An Accelerated Approval submission is an argument that the agency should grant marketing authorization on the basis of an endpoint that is not, itself, the thing patients care about. The drug shrank the tumor, lowered the viral load, cleared the protein, moved the biomarker, and the sponsor is asking the FDA to accept that this surrogate is reasonably likely to predict the clinical benefit that the confirmatory trial has not yet proven. Everything in the dossier hangs on the credibility of that one inferential leap, and since the Food and Drug Omnibus Reform Act of 2022 (FDORA) rewrote the FDA's authority over Accelerated Approval, the leap now comes with a non-negotiable companion: a confirmatory trial that is required to be underway, or specified with a timetable, at the time of approval, with the agency holding new expedited-withdrawal teeth if the sponsor fails to deliver. This lesson builds the surrogate-endpoint defense package end-to-end with AI assistance: the literature corpus, the precedent base of prior Accelerated Approvals, the statistical surrogate-validation argument, the Module 2.5 and 2.7 surrogate narrative, the AdComm-ready surrogate-defense brief, and the confirmatory trial plan that FDORA now demands. The anchor artifacts are the Accelerated Approval package and the confirmatory-trial protocol synopsis filed at the NDA or BLA, and the governing discipline is that AI accelerates retrieval and drafting while the named author owns every link in the inferential chain.

What FDORA Changed and Why the Confirmatory Plan Is Now Load-Bearing

For most of Accelerated Approval's history the central tension was a soft one: sponsors won approval on a surrogate and then, sometimes, took years to complete or even begin the confirmatory study, and the FDA's only remedy was a withdrawal procedure so cumbersome that it was almost never used. FDORA, signed into law in December 2022, hardened that tension into statute. The agency now has explicit authority to require that confirmatory trials be underway at the time of Accelerated Approval, or to specify the conditions and timetable under which they must be conducted, and it gained a streamlined withdrawal procedure for products whose confirmatory trials fail to verify benefit or are not conducted with due diligence. The practical consequence is that the confirmatory trial plan is no longer an appendix the sponsor promises to flesh out later. It is a load-bearing element of the approval itself, scrutinized at filing, and a weak or vague plan can sink an otherwise strong surrogate argument.

This reframes the whole submission. A sponsor team that treats the surrogate defense as the real work and the confirmatory plan as paperwork will be surprised by how hard the reviewing division and the advisory committee press on the plan: its feasibility, its timeline, its endpoint, its power, its enrollment assumptions, and whether it is designed to actually answer the question the surrogate left open. The surrogate argument earns the conditional approval; the confirmatory plan is the condition. AI assistance is genuinely valuable across both, but the value is concentrated in the retrieval-heavy and consistency-heavy parts of the work, and the danger is concentrated exactly where the inferential leap lives, so the design of the AI workflow has to respect that asymmetry.

Building the Surrogate-Validation Evidence Base

A surrogate endpoint is only as defensible as the evidence that it tracks the clinical outcome, and the strength of that evidence falls on a well-understood hierarchy that the reviewing division will apply whether or not the sponsor names it. At the strong end sits a surrogate validated at the trial level, where a treatment's effect on the surrogate has been shown across multiple randomized trials to predict its effect on the clinical outcome, ideally with a quantified surrogacy relationship. In the middle sits a surrogate supported by consistent individual-level association and a coherent biological mechanism. At the weak end sits a surrogate that is merely correlated with the outcome in observational data, which the division generally will not accept as reasonably likely to predict benefit. The first job of the evidence base is to locate the sponsor's surrogate honestly on that hierarchy, because an argument that overclaims its position is the fastest way to lose credibility with a statistical reviewer.

This is where AI retrieval earns its place. Assembling the surrogate-validation evidence base means finding every relevant meta-analysis of the surrogate-outcome relationship, every prior trial that measured both the surrogate and the clinical endpoint, every regulatory precedent where the same or an analogous surrogate was accepted or rejected, and every mechanistic study that supports the biological plausibility of the link. A retrieval-augmented workflow over a curated corpus of the published literature, FDA review documents, and the sponsor's own prior submissions can surface this material far faster than manual search, and it can do so with citations to the specific source for each retrieved claim. But the retrieval is the beginning, not the end. Every retrieved paper must be read by a human for what it actually shows, because the model will happily retrieve a meta-analysis whose conclusion is more equivocal than its abstract suggests, or characterize a trial-level validation as established when the source describes it as exploratory. The corpus gives you reach; the human gives you fidelity.

The Precedent Base, and Why Analogy Is the Strongest and Most Treacherous Tool

Accelerated Approval is a precedent-driven program. The most persuasive thing a sponsor can show a reviewing division is that the agency has already accepted this exact surrogate, or one closely analogous, for an Accelerated Approval in a comparable disease setting, because it converts an abstract argument about reasonable likelihood into a concrete demonstration that the agency itself has drawn the same inference before. Building the precedent base means assembling the prior Accelerated Approvals that relied on the surrogate, the FDA review documents and advisory committee discussions that surrounded them, the cases where the surrogate was accepted, and, just as important, the cases where a similar surrogate was rejected or where a confirmatory trial later failed to verify benefit and the approval was withdrawn. The withdrawals matter because the division reads them too, and a precedent base that cites only the favorable cases is one a reviewer will distrust the moment they recall the unfavorable one the sponsor omitted.

AI-assisted precedent retrieval is powerful precisely because the corpus of Accelerated Approval decisions is large, dispersed across review documents and AdComm transcripts, and tedious to search by hand, and a well-grounded retrieval workflow can surface analogous cases a manual search would miss. It is also where the most dangerous hallucination in this entire workflow lives. A precedent is a factual claim about a specific prior regulatory decision: that a named drug received Accelerated Approval in a named indication on a named surrogate in a named year, with a named confirmatory outcome. A model that fabricates or misremembers any element of that claim produces a precedent that is rhetorically perfect and factually false, and a false precedent in a surrogate-defense brief is catastrophic, because the reviewing division knows the real precedent base better than the sponsor does and will recognize the error instantly. Every precedent must be verified against the primary regulatory record, the approval letter, the review document, the AdComm record, before it enters the brief. A precedent the team cannot verify is treated as wrong until proven right, exactly as an uncheckable TLF citation is.

Drafting the Module 2.5 and 2.7 Surrogate Narrative

With the evidence base and the precedent base assembled and verified, the surrogate narrative threads them into the Module 2.5 Clinical Overview and the Module 2.7 Clinical Summaries as a single coherent argument: this surrogate, in this disease, supported by this mechanistic and statistical evidence and this regulatory precedent, is reasonably likely to predict the clinical benefit, and here is the confirmatory trial that will convert reasonable likelihood into proof. AI is a strong drafting partner here because the argument has a stable structure and the sponsor wants it expressed consistently across 2.5.4, the relevant 2.7.3 efficacy summary, and the surrogate-defense brief, and consistency across documents is exactly what a model maintains well when it is grounded in a single verified source set. The cross-module consistency the model provides is a genuine asset, because an inconsistency between the surrogate claim in 2.5 and the supporting detail in 2.7 is the kind of seam a reviewer pulls.

The danger in the narrative is the inferential connective tissue. The model can state the surrogate result, transcribe the magnitude, and cite the validation literature, but the sentence that says this surrogate is reasonably likely to predict clinical benefit is a judgment, not a retrieval, and it is the single most consequential sentence in the dossier. The model will produce that sentence fluently at whatever strength the surrounding text suggests, including a strength the evidence does not support, because it is completing a pattern and has no independent grasp of how strong the inference actually is. The named author owns the calibration of that claim: whether the evidence supports reasonably likely, or only possibly, or whether it overreaches and should be softened. This is the same boundary the program draws everywhere: the AI structures the argument and maintains its consistency; the human owns the conclusion and its calibration. A surrogate narrative that reads as more certain than the evidence base supports is not a stronger submission. It is a more fragile one, because the gap between the prose and the evidence is exactly what the statistical reviewer and the advisory committee are trained to find.

The Confirmatory Trial Plan FDORA Now Requires

The confirmatory trial plan must do something the surrogate argument cannot: it must lay out a study that will measure the clinical outcome the surrogate stands in for, with enough rigor and on a timetable credible enough that the agency will accept it as the condition of approval. The plan and its protocol synopsis specify the clinical endpoint, the design, the population, the comparator, the sample size and power, the timeline to key milestones and to final readout, and the relationship between the confirmatory population and the Accelerated Approval population. Under FDORA's hardened posture, the agency expects this trial to be underway or to have a concrete, near-term timetable at the time of approval, and a plan that proposes to begin enrollment at some unspecified future point invites the division to question whether the sponsor is serious about verifying benefit at all.

AI assistance in the confirmatory plan is most valuable in the consistency and completeness dimensions. The model can check that the confirmatory endpoint is the clinical outcome the surrogate was argued to predict, not a second surrogate, which is a recurring deficiency. It can check that the population and comparator are consistent with the Accelerated Approval population and with the broader development program. It can draft the protocol synopsis in the structure the reviewing division expects, aligned to ICH M11 where applicable, and it can cross-check the synopsis against the surrogate narrative so the two documents tell one story. What the model cannot do is decide whether the trial is feasible, whether the timeline is honest, whether the power assumptions are defensible, or whether the design will actually answer the question, and those are precisely the judgments the division and the advisory committee will press hardest. A confirmatory plan is a commitment with statutory consequences for failure, and the sponsor's named clinical and biostatistics leads own every parameter in it. The model helps ensure the plan is complete and consistent; it does not own whether the plan is real.

The AdComm-Ready Surrogate-Defense Brief

Many Accelerated Approvals on novel or contested surrogates go to an advisory committee, and the surrogate-defense brief is the artifact that has to survive a room of external experts whose explicit charge is to probe the inferential leap. The brief is not a softer version of the Module 2.5 argument; it is a sharper one, built to anticipate and answer the hardest questions a skeptical committee will ask. Is the surrogate validated at the trial level or only associated at the individual level? What is the magnitude and durability of the surrogate effect, and is it clinically meaningful or merely statistically detectable? What does the precedent base actually show, including the cases where a similar surrogate failed to predict benefit? Is the confirmatory trial designed and powered to verify, and is it credibly underway? AI assistance, including the persona-engineered reviewer technique from the prior lesson, is well suited to generating the anticipated-question set and stress-testing the brief's answers against the worst reasonable version of each.

The brief must also honestly present the uncertainty, because an advisory committee that senses a sponsor hiding the weakness of its surrogate becomes adversarial in a way that is very hard to recover from. The most defensible surrogate-defense brief states clearly where the evidence is strong, where it is moderate, and where it depends on the confirmatory trial, and it frames the Accelerated Approval as exactly what the statute intends: a conditional grant that the confirmatory trial will resolve. The model can help draft this honestly calibrated brief, but the calibration itself, the decision about how strongly to claim and where to concede, is the named regulatory and clinical leadership's judgment, made with knowledge of the division's current posture and the committee's likely composition that no model possesses. The brief is where the surrogate argument meets its hardest audience, and the human owns how that argument is pitched.

The Audit Trail for a Conditional Approval the Agency Can Now Withdraw

Because an Accelerated Approval is conditional and FDORA gives the agency expedited authority to withdraw it if the confirmatory trial fails or is not conducted with due diligence, the audit trail around the AI-assisted surrogate package carries unusually long-lived consequences. The confirmatory commitment can be revisited years after approval, and the sponsor may need to demonstrate, long after the original team has moved on, that the surrogate argument was built on verified evidence and genuine precedent rather than on fluent fabrication. Every AI-assisted step in the package inherits the ALCOA+ and 21 CFR Part 11 obligations: the retrieval runs that assembled the evidence and precedent bases, with the sources actually surfaced and the human verification of each; the drafting runs that produced the surrogate narrative, with the model, version, temperature, and loaded sources captured; and the named human sign-off on every calibrated inferential claim and every confirmatory-plan parameter.

This matters with particular force for the precedent base and the surrogate-validation claims, because those are the elements most likely to be re-examined if the confirmatory trial disappoints. A sponsor that can show, years later, that every precedent in its surrogate-defense brief was verified against the primary regulatory record at the time, and that every surrogate-validation claim traced to a human-checked source, stands on solid ground even if the confirmatory trial ultimately fails to verify benefit, because failing to verify is a permitted outcome of a genuine conditional approval, whereas building the original argument on an unverified or fabricated foundation is a credibility and integrity failure that taints everything. The audit trail is what distinguishes an honest conditional approval that did not pan out from a submission that overclaimed. Build the surrogate package so that every inferential link, every precedent, and every confirmatory commitment would read as verified and human-owned to a reviewer examining it years after the fact, because under FDORA, one of them might.

Key Takeaways

• FDORA made the confirmatory trial plan load-bearing, not paperwork. The FDA now has authority to require confirmatory trials to be underway at the time of Accelerated Approval and a streamlined procedure to withdraw approvals whose trials fail or are not conducted with due diligence, so a vague confirmatory plan can sink an otherwise strong surrogate argument. The surrogate argument earns the conditional approval; the confirmatory plan is the condition.
• Locate the surrogate honestly on the validation hierarchy, because overclaiming is the fastest way to lose a statistical reviewer. Trial-level validation is strong, individual-level association plus mechanism is moderate, mere observational correlation is generally not accepted. AI retrieval gives you reach across the meta-analyses and mechanistic studies; the human gives you fidelity by reading what each source actually shows rather than what its abstract suggests.
• The precedent base is the strongest tool and the most dangerous hallucination site in the workflow. A precedent is a factual claim about a named prior decision, surrogate, indication, year, and confirmatory outcome; a fabricated or misremembered precedent is rhetorically perfect and factually false, and the reviewing division knows the real precedent base better than the sponsor. Verify every precedent against the primary regulatory record, and include the unfavorable cases.
• The model structures and keeps the surrogate narrative consistent; the human owns the inferential connective tissue. The sentence that says the surrogate is reasonably likely to predict clinical benefit is a calibrated judgment, not a retrieval, and the model will write it at whatever strength the surrounding text suggests, including one the evidence does not support. A narrative that reads as more certain than the evidence is more fragile, not stronger.
• Build the audit trail for a conditional approval the agency can now withdraw years later. Capture every retrieval and drafting run under ALCOA+ and 21 CFR Part 11, with verified sources and named sign-off on every inferential claim and confirmatory parameter, so that an honest conditional approval that did not pan out is cleanly distinguishable from a submission that overclaimed on an unverified foundation.