When we describe TrialNexus to DM leads, the question we get most often is not "how accurate is the SDTM mapping?" It is "who is responsible when something goes wrong?"
That is the right question. And the answer is not "the agent" or "TrialNexus." It is the same person who was responsible before: the DM lead who approved the output.
The agency dial is the product primitive that makes that answer coherent — not just as a policy statement, but as an observable, auditable, adjustable property of the system itself.
The problem with binary autonomy
Most AI tools offer a binary choice: the AI does the thing, or you do the thing. The implicit promise is that you can "turn it on" when you trust it and "turn it off" when you don't. The problem with this model in clinical trials is that trust is not a single variable. You might trust the SDTM mapper's term-level MedDRA coding at 95% confidence and not trust its study-level domain selection. You might trust it for straightforward AE terms and not for lab results with complex derivations. You might trust it in a Phase I where the stakes are lower and pull the dial back for a Phase III pivotal.
A binary on/off switch forces you to make a single judgment call for the whole system. That is not how clinical judgment works. It is not how DM leads think about their own process. The agency dial formalizes what experienced DM leads do naturally: calibrate trust to context.
Three modes, one dial
TrialNexus implements three operating modes, per agent, per study:
Manual (full oversight). Every agent output enters the Decision Queue before it is committed. Nothing is written to any artifact without explicit approval. The DM lead sees every proposal and acts on each one. This is the default for all new studies and all agents.
Supervised (confidence-gated autonomy). Outputs above a configurable confidence threshold — set per agent, per study — are auto-committed. Outputs below the threshold enter the Decision Queue. The DM lead reviews the flagged decisions and can audit the auto-committed ones at any time. This mode is appropriate when an agent has operated on enough studies of a similar type that the DM lead is confident in its behavior on routine decisions.
Autonomous (monitor only). The agent runs fully. Outputs are committed automatically. The DM lead receives a daily summary of what was committed and can flag any item for review within a configurable window. This mode is appropriate for well-understood, high-volume, low-ambiguity tasks — like applying controlled terminology to fields where the mapping is unambiguous and the study team has run the same protocol type multiple times.
The dial setting, every change to it, the identity of whoever changed it, and the timestamp are all logged. If a study has an inspection, the sponsor can show exactly what oversight mode was active for which agent at which point in the study — and why it was changed.
Why we built it per-agent, not global
An early version of the agency dial was a single study-level setting. Every agent moved together. We changed this after three conversations with DM leads who pointed out the obvious: the MedDRA coder and the protocol deviation categorizer are not the same task and should not have the same confidence threshold.
The per-agent architecture also makes the learning curve more manageable. A new TrialNexus user should not be asked to commit to a single dial position for the whole pipeline before they have seen how each agent performs on their specific protocol type. Starting every agent in Manual, then moving individual agents to Supervised as confidence is established, is a natural progression. It mirrors how any competent DM lead would approach onboarding a new programmer: full review, then supervised autonomy, then trust — earned incrementally, per task.
The compliance architecture behind the dial
The agency dial is not just a UX feature. It is the front end of a four-layer control architecture that we think about as the real product.
Layer one is CDISC standards enforcement at the model level. Agents are constrained to produce outputs that conform to the relevant CDISC standard — SDTM IG, ADaM IG, Define-XML v2.1. An output that would violate the standard cannot be committed, regardless of dial position.
Layer two is context-window scoping. Each agent's context is restricted to the data it needs for its specific task. The SDTM mapper sees the protocol USDM object and the relevant controlled terminology — it does not see unrelated patient data, other studies, or information outside its domain. This is not just a privacy measure; it is a consistency measure. An agent that can only see the relevant context cannot make mapping decisions based on irrelevant data.
Layer three is the Decision Queue — the human approval gate that the dial controls.
Layer four is the immutable audit trail, which logs the state of layers one through three for every decision, regardless of dial position.
The dial is the control the DM lead can move. The other three layers are constants. A sponsor who moves every agent to Autonomous mode still has CDISC validation, scoped context, and a complete audit trail. What they have given up is the synchronous human review step — and that is a judgment call that belongs to them, not to us.
The question we are still thinking about
We do not yet have a good answer to one question: how should confidence thresholds be calibrated across different sponsors and trial types?
The threshold that is appropriate for a Phase II oncology study at a large biotech with an experienced DM team is different from what is appropriate for a Phase I at a small academic center with one DM lead who has never used an AI tool. The dial is the right primitive. The right calibration for each position is still partly empirical — it is something we are learning from the studies running in production.
What we are confident about is the direction: in clinical trials, the question is never "can the agent do this autonomously?" It is "can you show a regulator exactly what the agent did, who reviewed it, and who approved it?" The agency dial is designed so that the answer to that question is always yes — regardless of where the dial is set.