When someone says, “We should use AI,” my first question is not “Which AI tool should we buy?” As an independent Quality Auditor in clinical research, I would first ask: “What exactly are we trying to improve—and what are we actually allowed to put into an AI system?” Because in our industry, AI is not just a technology question. It is also a quality, data integrity, confidentiality, security, validation, access-control and governance question. For example, a small CRO might experiment with ChatGPT or Gemini, but later restrict those tools because of concerns around confidential client information, personal data, intellectual property or uncontrolled use. A large CRO may take a different approach and permit Microsoft 365 Copilot because it can operate within an enterprise-controlled environment with organizational permissions and security policies. But even then, Copilot does not magically solve everything. A clinical research professional may work in Microsoft 365, while the actual study data, TMF, CTMS, EDC or other client systems sit somewhere else—often in a client-controlled environment with its own applications, permissions and authorised third-party access. An AI assistant inside our Microsoft environment may therefore have no legitimate access to the information sitting inside those client systems. So when I hear “We should use AI,” I would unpack it into five questions: 1. What is the problem? Are we trying to reduce repetitive work, improve document review, identify inconsistencies, summarize information, improve decision-making, or simply save time? 2. What data is involved? Is it public information, internal company information, confidential client information, personal data, study data, TMF content or regulated records? 3. Where is the AI allowed to operate? Is the tool approved by the organization? Is it inside the enterprise environment? What systems can it access? What permissions does it inherit? What happens to the prompts and outputs? 4. What is the risk of getting the answer wrong? Using AI to draft a meeting agenda is very different from using AI to identify a potential GCP issue, assess a deviation, review a critical TMF record or support a regulatory decision. 5. What human control remains? Who verifies the output? Who makes the final decision? Is there an audit trail? Can we demonstrate what information was used and how the conclusion was reached? So, for me, “We should use AI” is not an AI strategy. It is a hypothesis that needs to be tested against the problem, the data, the environment, the risk and the controls. And sometimes the right answer may be: “Yes, use AI.” Sometimes it may be: “Use only the organisation-approved AI.” Sometimes: “Use AI, but only with de-identified or non-confidential information.” And sometimes the most quality-focused answer is: “No AI here. The risk outweighs the benefit.” That is how I would unpack the statement. Because in clinical research, the question is not simply whether AI can do something. The question is whether AI can do it in a way that is appropriate, controlled, explainable and defensible. Saving Changes...
When someone says “we should use AI,” the conversation is rarely about technology itself. It is usually about pressure for speed, efficiency, innovation, or competitive leverage. The first step is to clarify intent.
Three signals help distinguish what is really being asked.
First, decision proximity. Is AI automating a task, augmenting human judgment, or moving toward managing objectives autonomously? These are fundamentally different categories of work. The closer AI gets to consequential decisions, the stronger the need for governance, traceability, and explicit oversight.
Second, problem clarity. Is there a clearly defined business problem with measurable impact, or is AI being treated as the starting point? When the solution precedes the problem, misalignment and inflated expectations follow.
Third, accountability design. Who owns the outcome if an AI-driven recommendation fails? When responsibility becomes diffuse, risk scales faster than performance.
In many organizations, “AI” simultaneously means efficiency, experimentation, and cost reduction to different stakeholders. Misalignment becomes visible when decision flows and ownership are unclear. A common tipping point is when stakeholders use the same word “AI” but describe different success metrics.
The real shift is not from manual to automated. It is from “man in the loop” to “man in control.” Without deliberate design of responsibility, capability increases while accountability erodes.
Clarity of purpose, category of AI work, and ownership separates disciplined transformation from technological noise.
Your breakdown of the three signals highlights exactly why so many AI initiatives stall. When organizations treat AI as a 'resolve everything' tool instead of performing only certain tasks with humans performing the oversight function, they default in finding solutions to a task. The importance of who is finally accountable is important too. AI cannot be held accountable for its mistakes . We have to be trained to fulfill oversight capacities. Saving Changes...
Anonymous
Mar 29, 2026 9:30 PM
Replying to Eric Simms
...
"What do you want to achieve: Cost savings? Increased production rate? Decreased risk?" - and so forth. In short, I identify the business objective they wish to realize, then examine whether AI is the proper solution, and if so how best to implement it.
All three but I want to frame the question is such a fashion as to have automation query all the publicly available sources, synthesize the results and provide me the explanation or discussion but maybe even nor the "right answer." Saving Changes...
Usually when people say 'we should use AI' without being specific they envision a blackbox that can magically provide solutions. We need to educate and clarify the problem along with appropriate AI applications.
Unfortunately, 80% of AI projects fail. People do not have a clear understanding of what is to be accomplished, e.g., if AI is really needed, jump right to tools because it is fun, over promise and under deliver, and what value is expected. In addition, AI is a set up and forget endeavor--no retraining, governance etc. Saving Changes...
PMI Team Member
