← Back to News & Insights
Gaspar Wong · Nov 11 · 8 min read

AI in Regulated Industries Doesn't Fail on Capability. It Fails on Validation.

The case for a hybrid of deterministic facts and agentic judgement in clinical development, and why that separation is what makes evidence reusable.

Two serious adverse events sit in the safety tables of a clinical study: liver failure and acute kidney injury. Both are coded as serious and severe, and the investigator assessed both as not related to treatment. A system that only reports the record would write exactly that, and it would be correct.

Our system wrote that too. Next to it, it raised a question: the therapy's best-characterised adverse effect is known to produce exactly that combination of organ injury. It did not change the causality assessment, and it should not have. It put the clinical context next to the record and sent both to the medical writer, who made the call.

That is the distinction that matters: not an AI that decides, but one that surfaces what a reviewer would want to know before they decide. It also captures the whole problem of AI in regulated work. Recording what the data say is easy to defend but adds little. Noticing what a senior reviewer would notice is valuable but hard to defend. Any AI used in a regulated process has to do both, and has to show which one it is doing at any moment.

The document is an output, not the whole picture

Everything a trial learns ends up in a document, and a document is the one format you can't ask a question of.

A clinical document communicates what a study found. What sits behind it, the evidence for each finding, the reasoning used to interpret it and the decisions made in review, is valuable in its own right. Today that material is scattered across tables, listings, narratives and reports. The same finding may be described in several places, yet working out how it relates to other findings, or to a conclusion reached two studies ago, still takes substantial manual reconstruction.

The challenge is not to produce documents faster. It is to keep the relationships between the data, the findings and the reasoning that gives those findings meaning. That means treating the underlying clinical knowledge, rather than each finished document, as the thing worth keeping connected. Doing that safely with AI is where the real difficulty starts.

Rules can't interpret. Models can't reproduce.

Deterministic rule systems give you reproducibility. Run them against the same source data in ten years and they return the same findings, and you can test that. But they cannot notice that a pattern matters.

Language models can notice. But their output depends on a model version that will be retired long before the product it describes leaves the market. You cannot re-run last year's reasoning and expect the same answer.

Most AI in clinical development picks one side and lives with its weakness. Rules-based automation is defensible but shallow. Generative drafting is impressive but cannot be validated at scale. Neither holds up when an inspector asks where a sentence came from and whether it would be written the same way again.

The answer is not a better model. It is an architecture that keeps these two kinds of work apart.

A hybrid architecture: deterministic facts, agentic judgement

We build on three layers. The boundaries between them are enforced by the system's design, not by a policy document.

The evidence layer is deterministic. Structured source data and defined expert rules establish the findings. Each finding links back to the data it came from, and the rules that produced it are versioned and approved. The same data and the same rules give the same findings every time, however the language used to describe them changes.

The investigative layer is agentic and constrained. Agents read the established findings and the clinical context around them, looking for relationships, inconsistencies and patterns a reviewer would want to see. They can read the evidence layer but cannot write to it. The only thing they can produce is a question for a qualified person to answer.

The language layer only communicates. Generation turns approved findings into the right structure for a CSR, a patient narrative or a lay summary. It writes about findings. It never creates them.

Above all three sits the decision that matters: a qualified person accepts, edits or dismisses what the agents raised. That decision is recorded as data, attached to the finding it concerns, together with the reasoning behind it.

Each layer needs different controls. Reproducible rules give consistency in how findings are established. Agentic interpretation needs grounding and review. Qualified medical writers keep responsibility for the scientific judgements that matter. The aim is not to remove professional judgement. It is to give that judgement a clearer evidential foundation.

This separation is not there to make the system tidier. Everything else depends on it.

Infrastructure, not a tool

In a regulated environment, the expensive part of reuse is not generating content again. It is re-establishing trust in it. Every time a statement moves from a CSR into a submission summary, a safety update or a response to an agency question, someone has to confirm that it is still supported. Today that is done by hand, document by document, and often by people who were not there when the original judgement was made.

When findings are deterministic and traceable, they are validated once, at the level of the finding, and that validation travels with them. A safety finding established for a CSR is the same object when it appears in the patient narrative, the lay summary and next year's integrated analysis. It is not a paraphrase of it that has to be checked again.

The hybrid model also lets the AI change without the evidence changing. Models will be replaced many times over a product's life. When the layers are separate, upgrading the agentic layer changes which questions are raised and how well the prose reads. It does not reopen the validation of the facts underneath.

That is the difference between a tool and infrastructure. A tool's validation expires with its version. Infrastructure keeps its validation and lets you build on it, study after study.

Validating what can't be proven

The deterministic half is the easier part, because regulated industries already know how to validate it. You re-run the rules against fixed inputs, check the outputs against expected results, and control every change. Our evidence layer meets that standard today: every finding reproduces identically from its source data, and every finding traces back to the records it came from.

The agentic half is the open problem, and the industry has mostly avoided talking about it. There is no ground truth for whether a question was worth raising. Experienced safety physicians do not always agree with each other on that, and nobody should expect them to.

So rather than look for a correct answer, the right measure is the disagreement itself. When several experienced reviewers assess the same material independently, the spread of their judgements defines the range within which reasonable experts differ. The system should be judged against that range, not against any single reviewer. If it raises what a credible share of those reviewers would have raised, it is performing like an expert. If it raises things no reviewer would, it is producing noise.

We treat precision as seriously as recall, which is less common than it should be. Recall gets the attention because a missed signal is the failure everyone fears. In practice, the more likely failure is the opposite: a reviewer who receives three unnecessary flags stops reading the fourth. A system that flags everything has not reduced risk. It has handed the risk back to the reviewer and made the review worse.

Validation should not stop at deployment either. Every flag, and every decision taken on it, belongs in the record, so that each review adds to the system's track record. Dismissed flags matter as much as accepted ones. They show where the system's sense of relevance differs from the experts', and the reasons given for dismissing them are clinical knowledge in their own right.

Go back to the liver and kidney case. Nobody can prove the question was correct. The right test is whether experienced reviewers would have raised it, whether the system avoids raising twenty questions like it that nobody needs, and whether a qualified person made the decision on the record. I do not think a stricter standard exists, and I think the industry should be willing to say so.

Connecting reasoning, not just systems

Others are working on related problems. Boston Consulting Group has described multi-agent approaches to biopharma medical writing that connect internal platforms, regulatory systems and scientific information. Novartis has described pairing generative AI with knowledge graphs across R&D, and AstraZeneca's Velocia initiative aims to carry knowledge forward between stages of evidence generation.

Most of these efforts connect systems. The harder problem is connecting reasoning: keeping the judgement behind a conclusion attached to the conclusion, in a form that outlasts the people who made it.

When the same pattern of organ injury appears in the next study, the question should not be raised as if for the first time. The earlier finding, the flag, the reviewer's decision and the reasoning behind it should be retrieved, not reconstructed from old reports. That is what reuse should mean in a regulated industry. It is not copying sentences between documents. It is carrying forward judgement that has already been reviewed.

Where we are today

TriloDocs runs this architecture today across clinical study reports, patient narratives and lay summaries. Carrying reviewed reasoning across documents and studies is what we are building next. I would rather say that plainly than claim it is already done.

Beyond medical writing

None of this is specific to clinical documentation. The pattern is to establish facts deterministically, apply agentic judgement on top of them, keep the two separable, and keep the reasoning. It applies anywhere AI meets a process that has to be defended afterwards: pharmacovigilance, quality, regulatory submissions, and well beyond life sciences.

Documentation is simply where getting it wrong shows up first and most visibly. Organisations that build this separation now will be able to change their models freely while their evidence stays put. Those that don't will find that every model upgrade sends them back to validation.

Continuing the conversation

Gaspar Wong presented these ideas at the CNS Summit in Boston, and the TriloDocs team continued the discussion at the AMWA Medical Writing & Communication Conference in Washington, DC. We welcome conversations with clinical development, medical writing and R&D technology leaders working on how AI can move from document automation to connected, trustworthy clinical intelligence.

Gaspar Wong is Head of Product at TriloDocs, where he builds hybrid deterministic and agentic systems that separate what can be proven from what has been inferred, and works on the harder question of how to validate the second.