Everyone Wants AI in TMF. Almost No One Knows What to Do With It Yet.

Notes from the TMF track at the 2026 CDISC US Interchange

Last week, roughly 500 people gathered for the CDISC US Interchange, and 158 of them came for the Trial Master File track. These are the people who live inside eTMFs every day: records managers, clinical operations leads, QA, inspection-readiness teams, and the vendors who build the systems they work in. If any group is positioned to tell you the real state of AI in TMF management, it’s this one.

So here is the real state of it. The industry loves talking about AI’s potential—but it’s still struggling to name what that potential actually is, beyond use cases it identified years ago.

The same use cases, rebranded

Session after session, the “proven” applications of AI in the TMF came back to the same short list: auto-classification of documents, metadata extraction, and quality-control support. 

Those are real wins. They are also the wins the industry was promising before most people in the room had heard the term “large language model.” 

One eTMF vendor’s presentation covered smart classification and filing and using AI to catch quality issues before a document enters review. A suggested workflow was: upload a document, let AI classify it, have AI flag missing metadata and possible duplicates, and send everything to a human for review. It suggests, but it doesn’t act. That is sensible, and it is also a description of auto-classification with a nicer interface.

Nobody’s getting replaced by AI. Is that a good thing?

The most interesting work we saw went slightly past that line. One presenter shared that they have historically spent about two hours per site reviewing site personnel records, the documentation proving site teams were trained. AI now flags missing and inconsistent records and risk-weights them so human experts know where to dig, flagging things like PI changes and missing medical licenses as high risk. The framing here was the most grounded of the conference: find where things don’t make sense and surface them to a trained human. “You don’t necessarily want AI making the call.”

It’s good work, but look at the shape of it: AI finds problems and humans fix them. Which raises a question nobody on stage really answered: if AI surfaces more issues than your team used to find, is it saving work or creating it? 

Vendors are asking customers to define the problem for them

The clearest signal came during one vendor’s session on agentic AI. The speaker opened by asking the room to explain their problems to vendors so vendors could solve them. Later he made the same appeal again: tell us what AI should do for you, and ask your vendor to make it so.

Customer input matters. But when the largest vendor in the space devotes stage time to asking the audience for a roadmap, that tells you something about where the roadmap is.

The rest of the session didn’t close the gap. The vendor claimed that they don't build LLMs, they build agents around them. On MCP, the protocol that lets AI agents connect to external tools and data, his take was “MCP, I kinda don’t care.” He framed it as relevant only to advanced users. That’s a bold dismissal of plumbing that, in the very first session of the day, had been praised for letting an agent pull that very vendor’s own help documentation.

An audience member asked the question everyone was thinking: what is the value of going to a vendor instead of going straight to Anthropic, or building it in-house, if the vendor is just layering an agent on top of a model it doesn’t own? The speaker’s answer was that “building the agents is the magic here.” A different audience member jumped in with a sharper response: it’s the same reason you buy specialized eTMF software in the first place instead of building it all in house. 

When a customer has to make your value proposition for you, the market hasn’t settled yet.

The people problem is the real problem

If vendors are unsure what to build, adopters are unsure how to use it. That might be the bigger story.

When one speaker asked the audience who knew what all the AI-related terms on his slide meant, three hands went up. Earlier, when a presenter asked who in the room paid for an LLM subscription, one hand. This is the TMF track at a data standards conference, about as engaged and forward-leaning an audience as the industry has.

The data tells the same story. A presentation citing the 2023 State of the TMF Industry report noted that 46% of teams blame “unmotivated contributors” for late filing. We’re struggling to get people to simply file documents on time, and the plan is to ask them to learn an entirely new skillset and supervise AI agents?

One speaker put this tension front and center in his talk on building AI competency at every level of a team. His sharpest point: when only some team members are AI-fluent, “my colleague’s dashboard said it’s fine” is one step away from “the AI said it’s fine.” That is how inspection findings happen. It’s also hard to measure. Everyone agrees teams need to be comfortable with these tools, but nobody presented a way to show a regulator that they are.

Another speaker was refreshingly blunt about it. Her organization’s AI work was driven by document volume and inspector pressure, and she named organizational change management as the biggest hurdle: “the biggest problem is the people.” And there isn’t much AI can do about human behavior. 

Where the honest conversations are heading

Through all the hype, a few ideas stood out as the right ones.

The first is that human oversight isn’t optional, but it can be calibrated. You don’t always need a human in the loop if you return to risk-based principles. That is the mature version of the conversation. AI oversight should be proportional to risk, the same way TMF QC has always been. One organization shared that they are tracking how often humans accept or reject AI recommendations, and investigating whether the AI or the human was wrong. That’s how you build evidence that an AI process deserves trust.

The second is that the audit trail matters more than ever. Teams are starting to mine audit trail data for both AI agent actions and human user actions. When an agent can file documents and set metadata, you need to know exactly who, or what, did what.

The third is that the TMF is getting more connected. Questions arose about whether agents will start reaching beyond the TMF into CTMS, safety, and other systems to give insights more context. The consensus was yes—but that future depends on the plumbing, including the integration standards some vendors would rather wave off.

Meanwhile, the standard itself is changing underneath everyone. The proposed new TMF standard covers hundreds of record types and moves to a non-hierarchical, relational model, with record groups that can sit in multiple zones. An implementation guide will come with it, and Danish inspectors are reportedly already asking sponsors whether they’re aware of it. Any AI strategy built on yesterday’s folder structure is about to need rework.

Where Egnyte stands

We left the Interchange with a clear conclusion. The industry doesn’t need more AI announcements. It needs AI that holds up in an inspection.

That means AI that suggests before it acts, with oversight scaled to the risk of the decision. It means an audit trail that records machine actions as rigorously as human ones. It means AI built on connected content, not locked inside a single application. And it means tools that people at every level of AI fluency can use responsibly, not just the three people in the room who knew what every term on the slide meant.

We’re building toward that, and we’ll have more to share. In the meantime, we were in the room, listening closely to what sponsors, CROs, and the TMF community said they actually need, and to what they’re still waiting for.