The first clinical trial we ever watched up close was a Phase II cardiovascular study at a mid-size biotech. The DM lead was one of the most competent people we had ever met — twenty years of experience, encyclopedic knowledge of CDISC standards, a clear-eyed sense of where the edge cases were in the SDTM IG. She spent most of her day in Excel.
Not Excel as a tool for doing something clever. Excel as a translation medium: copying values from one system, reformatting them, pasting them into another system. Protocol PDF to CRF specification. CRF specification to EDC configuration. EDC data to SDTM mapping template. SDTM mapping template to SAS program inputs. Her expertise was relevant for maybe 20% of the work — the decisions that required judgment, domain knowledge, and the ability to read a regulatory guidance document and understand what it implied for a specific study design. The other 80% was translation work that any reasonably careful person could do, given enough time.
That is not a training problem. It is not a hiring problem. It is a tooling problem. The clinical trial industry has, for thirty years, compensated for the incompatibility of its systems by hiring highly trained people to translate between them. That strategy worked when the alternative was worse. It is not working at the pace the industry needs now.
Why agents, and why now
The pattern we are describing — a sequence of translation steps that could, in principle, be automated, but have not been because the translation requires enough domain knowledge that generic automation fails — is exactly the pattern that modern LLMs are suited to address.
LLMs are not magic. They are very good at tasks that require broad knowledge, semantic understanding, and the ability to apply documented rules to novel inputs. SDTM mapping is that. DMP drafting is that. CDISC controlled terminology assignment is that. These are tasks where a senior programmer holds a large amount of knowledge in their head and applies it, quickly and correctly, to a specific input. An LLM that has been grounded on CDISC documentation, constrained to produce standards-conforming outputs, and equipped with the right tools can do the translation work at a quality level that supports expert review.
What LLMs are not good at — and what we have been careful not to ask them to do — is final clinical judgment. The decision about whether an adverse event should be classified as related to study drug. The decision about whether a protocol deviation is significant. The interpretation of an ambiguous eligibility criterion in a specific patient case. These require human expertise and human accountability. An agent can prepare the materials for those decisions and present them clearly. The decision belongs to the expert.
The timing is not accidental. The 2020s version of LLMs is the first generation that is reliable enough, in a well-constrained domain, to produce outputs that a qualified expert can review rather than redo. That is the threshold. Not "the AI is always right." "The AI is right enough that expert review is confirmation, not reconstruction."
The translation-layer problem
Clinical data management is, structurally, a series of translation steps between incompatible representations of the same underlying information. A protocol describes a study in clinical language. An EDC collects data in a form designed for site convenience. SDTM represents that data in a standardized tabular structure. ADaM transforms SDTM into analysis-ready datasets. Define-XML provides machine-readable metadata about all of it.
Each translation step has well-documented rules. The rules are maintained by CDISC and updated regularly. The rules are specific enough that, given a protocol and a set of CRF data, a deterministic mapping to SDTM is theoretically possible — and in practice, the cases where multiple valid mappings exist are a fraction of the total decisions.
The gap between "theoretically possible" and "actually automated" has been the domain knowledge required to apply the rules correctly. You need to know SDTM. You need to know CDISC CT. You need to know the SDTM IG well enough to handle the edge cases. LLMs, trained on the CDISC documentation corpus and constrained to operate within the SDTM framework, can now apply those rules at a quality level that supports expert review rather than requiring expert production.
The four things we will not compromise on
We could have built TrialNexus to be more autonomous. Every product instinct in AI software pushes toward "set it and forget it" — the tool that does more, requires less oversight, gets out of your way. We made the opposite choices deliberately.
First: a human stays at every decision point. The Decision Queue is not an optional feature. It is the architecture. Agents propose; humans decide. Every output that enters a clinical trial artifact carries the signature of the expert who approved it — not the agent.
Second: CDISC compliance is the product. If an output does not conform to the relevant CDISC standard, we have not shipped. Define-XML that fails Pinnacle 21 is not a "version 1" — it is a failure. The standard is the bar, not a target.
Third: every decision has provenance. The reasoning behind each agent output — the source passages, the controlled terminology codes consulted, the confidence score, the model version — is an immutable artifact. Not metadata. Not a log no one reads. The thing an regulatory reviewer can follow without a guide.
Fourth: no data lock-in. Your data is yours. Every artifact TrialNexus produces is in open, standards-based formats. USDM makes protocols portable. If you stop using TrialNexus, you leave with everything. We do not compete on dependency.
Where we go from here
TrialNexus Core is live in production today. Signal, the read-layer product for clinical operations and medical monitoring, is live. Agent Studio — the environment for sponsors who want to build their own custom agents inside the same 21 CFR Part 11 compliant runtime — is available now.
We are a small team. We are not trying to be the enterprise software company that owns the clinical trial stack. We are trying to build the right abstractions — agent-based, standards-native, human-supervised — before the industry decides what the right abstractions are. We think we are in the right place at the right time with the right constraints. The next ten years will tell us if we're right.
The DM lead who spent her days in Excel is still out there. She is still translating between incompatible systems, and she is still more valuable doing that than not doing it — because the alternative is worse. We built TrialNexus so that, one day soon, she can spend her days doing the 20% of work that requires her expertise, and let the agents handle the rest. That is the point. That is all of it.