When AI handles the work, humans own the decision

The clinical trial industry has spent twenty-five years solving a problem that once felt unsolvable. Paper binders. Manual transcription. Data that took weeks to collect and days more to share across organizations. The bottleneck was distance, and the answer was digitization. Now, AI is arriving in clinical trials, and the question is no longer what it can do. It is where it fits.

Paul Bleicher understood this before most of the industry did. EDC eliminated geography. Data that once required a site visit could be captured in hours, structured, queryable, visible from anywhere. CTMS, eCOA, ePRO, IVRS followed. Each one added precision to what the industry could collect and store. That transformation was not trivial. It changed the basic physics of how clinical trials operate, and the people who built it deserve the credit.

But captured data is not a decision.

The orchestra without a conductor

I keep coming back to this image. A concert hall, fully prepared. Chairs arranged. Music stands in position. Sheet music on every stand, annotated, ready. The instruments are there. Every element of a symphony is present. But no musicians. No conductor. Nothing moving.

Is this a symphony? No. It is the architecture of one, waiting.

That is where clinical trials stand today in relation to their data. We have captured it, structured it, made it accessible across systems and organizations. What we have not built, in most trials, is what happens next. Who receives the data. In what sequence. With what context. With whose authority to act. And with what record of every step.

Clinical trials are not data collection exercises. They are decision-making processes. Complex, regulated, multi-party processes in which the right expert must receive the right information at the right moment and take a defined, accountable action. The orchestra needs players. More than players, it needs a conductor.
‍

‍

What the conductor actually does

When a symphony performs, the conductor is not playing an instrument. They are managing something harder: the sequence, the timing, the handoffs between sections, the real-time judgment about when to accelerate and when to hold. The violins have their music. The brass have theirs. Without the conductor, each section has what it needs to play, but not what it needs to play together.

Clinical trials have the same structural problem. Protocols define how a trial should run. Organizations, sponsors, CROs, investigator sites, expert committees, each have defined roles. Systems capture what each organization does. What most trials do not have is something that governs the handoffs. Something that knows, at any given moment, who needs to act, on what information, in what sequence, and with what accountability trail.

Without that infrastructure, we improvise. Trackers. Email. SharePoint. Excel spreadsheets that one person built and maintained and only they fully understand, until they go on vacation. This works, barely, and only because the people involved are extraordinary. But it is heroic improvisation, not governed execution. And the cost of heroic improvisation, across thousands of decisions in a complex trial, accumulates.

Delays in complex trials are rarely caused by missing data. They are caused by data that arrived but was not routed, by decisions pending because nobody knew whose desk they should land on, by audits where compliance must be reconstructed rather than demonstrated. A one-year delay on Lipitor, for example, would have cost Pfizer roughly $13 billion in peak-year revenue. Other blockbusters carry comparable stakes. The patients waiting for them are not an abstraction.

A new musician arrives

AI is arriving in clinical trials. Not as a future possibility, as working capability. Dossier summarization. PHI redaction. Eligibility screening. Anomaly detection. These are tools being deployed today, producing real outputs that real trial teams are expected to act on.

But an AI output is not an outcome. The output needs to travel somewhere. It needs to reach the right person, at the right stage of a defined process, with the right context, in a way that can be audited. If the infrastructure to govern that journey does not exist, the AI output floats, valuable, complete, and disconnected from the trial it was meant to serve.

Think about what this looks like concretely. A gifted soloist arrives at the orchestra. Someone who can contribute something the ensemble cannot produce on its own. But there is no conductor. The soloist does not know when to come in. They can hear the music, but the tempo drifts, the cues are unclear, no one coordinates the moment of their entry. The capability is real. It has no place to land.

That is the situation most trial teams face when they try to integrate AI today. The tool works. The workflow that receives it does not exist.

AI as a participant in governed workflows

When process management infrastructure exists, when there is something governing the execution of a trial's decision workflows, AI becomes something much simpler to deal with: a participant in a process.

In a governed workflow, a step is a step whether it is performed by a physician, a CRA, or an algorithm. The platform routes information to that step, enforces what must happen, records what did happen, and moves the process forward. The identity of the participant, human or AI, is secondary to their role in the sequence. Adding a new AI capability to a governed process is not a transformation project. It is configuration.

That is the delivery vehicle that most conversations about AI in clinical trials are missing. The question is not only whether the AI is capable. The question is whether the infrastructure exists to receive it: to tell it when to act, to capture what it produces, to route that output to the next decision-maker, and to maintain the audit trail that regulators will eventually review.

When that infrastructure exists, something else becomes possible. The trial becomes legible in ways it never was before. Which sites are fastest at providing documentation, and are they also the ones generating the most queries? Which expert's decisions consistently diverge from committee conclusions, and what does that pattern mean for how you staff future reviews? These are not dashboard questions. They are operational intelligence questions, and they can only be answered when every decision in a trial has traveled through a governed, structured process.

Humans own the decision

For the foreseeable future, the consequential decisions in clinical trials will be made by humans. Not because the technology is insufficient, though in some respects it remains so, but because regulated clinical environments require accountability that traces to a person who exercised judgment. AI will increasingly handle the preparation, the summarization, the screening, the detection. Humans own the decision.

That is the right design.

The trials that will move fastest, comply most cleanly, and integrate new capabilities most effectively will not be the ones with the most AI. They will be the ones with the governance infrastructure that allows AI and human expertise to operate in sequence, each contributing what they are suited to contribute, at a defined moment in a process that is visible, auditable, and improvable over time.

You have the data. You have the AI. You have the experts. The question worth sitting with is whether you have the architecture that connects them.

That architecture is not on the horizon. It is already running trials.

‍

Abraham Gutman, Founder & CEO, founded AG Mednet with a focus on the operational challenges that slow clinical trials down after data is captured. He has over 20 years of experience at the intersection of clinical research and technology.