A clinical trial manager opens a spreadsheet of 60 protocol deviations logged across 12 sites in the last quarter and has to do three things with it before the steering committee meets. She has to classify each deviation as major or minor against the protocol's own criteria, because that label determines reporting, escalation, and whether a participant's data can be used. She has to draft a corrective and preventive action narrative for every major deviation, because a major deviation without a documented CAPA is a finding waiting to be written by an inspector. And she has to write the Quality Tolerance Limit excursion memo, because the deviation pattern has pushed one parameter past the threshold the trial defined in advance under ICH E6(R3). An enterprise large language model can triage the 60, propose classifications, and draft the narratives faster than she could read the log twice. It can also confidently mislabel a major deviation as minor, draft a CAPA that treats a symptom as a root cause, and produce a QTL memo that reads correct and reasons wrong. The major/minor judgment and the ownership of the CAPA stay human, and this lesson is about why those two boundaries do not move.

Why the Major/Minor Line Is the Whole Game

A protocol deviation is any departure from the approved protocol, and the single most consequential thing done with one is its classification as major or minor, because nearly everything downstream hangs on that label. A major deviation, sometimes called important or significant, is one that affects, or has the potential to affect, participant safety, the integrity of the study data, or the rights and welfare of participants, and it carries reporting obligations, may require corrective action, and can affect whether the affected data are included in the analysis. A minor deviation is a departure that does not rise to that threshold. The line between them is not a matter of taste; it is a judgment about safety, data integrity, and participant welfare, made against the specific protocol's own deviation-handling criteria.

This is the heart of why classification cannot be delegated to the model. Determining whether a particular deviation, a dose given a day late, a missed laboratory assessment, an out-of-window visit, an eligibility criterion not met, affected or could have affected safety or data integrity requires reasoning about the specific protocol, the specific participant, and the specific clinical context, and that reasoning is exactly what a pattern-completing model does not perform. The model can recognize the shape of a deviation and produce a plausible major-or-minor label, and the label will be wrong often enough, and confidently enough, that accepting it is a clinical-judgment failure dressed up as efficiency. The named clinical trial manager owns the classification, because the classification is a safety and integrity determination, not a text-generation task.

The protocol-specific dimension is what makes generic AI classification particularly hazardous here. Two trials can treat the same nominal deviation differently: an out-of-window visit that is minor in a trial with flexible visit windows can be major in a trial where the visit timing is critical to the endpoint, and only the protocol's deviation-handling plan resolves which it is. A model classifying from the deviation description alone, without reasoning from the specific protocol's criteria, applies a generic notion of severity that may directly contradict the trial's own definition, producing a tidy classification that is wrong by the only standard that matters. The human classifies against the protocol's criteria, every time, and treats the model's proposed label as a prompt to check the criterion, not as an answer.

Where AI Genuinely Helps: Triage, Not Judgment

Saying the model cannot own classification does not mean it has no role; it means its role is triage and drafting support, not the safety judgment itself. Across 60 deviations and 12 sites, there is real, tedious work the model does well. It can read the deviation log and organize it, grouping deviations by type, by site, by protocol section, surfacing the patterns a human would otherwise assemble by hand. It can flag the deviations that share characteristics with clearly major categories, the safety-relevant ones, the eligibility violations, the consent issues, so the human's attention goes first to the ones most likely to need it. It can draft a first-pass description of each deviation in consistent language that the human then classifies.

This triage role is genuinely valuable because it inverts the time problem. Without AI, the CTM reads all 60 deviations at the same depth to find the few that matter, spending most of her attention on the obviously minor. With AI doing the organizing and flagging, she can spend her judgment where judgment is needed, on the deviations whose classification is genuinely uncertain or whose severity is high, while the model handles the assembly that does not require clinical reasoning. The acceleration is real, and it is the right kind of acceleration, because it concentrates rather than replaces the human judgment.

The discipline is to keep the boundary sharp between what the model organizes and what the human decides. A model that groups and flags is an assistant; a model whose proposed classifications are accepted because they look reasonable has crossed from triage into judgment, and the crossing is silent because a proposed label and a verified label look identical on the page. The CTM uses the model to bring structure and candidate flags to the 60 deviations, and then performs the major/minor determination herself against the protocol, treating the model's organization as a workspace and never as a verdict. The flag says look here; the classification is hers.

Drafting the CAPA Without Confusing Symptom for Cause

For every major deviation, a corrective and preventive action narrative documents what went wrong, why, and what will be done to correct the immediate problem and prevent its recurrence. The CAPA is where the trial demonstrates that it does not merely log problems but responds to them, and a weak CAPA is one of the most common inspection findings, because inspectors read CAPAs to judge whether the sponsor and site actually understand their own failures. The structure is well defined: a description of the deviation, an investigation into its root cause, the corrective action that addresses the immediate instance, and the preventive action that addresses the underlying cause so it does not recur.

The model drafts this structure fluently and falls into a specific, dangerous trap at the center of it: the root cause. A genuine root-cause analysis reasons from the deviation back to the underlying systemic failure, the training gap, the process ambiguity, the resourcing problem, that allowed it to happen, and that reasoning requires understanding what actually occurred at the site, which the model does not have. Asked to draft a root cause, the model produces a plausible one, and a plausible root cause is frequently a restatement of the symptom: a deviation where a visit was missed gets a root cause of the visit being missed, dressed in causal language, with a corrective action that fixes nothing because it never reached the actual cause. The human owns the root-cause determination, because a CAPA built on a symptom-as-cause is a CAPA that will fail to prevent recurrence and will be read by an inspector as a sign the failure is not understood.

The preventive action inherits the same risk. A preventive action is only as good as the root cause it addresses, so a model that misidentified the root cause will draft a preventive action that prevents the wrong thing, and it will read as a complete, responsible CAPA while doing nothing to stop the deviation from recurring. The CTM, working from what actually happened, supplies the genuine root cause and confirms the preventive action addresses it, using the model to draft the narrative structure and consistent language once the causal reasoning is human-owned. The CAPA ownership stays with the human not because the model writes badly but because the model cannot know why the deviation happened, and a CAPA is fundamentally an answer to that question.

The QTL Excursion Memo Under ICH E6(R3)

Quality Tolerance Limits are a feature of the quality-by-design approach to clinical trials that ICH E6(R3) reinforces: at the study level, the sponsor defines in advance the parameters that matter to participant safety and data reliability and sets tolerance limits, thresholds that, when crossed, signal that the trial is operating outside its expected quality range and warrant investigation. The protocol-deviation rate is a common QTL parameter, and when the pattern across the 60 deviations pushes the deviation rate past the predefined limit, a QTL excursion has occurred and must be documented in a memo that records the excursion, investigates whether it reflects a genuine systemic problem, and describes the response.

The QTL excursion memo is a reasoning document, and that is where the model's limits bite. The hard question in a QTL excursion is not whether the threshold was crossed, which is arithmetic, but what the excursion means: whether it reflects a real systemic quality problem requiring action, or an artifact of how deviations were counted, a single aberrant site, or a threshold set too tight, and that interpretation is a judgment about the trial's actual quality state. A model drafting the memo can produce a fluent narrative that asserts a cause and a significance for the excursion, and that narrative reads authoritative while being a generated interpretation the underlying data may not support. The human owns the interpretation of what the excursion means, because the entire purpose of the QTL system is to convert a threshold crossing into an informed quality judgment, which is precisely what the model cannot supply.

There is a documentation precision the model also tends to miss. A QTL excursion memo under E6(R3) should tie the excursion to the predefined limit and the rationale for that limit, describe the investigation actually performed, and connect the response to the trial's risk-based quality management, rather than presenting a generic quality narrative. A model drafting from the deviation pattern alone can produce a memo that describes an excursion without grounding it in the predefined QTL framework the trial established, leaving the memo disconnected from the quality system it is supposed to document. The CTM grounds the memo in the trial's actual QTL definitions and the investigation she conducted, using the model for structure and language once the quality judgment and the investigation are hers.

The Stakes of a Misclassification, Traced Forward

It is worth tracing what a single accepted misclassification does, because the abstraction of major versus minor hides a concrete chain of consequences. Suppose the model labels a deviation minor that is, by the protocol's criteria, major: a participant was dosed despite meeting an exclusion criterion, but the model, reasoning from the deviation description alone, treats it as a routine eligibility note. Because it is logged minor, it does not trigger the reporting the major classification requires, it does not get a CAPA, and it does not feed the safety review that a major safety-relevant deviation would. The deviation that should have prompted a question about participant safety instead sits quietly in the log as one of the routine 60.

The chain runs further. The misclassification suppresses the deviation's contribution to the QTL parameter, so the deviation-rate excursion may be understated, which means the systemic problem the deviations signal is masked at exactly the level designed to catch it. When an inspector later reviews the deviation log against the protocol's classification criteria, the misclassified deviation is visible as a deviation that should have been major and was not, and that single finding does what one fabricated citation does in a submission: it shifts the inspector to distrust the entire classification process, because a function that misclassified one safety-relevant deviation may have misclassified others. The credibility cost of the misclassification exceeds the single case, exactly as the credibility cost of a fabricated TLF reference exceeds the single table.

This is why the human-owned classification is not a conservative ritual but the load-bearing control of the whole workflow. Every downstream artifact, the CAPA, the QTL memo, the safety review, the inspection-readiness of the deviation log, depends on the major/minor label being right, and the label is right only when a human made it against the protocol's criteria. The model's speed in triaging the 60 is worth a great deal when it concentrates the human's judgment on the classifications that matter, and worth less than nothing when it substitutes a plausible label for that judgment.

The Audit Trail and the Two Human-Owned Boundaries

The deviation log, the CAPAs, and the QTL memo are controlled records that live in the trial master file and the sponsor's quality system, and an AI-assisted version of each must carry the same defensibility as any other. The record should capture how the AI was used: that the model triaged and organized the 60 deviations and drafted narrative structure, and that a named human performed the major/minor classification against the protocol, determined each root cause, and interpreted the QTL excursion. This is the ALCOA+ discipline applied to deviation management, and it draws the line in the audit trail exactly where the lesson draws it: the model assisted the assembly, the human owned the judgment.

Two boundaries in particular must be unmistakable in the record. The first is the classification boundary: every major/minor determination is attributable to a named human who made it against the protocol's criteria, not to the model that proposed it. The second is the CAPA ownership boundary: every root cause and preventive action is owned by a named human who reasoned from what actually happened, not by the model that drafted the narrative. When an inspector asks how the deviations were classified and how the CAPAs were determined, the defensible answer names a human at each judgment point and shows the AI as a drafting assistant, and the audit trail is the evidence that the two boundaries held. A record that cannot show where the human judgment entered cannot defend the classifications it contains.

What This Means for the CTM on Monday

Deviation management is a strong fit for AI assistance precisely because the volume is high, the assembly is tedious, and the judgment is concentrated in a minority of cases, which is the pattern where triage support pays off most. A CTM facing 60 deviations across 12 sites reclaims real time by letting the model organize the log, flag the likely-major cases, and draft the narrative structure of the CAPAs and the QTL memo. But the time is reclaimed only because the model concentrates her judgment, not because it replaces it, and the two boundaries that cannot move are the two on which everything downstream depends.

So she lets the model triage but classifies every deviation herself against the protocol's criteria, treating each proposed label as a flag to check rather than an answer to accept. She lets the model draft the CAPA structure but determines every root cause from what actually happened, rejecting the symptom-as-cause the model gravitates toward, and confirms each preventive action addresses the real cause. She lets the model draft the QTL memo but owns the interpretation of what the excursion means, grounding it in the trial's predefined limits and the investigation she conducted. She captures the run so the audit trail shows the classification and the CAPA ownership were human, and then she signs. The model triaged the 60 in minutes and drafted the narratives; the named CTM made the safety judgments, and the major/minor line and the CAPA ownership never left her hands.

Key Takeaways

• The major/minor classification is a safety, data-integrity, and participant-welfare judgment made against the specific protocol's criteria, and it cannot move to the model. The model produces a plausible label by recognizing a deviation's shape, but the same nominal deviation can be major in one trial and minor in another, so a generic classification can directly contradict the trial's own definition. The named CTM classifies against the protocol every time.
• AI's real role is triage, not judgment: organizing, grouping, and flagging the 60 deviations to concentrate the human's attention. This is the right kind of acceleration because it inverts the time problem, letting the CTM spend judgment where judgment is needed. The boundary is sharp: a model that groups and flags is an assistant, and a model whose proposed labels are accepted has silently crossed into judgment.
• A CAPA's root cause is the trap, because the model restates the symptom as the cause in causal language. A root-cause analysis reasons back to the underlying systemic failure, which requires knowing what actually happened at the site, and the model does not. A preventive action built on a misidentified root cause prevents the wrong thing while reading as a complete CAPA, so the human owns the root cause and confirms the preventive action addresses it.
• The QTL excursion memo is a reasoning document, and the hard question is what the excursion means, not whether the threshold was crossed. A model produces a fluent interpretation the data may not support, so the human owns whether the excursion reflects a real systemic problem and grounds the memo in the trial's predefined QTL limits and the investigation actually performed under ICH E6(R3).
• A single accepted misclassification propagates: it suppresses reporting, skips a CAPA, masks a QTL excursion, and shifts an inspector to distrust the whole classification process. The credibility cost exceeds the single case, exactly as a fabricated TLF reference does in a submission. The audit trail must show two unmistakable boundaries: a named human owns every classification and every root cause, with the model as a drafting assistant.