Trialt’s Dr. Yuri Kartashov spotlights AI limits in clinical data management

Trialt Chief R&D Officer Yuri Kartashov delivered a keynote this week at COG CRO Summit Europe in Amsterdam on what clinical data management can safely delegate to AI and what must stay under human control. The talk focused on evidence-based rules for using AI in routine data tasks, as CROs and sponsors weigh faster workflows against trust, oversight and risk.

Why it matters: - Clinical data teams are under pressure to use AI without sacrificing quality, oversight or auditability. - Trialt’s message centers on a practical question for CROs and sponsors: which tasks can AI handle now, and which require human approval. - The stakes are higher in data management because small workflow errors can affect trial timelines, site relationships and study decisions.

What happened: - Trialt said Chief R&D Officer Yuri Kartashov, PhD, delivered a keynote this week at COG CRO Summit Europe at Novotel Amsterdam City in the Netherlands. - The presentation was titled “Assisted to Automated to Autonomous: Evolution and Impact of AI in CDM.” - The event brought together senior leaders from across the European CRO sector to discuss technology, partnerships, resource optimization, site relationships and other priorities in clinical research.

The details: - Kartashov separated AI use in clinical data management into three modes: assisted, automated and bounded autonomous. - In assisted mode, AI drafts work and people make the decisions. - In automated mode, a predefined workflow runs inside agreed rules and checkpoints. - In bounded autonomous mode, an agent chooses its next steps within limits set and overseen by people. - He used a data query as the working example, including scenarios where AI drafts a query, sends it to a site or closes it. - Kartashov said each added permission changes the evidence a team should require. - He compared current AI claims with documented outcomes in medical coding, data review, report generation and trial screening. - He also placed those claims alongside documented failures and near misses. - A pre-conference survey of CRO and sponsor data-management leaders found that drafting is widely trusted, while sending and closing queries are not. - The same survey found that supervised use, not autonomy, is what most respondents expect by the end of 2027. - Kartashov said, “A query drafted in seconds can still take days to resolve.” - Kartashov said the key unit of value is an accepted outcome, including the time people spend checking, correcting and recovering. - He said permission should follow evidence, starting with one repeated task, keeping a person approving each action, measuring the work left for people and expanding only when evidence supports it. - The presentation also covered practical setup for agentic workflows in routine data management tasks, including query drafting, data-transfer comparison and reconciliation. - Kartashov outlined how to give an agent narrow, named tools with read-only access to study data. - He described where human checkpoints should sit in the workflow. - He said each approval should be tied to the current data version. - He said difficult cases should be tested before any live use. - He also discussed an initial AI-assisted script, a learning routine that stays with the work, and the habits, skills and cost decisions tied to AI adoption. - Attendees were urged to leave with one recurring task that can become a tested, reviewable workflow. - Yaroslav Sud, Trialt’s VP of Biometrics, said Kartashov’s view of AI is grounded in evidence rather than hype. - Sud said the talk gave data management leaders a practical way to decide what to delegate now, what to hold back and what evidence to request before expanding AI’s role.

Between the lines: - Trialt is positioning AI adoption as an evidence-gated process, not a broad automation push. - The company’s framing suggests the near-term opportunity is in low-risk drafting and support work, while higher-risk actions still need tighter controls. - The emphasis on acceptance, checkpoints and version binding reflects a focus on traceability, which matters in regulated clinical research.

What’s next: - Trialt said attendees were invited to identify one recurring task and turn it into a tested, reviewable workflow. - The company also pointed readers to its clinical data management page for more information and to request a copy of the presentation. - COG CRO Summit Europe included artificial intelligence for CROs among the featured topics for its 2026 program. - More information is available in the summit program and Trialt’s clinical data management page.

The bottom line: - Trialt’s keynote argues that AI can speed clinical data work, but only if each new task is backed by evidence, human oversight and clear operational limits.