What Clinical Trials Teach Us About AI Implementation

For many organizations, AI has become a mandate. Leadership understands that AI will ultimately aid efficiency and want to see it used. Meanwhile, teams are responding to these mandates, often by experimenting and trying to find places where AI might fit their workflows. But “using AI” is not much of an objective. What many folks are finding is that, when adoption itself becomes the goal, it is easy to overlook important questions: Where does AI actually improve the work? For whom? What does success look like? And what needs to change to realize that value? 

We encountered those same questions when we looked into how best to introduce Data Advisor, our AI-powered, in-context assistant embedded directly into the elluminate® Clinical Data Intelligence Platform. We knew the technology worked—we had tested it internally, including with our own biometrics team. We knew the context already available within elluminate could enable users to conversationally explore clinical data while keeping that data within its existing governed environment. But we didn’t yet know where Data Advisor would matter most. 

Would Data Advisor make data review faster? Would it help medical monitors investigate patient trends and outliers? Could it help teams connect information across domains or test hypotheses more easily? Would certain users benefit more than others? And what would organizations need to put around the technology to drive adoption?  

As we set out to find answers to these questions, we also learned a great deal about how best to implement AI tools into the clinical trial process.  

What We Learned With Data Advisor 

To discover where Data Advisor might be best leveraged, we worked with early adopters to put our latest AI tool into real workflows and observe not only whether people could use it, but where it changed the work.  

As teams applied Data Advisor to data cleaning and discrepancy review, cross-domain reconciliation, medical monitoring, patient trend analysis, hypothesis testing, and post-database-lock exploration, some reported significant time savings while others reported only moderate gains. One client surfaced a different outcome entirely: adoption. Data Advisor gave users who didn’t live in elluminate every day more opportunity to engage with their clinical trial data. The company saw daily use of their platform soar. 

As we anticipated, the outcomes weren’t uniform. Where some companies saw 50% improvements in efficiencies, other companies saw anywhere from 25% to 10%, and these discrepancies became one of the most useful findings. We discovered that the level of efficiency customers saw while using Data Advisor came down to how they went about implementing the solution. As we worked to identify which implementation practices were seeing the best results, we realized that the model for implementation looked incredibly familiar.  

In fact, it looked almost exactly like the way our industry sets up its clinical trials.  

Implementing AI Tools the Way We Conduct Clinical Trials 

If we set up our clinical trials the way many of us set up our AI tools, it would be akin to opening up a trial and, on launch day, saying, “I wonder all the things this drug can do and all the effects it has.” Instead, we start with a hypothesis, define what we need to learn, select an appropriate population and—arguably the most important step—we establish measures of success. We then use what we learn to determine what happens next. 

What we discovered when we launched Data Advisor is that AI programs should work along similar lines.  

1. Pre-Trial: Design the experiment 

Start with an AI protocol and answer the following questions: 

• What problem are you trying to solve?  
• Where is there a meaningful gap or inefficiency today?  
• Who are the right users to test the technology?
• And what would success actually look like? 

That last question is particularly important, because success with AI isn’t necessarily about efficiency. Depending on the problem, the ideal outcome might be better ability to investigate difficult questions, more confidence in understanding the data, or increased adoption.  

2. Phase 1: Implement AI 

• Choose one or two workflows where you’ve identified a clear problem you wish to solve.  

• Select users who can meaningfully evaluate the technology. 

• Just like we chose early adopters who were willing to lean in and excited about testing, you should identify your own early adopters—the employees who are wanting and willing to lean into something new. Adoption ends up best coming from peer case studies. 

• Decide what you need to learn and what evidence would justify moving forward. 

3. Phase 2: Expand and refine

• Bring in additional users, roles, studies, or workflows.  
• Test whether the value seen initially holds across a broader population.  
• Refine the measures of success as you learn more about where AI fits and where it doesn’t. 

4. Phase 3: Change the way work gets done 

As we worked with our early adopters, there were certain use cases where we realized only mild success. This tended to happen when inserting AI into the current process became somewhat disruptive to a rigid workflow. In those cases, we realized the process itself might be worth revisiting to drive even more efficiency. 

Realizing more value from AI can often require us to reconsider the processes themselves. Once you have evidence of where AI can add value, you can begin looking beyond how the technology fits into today’s workflow and ask how the workflow might operate differently with that capability available.  

Turning AI Implementation From a Roll-Out Into a Learning Process  

What’s exciting about this approach to AI adoption is that we already know how to work this way. We know how to begin with a question, collect evidence, learn from a defined population, and expand based on what that evidence tells us.  

So why isn’t this approach to AI adoption a foregone conclusion in our industry? Maybe it’s because we know how long the clinical trial process can take. But will it take a similarly long time to fully realize the ROI of a new AI tool? Again, I think we may be asking the wrong question. Instead, I’d ask: Can we expect a new technology to prove its full value before teams have learned where and how it works best? 

What we’ve seen is that a phased approach to AI adoption creates the opportunity to build your ROI case deliberately. Define what value should look like before you begin. Demonstrate it in a focused set of workflows. Use those results to build confidence and buy-in. Then expand into the teams and processes where there is evidence that AI can make a meaningful difference. 

Learn more about Data Advisor here.