The industry is buying sports cars and ignoring the roads

Within the domain of clinical trials, every technological revolution appears to correspond to a pattern. Whenever a new and cutting-edge technology is introduced to the market, vendors produce demos and present at numerous conferences. The industry has generally reacted with excitement, but roughly six months to 1.5 years after release, companies find that their teams have the same discouraging realisation that the technology is only successful in controlled settings and proves problematic in their actual environments. This cycle has already occurred with RPA and first-generation NLP implementations and will continue with agentic AI unless the clinical industry adopts a new strategy. 

More than 40 sessions at the PHUSE US Connect 2026 conference involved educational discussions related to AI, LLMs or agentic workflow. Therefore, it is evident that there is a great deal of focus and attention on transformative technologies. Examples include multi-agent pipelines that produce ADaM datasets, generating SAPs for protocols using LLMs (this is no longer theoretical) and discover outliers in safety data before human intervention.

The sessions on data standards had titles such as "SDTMIG v4.0 — Are You Ready For It?", "Reimagining SDTM Mapping with Agentic AI and Automation" and "Transforming Legacy Data into CDISC Standards" A separate agenda included hands-on workshops detailing how to successfully implement and utilize CDISC 360i metadata-driven automation as well as an extensive CDISC 360i topic over the three-day conference. While these sessions may not have been the most exciting, they sent a clear message: a large portion of the industry continues to struggle with major challenges in terms of finding accurate and structured source data, obtaining consistent, accurate and easily understood metadata, obtaining SDTM mapping that may have been created by someone who was available and for which no documentation exists, and understanding the interpretation of ADaM Datasets that are technically valid, but require special expertise to interpret. Agentic AI requires clean, uniform and machine-readable structured data to provide accurate results. Without this possibility, agentic AI may produce wrong results, while demonstrating great confidence in the responses provided. From a regulatory submission standpoint, this creates an unacceptable level of risk.

The hidden cost of legacy clinical data infrastructure

At this year's PHUSE US Connect 2026, a recurring theme across sessions was the challenge of clinical data infrastructure that hasn't kept pace with modern demands. Presentations highlighted how evolving submission requirements, including CDISC standards for metadata and SDTM mapping are placing increasing pressure on existing data workflows. When AI-driven solutions enter these environments, their effectiveness depends entirely on the quality and consistency of the inputs they receive. A well-designed AI pipeline cannot compensate for missing context, ambiguous mappings, or undocumented legacy decisions. The result is that organizations investing in AI tooling often spend significant time on data remediation before realizing the expected efficiency gains, making infrastructure modernization not just a technical priority, but a strategic one.

When clinical data management processes were simple (i.e., you receive programming specs, program code, check code output, submit), investigators could compensate for structural inconsistencies (or "know where the skeletons were buried") by asking people about it. AI cannot call up someone who knows why there is non-standard definition for DOSE based on how it was created historically, nor can AI assume that Tab #2 ("SDTM mapping") of an Excel file ("FINAL_v3_USE THIS ONE") will contain the mapping document for the data from this study. AI can only use what is given to it, and do so quickly (because AI systems are optimized for rapid execution on well-defined inputs, not for reconstructing missing context). Companies spend months building an AI TFL generation pipeline and then a few more manually "cleaning" the input required to run the AI. The efficiency of AI quickly diminishes. Executive teams often ask why their investment has not yielded a decent return. Vendors receive all of the blame for the problems despite they were actually just being given broken inputs to work with in the first place. The problems are historically associated with the integrity of the data used to create the submissions for each trial.

Standards as infrastructure, not compliance

To the industry, achieving CDISC compliance is seen as a regulatory requirement that must be accomplished before submission, must be audited, and must be visible on your QC list. Standards provide the foundational infrastructure upon which an artificial intelligence tools will operate effectively. Even with the highest level of sophistication in building the most intelligent autonomous vehicle, if the roads are not paved, clearly delineated or are set up differently from one city to another, the autonomous vehicle will not operate properly. It will operate with a high level of efficiency when you are demonstrating how it works, but will only operate with a driver present in production mode. CDISC 360i represents the most ambitious effort to create a standard of infrastructure around the linkage of concepts, data collection, analysis and submission in a functional, machine-able chain between the design phase of a study and the final regulatory submission. The session at PHUSE this week on 360i include "Agentic AI-Powered SDTM Automation: A CDISC 360i Metadata-Driven Approach," and "CDISC 360i: From Vision to Implementation — Lessons Learned." These sessions signify progress in the area of CDISC compliance.

However, while 360i addresses how to achieve CDISC compliance going forward, it does not address the problem of remediating over 10 years of disparate data that exists in archives, the data from active studies that were created using older conventions and the operational procedures that companies continue to follow and operate under as if their standards are delivered once the study is completed.

What bridges now and next

At entimo we approach this challenge with a different perspective. The question is not simply about how the data will be organized, it's about ensuring that the correct action takes place at the proper time, consistently, with no human memory required for activation.

Clinical data pipelines are filled with implicit trigger points. When a database lock occurs, it can only be resolved once someone notices the error, notifies another person, who then starts extraction, who then sends the file to programming, etc. If a protocol amendment is approved, there's a need to update metadata, review mappings and flag downstream specifications before being able to execute the correct workflow step. If there’s a Safety Signal in interim data, who will be the one to identify it? The nature of these hand-offs is such that that's where errors accumulate, and delays are compounded. If the trigger is a human memory or an email thread, the AI system has a high likelihood of failing to perform safely. Business Event triggers perform the same function as the implicit hand-offs; replacing them with explicit, defined, and system-wide events. If something meaningfully occurs within the data landscape, a signal is emitted, the next step of the workflow is executed, with full traceability, without any dependency on a person being physically present/available and attentive, at exactly the right moment. This concept isn't new within the software architecture world; however, it is an extremely new concept within clinical data operations and the timing of this shift is purposeful.

AI requires more than clean data to be effective.

What the PHUSE US Connect 2026 agenda is actually telling us

The PHUSE US Connect 2026 agenda represents the pharmaceutical industry at a pivotal moment in terms of advances in AI technology. As AI sessions demonstrated what is achievable, the standards sessions revealed persistent issues with the foundation of the industry. Organizations need to prioritize solutions to these foundational issues if they want to achieve reliable and trustworthy results for both patients and regulators. The teams who will derive the greatest benefit from agentic AI in the next three years will not be those who moved quickly to adopt agentic AI. They will be the teams that take advantage of this time to address foundational issues, consider standards as part of their infrastructure; design processes around events, rather than manual hand-offs, ask themselves “What's a reliable AI operating model?”, prior to asking themselves “How do I find the best AI vendor?”.

This type of internal dialogue is much more difficult than evaluating a vendor's demo. It involves assuming responsibility for many years of accumulated issues and funding what will probably not appear as innovative as it will a building block of successful innovation.

Unless there is an effective road to support a sports car, it will not get you anywhere.

Entimo sponsored the Poster Session at PHUSE US Connect 2026 in Austin last week. We presented on AI-supported validation of open-source software and the next generation of automation through business event triggers. Reach out to us to learn more.