GLG Frontier
>
Research

With Clinical Trial AI Winning at the Bookends, the Middle Is the Next Test

Four practitioners who have run trial operations inside large pharma and contract research organizations (CROs) explain where AI is actually live in clinical development, where it stalls, and why the hardest problems are consistency, accountability, and incentives – not model capability.
GLG Frontier
Frontier Briefing
November 2026
CONTRIBUTORS
GLG Frontier Research Team

In clinical trials, experts agree that AI has real traction in the document-heavy, data-rich work sitting at either end of a study. The middle of the trial, on the other hand – monitoring, site management, and the daily work of keeping a study clean – is still very hands-on, with AI acting as an assistant at best.

As for what’s holding the middle back, and who captures the savings when it moves – questions that should be front of mind for any model builder – the experts are split.

Where AI Already Runs: The Bookends

At the front of a study, AI reliably supports protocol design and feasibility. The target is fewer protocol amendments – the mid-study changes that push out timelines and blow through budgets. The same data feeds site selection, which ranks sites by investigator history, therapeutic experience, past enrollment, and proximity to eligible patients. A former CRO operations VP told us he sees AI adoption concentrating most at site selection and recruitment right now.

On the back end, generative AI drafts clinical study reports and assembles regulatory submission dossiers from the protocol and the captured data. The former senior CRO executive sees dossier preparation as one of the greatest time savings in the trial life cycle.

A layer of operational work sits between both ends, where practitioners are quietly automating meeting minutes, document drafts, and presentations. On the data-capture side, imaging algorithms flag tumors or lung scarring for human review, and real-time quality checks on remote spirometry catch bad readings before they become lost data.

The Middle Is Still Augmentation

While developers see great potential for AI in protocol monitoring, progress remains stunted. It’s labor-intensive work, and technology vendors are marketing agentic monitoring. But the CRO operators we spoke to see monitoring and data management still stuck in pilots and staff augmentation. When that agent does arrive, one said, there will be fewer monitoring roles overall, with the remaining CRAs shifting focus from document and data review toward site relationships, which are much harder to automate. That matters because monitoring is where the human layer touches the site and, through it, patients.

A former pharma emerging technology lead suggested that any agent meant to replace a CRA role should have to prove itself before the role goes away, which would require absorbing costs up front and running a series of pilots against baselines from comparable studies to achieve a success threshold.

Consistency Beats Cleverness

AI that reviews patient responses or medical data has to judge inputs consistently, so results stay statistically comparable, and an auditor can see why each call was made. The goal, the former pharma technology lead said, is “trying to make the solution as deterministic rather than generative as possible.”

The same logic applies to recruitment. Patient-matching tools and registries can push more candidates to sites, but volume without fit produces screen failures and frustrates site staff who are already stretched thin. The number of patients who qualify matters much more than referrals sent.

What’s Actually in the Way

When it comes to scale, the practitioners emphasize different hurdles. The former senior CRO executive sees regulation as the main barrier. Requirements for validating AI tools vary by geography, with no consistent global rulebook. He considers data fragmentation and systems integration to be solvable technical problems. “I don’t think user trust will be much of an issue,” he said.

The former pharma technology lead identified two key challenges: people and plumbing. On the people aspect, every AI tool or agent embedded in a workflow needs a named system owner, clear responsibilities, an audit trail that satisfies good clinical practice (GCP), and a plan for when it fails – not to mention change management required with any staff reductions.

On the plumbing side, legacy platforms often can’t build AI-native integrations through APIs and MCP connections like newer vendors. Incumbents may make third-party integration harder when they sell a competing feature.

Consent and data residency add a third constraint. In his view, using trial data to train models calls for a separate, optional consent that patients can decline. Residency rules in the EU and elsewhere limit where data can be hosted, and they can change during a single trial.

Where They Split: Who Captures the Savings

Even more than what’s in the way, the practitioners disagree on who stands to gain the most. The former pharma technology lead points to an incentive gap: CROs are paid per line item, so automation lowers their costs without necessarily lowering what sponsors pay. Over time, he expects some sponsors to assemble enough tools to run more of a trial themselves, in effect, becoming an in-house “AI CRO.”

The former senior CRO executive expects the opposite. A sponsor’s core skill is developing drugs, not deploying IT, so he sees sponsors buying AI from CROs and technology vendors rather than building it. He expects savings to fund more studies, since budget cycles at mid-size and large pharma routinely cut trials the company considers worthwhile.

A former CRO medical director describes sponsors already pulling key functions in-house and outsourcing others à la carte. In his view, sponsors can reasonably expect live access to the CRO’s trial database rather than after-the-fact reports. He sees speed as easy to price, because each month of delay has a value in lost revenue once an asset is approved, while quality remains a judgment call. The former pharma technology lead reaches a similar conclusion from a different direction: business cases built on speed and quality hold up, while those built on cutting another department’s headcount turn political.

Table Stakes, Not the Tiebreaker

For now, AI doesn’t win the contract. The former CRO operations vice president compares it to earlier tech waves, like electronic data capture (EDC), risk-based monitoring, electronic clinical outcome assessments, and decentralized trials. “Everyone felt existential before it arrived,” he said. Sponsors want to see a CRO moving on AI to stay credible, but they still choose partners on execution, therapeutic depth, and track record. The former medical director agreed that experience and word of mouth still outweigh AI claims in CRO selection, and that hitting projected enrollment timelines is what wins repeat business.

Our Take: What This Means for Teams Building Clinical AI

The four interviews reveal the bottlenecks: consistency, traceability, integration, and incentive design. As such, evaluators should test the following questions well before deploying a clinical AI system in a live study:

  • Does the system make the same call on equivalent data for the first patient and the 500th, across model versions?
  • Was the human baseline measured on comparable studies before the pilot, with a success threshold agreed in advance?
  • For matching and recruitment, is the eval target qualified patients per site rather than referral volume?
  • Can every output be traced for an auditor – inputs, model version, and reviewer – and does each agent have a named human owner?
  • Are per-patient consent flags and data residency enforced at the data layer, before training or inference?
  • Does the system degrade gracefully when a site runs on legacy systems, unstructured data, or paper?

These questions become useful evals only with domain input. Where the consistency threshold sits, which edge cases break a monitoring agent, and what an inspector will ask to see are judgments that come from people who have run trials.

Go Deeper

This Briefing draws on four GLG Library interviews: AI-Driven Clinical Trial Operations (September 18, 2026), Impact of AI in CROs (August 13, 2026), Contract Research Organization Updates (August 20, 2026), and CRO – Sponsor Perspective (September 27, 2026). GLG clients can read the full transcripts in myGLG. To pressure-test your own approach, GLG can connect you with practitioners in clinical operations, data management, and trial monitoring through a one-on-one expert call. Teams designing evaluations for clinical AI can work with GLG Frontier to build benchmarks and test scenarios grounded in real practitioner judgment.

About this piece: This Frontier Briefing draws on four GLG Library interviews recorded in August and September 2026. The experts are a former head of emerging technology for global clinical solutions at a large global pharmaceutical company; a former senior vice president and managing director at one of the world’s largest CROs, with nearly four decades advising pharma and biotech; a former vice president of operations at a large CRO’s biotech unit, with more than 25 years in CRO operations; and a former senior medical director at a mid-size CRO who now advises sponsors on clinical development. The frameworks and recommendations for sponsors and CROs are theirs. The takeaways for builders are GLG Frontier’s.

Get connected to the world's leading platform for trusted human expertise

Enter your information below to get in touch with our team.