case with the people doing the work. AI democratizes access to knowledge and analytical capability, so relatively simple front-line problems can become high-value opportunities when solved at scale.

2. Leadership sponsorship matters most after the pilot

Launching a pilot is comparatively easy. Productionization exposes assumptions that were wrong, workflows that need redesign, integrations that were underestimated, and user requirements that were not visible at the beginning. Leadership must allow sufficient time and resources to adapt the solution, tolerate controlled mistakes during learning, and obtain outside expertise where internal capability is still developing. A pilot should not be judged only on whether the first version worked perfectly. It should also be judged on what the organization learned about the process, the users, the data, and the controls required for scale.

3. The business process owner should own the AI project

Information technology, digital, data science, validation, and quality functions are critical enablers, but they should not become substitutes for business ownership. If an AI application is intended to improve investigations, batch review, maintenance, planning, or manufacturing performance, the owner of that business process should remain accountable for whether the redesigned workflow delivers the intended result. Success measures should therefore be operational: cycle time, right-first-time performance, rework, review effort, consistency, throughput, or earlier identification of risk. Technology adoption is not the outcome; improved operations are.

4. Upskill employees to become critical users, not passive users

Training should extend beyond prompt writing. Employees need to understand how the quality of context affects the output, why AI can generate plausible but unsupported content, how to identify missing evidence, when to escalate uncertainty, and which decisions require subject-matter judgment. This changes the narrative around jobs. The objective is not to outsource expertise to AI; it is to outsource selected repetitive cognitive activities while increasing the value of human judgment. The most effective user is not the person who accepts AI output fastest, but the person who knows how to interrogate, verify, and improve it.

Conclusion

The pharmaceutical industry does not lack AI ideas. The more difficult task is converting an idea into a controlled, useful, and adopted way of working. Experience suggests that the strongest path starts with a material operational problem that, involves users early, pilots the redesigned workflow rather than only the technology, and scales with clear business ownership, rapid learning, appropriate governance, and sustained leadership support.

Regulatory expectations are also becoming clearer in principle even while the details continue to evolve. The US FDA and MHRA publications do not point toward removing humans from accountable GxP activities. They point toward defined context of use, risk-based controls, reliable data, life-cycle management, appropriate human review, and organizational accountability.3456 Those principles are compatible with innovation; in fact, they provide a framework for doing it responsibly.

The question for pharmaceutical manufacturers is therefore moving beyond “Can AI perform this task?” A more consequential question is: “Can we redesign this process so that AI handles the activities it is good at, people remain accountable for the decisions that require expertise, and the combined system produces a measurable improvement in quality, productivity, and consistency?” That is the point at which AI moves from demonstration to production.7