Once FDA read with AI, sponsors’ inconsistencies got cheaper to find

The week

On June 2, 2025, the Food and Drug Administration announced the agency-wide launch of Elsa, a generative AI tool meant to help employees, from scientific reviewers to investigators, work more efficiently. For decades, a drug company’s application to FDA had been read the old-fashioned way: by expert humans working through thousands of pages, sampling what they could. Now the reader had an assistant.

Generative AI is the kind of software behind chat assistants: you type a request in ordinary words, and it produces text, summaries or code in response. It has learned patterns from enormous amounts of writing, which makes it fast and fluent, though not always right.

FDA said Elsa could help read, write and summarize documents, compare labels, summarize adverse events (reports of side effects and other problems patients experience) and generate code for internal databases. It said the tool could also help identify high-priority inspection targets, meaning which manufacturing sites or clinical study locations most deserve a visit.

The agency addressed the obvious worry about confidential company data. It said Elsa ran in a secure government cloud environment, computing capacity rented from a provider but walled off to meet federal security standards, and that it did not train on data submitted by regulated industry. So a company’s secrets were not being fed into the tool to make it smarter for everyone else.

The launch had a short run-up. On May 8, 2025, FDA said it had completed its first AI-assisted scientific review pilot. Commissioner Marty Makary directed every center to begin deploying generative AI immediately, with full integration by June 30.

After that date, the agency said, all centers would work on one common, secure generative AI system connected to FDA’s internal data platforms. It named Jeremy Walsh as its new chief AI officer, coordinating the rollout with Sridhar Mantha, who had led business informatics in the Center for Drug Evaluation and Research.

The May announcement also carried the claim that framed the effort. A deputy director in the drug center’s Office of Drug Evaluation Sciences said the pilot tools let him finish some review tasks in minutes that had previously taken three days. One reviewer’s account of particular tasks is not a measured agency-wide result, but it showed where leadership expected the gains.

On June 2, FDA said Elsa had arrived ahead of that schedule and under budget, built with in-house experts across its centers. It said staff were already using the tool to speed clinical protocol reviews and shorten scientific evaluations. It described Elsa as a first step, with more AI to follow in areas such as data processing.

The speed was itself a signal. Elsa was a working tool, in the hands of the people who decide whether medicines reach the market, rather than a distant research project.

Two pieces of context shaped how the industry read the news. In late March 2025, the Department of Health and Human Services had announced a restructuring that included cutting about 3,500 FDA jobs. HHS said the cuts would not touch drug, device or food reviewers or inspectors. Even so, a tool meant to make each employee more productive arrived as the agency was getting smaller.

The second piece was FDA’s own rulebook. Five months earlier, in January 2025, it had published draft guidance on how sponsors should establish the credibility of AI models used to support regulatory decisions. By June the agency was both a user of generative AI and the body setting expectations for everyone else’s.

The announcement left a good deal unsaid. It did not name the underlying model, describe how reviewers were trained to use it, or explain how Elsa’s outputs would be recorded in review files. Sponsors had to reason from what such tools generally do.

For life sciences executives, the week changed a familiar question. The industry had long asked how it should use AI. Now it also had to ask what happens when the regulator reading its work uses AI too.

Life sciences companies have spent years asking how they should use AI. Elsa reversed the direction of the question. From June 2025, the people reviewing a submission had a generative AI assistant at hand.

FDA did not say Elsa made decisions, and sponsors cannot see how any reviewer used it. (A sponsor is the company or organization that takes responsibility for a clinical study and applies for approval of a product.) They can still reason about what changes when the reader has a fast tool for summarizing, comparing and searching.

We see three consequences, each with a different owner. The first is what a machine reader surfaces in the dossier. The second is how a sponsor corrects a machine misreading. The third is how the sponsor’s own AI use will be judged by a regulator that now uses the same technology. All three follow from what these tools do well and what they do badly.

A fast, well-read and fallible reader

A generative AI tool of the kind FDA described behaves like an extremely fast, well-read research assistant who has never been in your building. Give it a stack of documents and a question, and within seconds it will find relevant passages, summarize them, line them up side by side or draft a response. It does not get tired at page 4,000, and it does not skim.

Its strengths are breadth and speed. Asking “show me every place this document mentions the starting dose” is trivial for it, whereas a human would need hours with a highlighter. Comparing two versions of a drug label, or pulling every reported side effect into one list, is the kind of chore these tools handle well.

Its weakness has a name: hallucination. Because the tool generates text by predicting what plausibly comes next, it can produce statements that sound authoritative and are simply wrong. It might merge two studies into one, flip a finding, or cite a page that does not exist. It does this with the same confident tone it uses when it is right. That is why responsible use always keeps a human checking the output, and why FDA described Elsa as an assistant to staff rather than a decision-maker.

How a document assistant finds things

FDA did not describe Elsa’s internals. Assistants built to answer questions about large document sets commonly work in two steps, though. First, a search step finds the passages most likely to be relevant to the question. Second, the language model reads those passages and writes an answer from them.

That design explains both the strengths and the typical errors. When the search step finds the right passages, the answer is fast and usually sound. When it misses them, the model still writes an answer, built from whatever it did retrieve. The reviewer sees fluent text either way.

Several features of a real dossier make misses more likely:

• Numbers locked in images. A table pasted as a picture, or a scanned legacy report, may be invisible to text search even when a human can read it easily.
• Similar names. Two main studies with neighboring protocol numbers and near-identical titles are easy for a search step to confuse, and easy for a summary to blend.
• Superseded versions. A review file accumulates amendments. A passage from an earlier sequence can be retrieved alongside the current one, and the summary may not say which is which.
• Meaning that depends on distance. A footnote three pages from its table, or a population definition in an appendix, may never reach the model with the number it qualifies.

Regulators know these problems well. FDA’s electronic submission specifications have long asked for text-searchable PDFs, bookmarks and consistent document structure, so that humans can find their way around. A machine reader raises the payoff from the same hygiene. A sponsor that follows those conventions carefully is making its dossier easier to search for every kind of reader.

Those two traits set the agenda for sponsors. A tool like Elsa makes it much cheaper to find genuine inconsistencies in a submission. It can also, occasionally, create apparent problems that are not really there. A well-prepared sponsor needs to be ready for both.

What a machine reader surfaces in the dossier

A marketing application in the Common Technical Document format can run to many thousands of pages across five modules. The Common Technical Document, or CTD, is the standard structure regulators in many countries use for drug applications. It has one module for regional administrative information, one for high-level summaries, and others for manufacturing quality, nonclinical (laboratory and animal) studies and clinical studies in people.

A human reviewer samples it. A reviewer with a summarization tool can ask for every mention of a dose, an endpoint or an exclusion criterion and see them side by side in seconds. An endpoint is the outcome a study measures to judge whether a drug works, such as blood pressure at twelve weeks. An exclusion criterion is a rule for who may not join a study.

That changes the economics of finding inconsistency. A safety population defined one way in the clinical summary and another way in a study report used to take a careful reader to notice. Now it can surface from a single prompt.

Where inconsistency usually enters

Most inconsistencies in a dossier are not errors of fact. They come from the way submissions are built. A clinical study report is often written by a contract research organization months or years before filing. Statisticians generate tables, listings and figures from locked datasets. Medical writers then summarize those outputs for the clinical overview and the summaries of safety and efficacy.

Each hand-off is a chance for a definition to drift. Data cutoffs differ legitimately. An integrated safety summary may pool studies with a later cutoff than any single study report. For a new drug application, FDA rules also call for a safety update after submission, the so-called 120-day safety update, which brings in still newer data.

Coding dictionaries add another source of drift. Adverse events are usually coded to MedDRA, the standard medical terminology for regulatory work, which is updated twice a year. Studies coded years apart may use different versions. A term that moved between categories can make counts disagree across documents even when the underlying patients are the same. Pooled analyses normally recode to one version, and the dossier should say which.

Each of those differences can be correct and documented. The trouble starts when the reader cannot see the reason for a difference without hunting for it. A search tool that puts the two numbers side by side removes the hunting, and makes the missing explanation the most visible thing on the screen.

Two safety populations, one dossier

Suppose a mid-sized drug company files for approval of a new treatment. In its clinical summary, the “safety population” is described as every patient who received at least one dose. In one study report, a late edit changed the definition to every patient who received at least one dose and had one follow-up visit. The numbers in the two places differ by a handful of patients. Neither definition is wrong, and the difference is explained in an appendix. But a reviewer who asks a tool to list every safety population count will see two numbers next to each other, and will reasonably ask why. The company now spends time on a question it could have prevented with one consistency check.

The practical response

The practical response sits in regulatory operations, the team that assembles and publishes the submission. Sponsors can keep a controlled vocabulary for each program: one definition per endpoint, population and analysis set, used in every module. A controlled vocabulary is simply an agreed dictionary that everybody must use, so the same thing is always called and defined the same way.

They can run their own automated cross-module checks before filing, comparing numbers in summaries against source tables. They can treat the proposed label, the integrated summary of safety and the clinical overview as one linked set, with a change log showing when a figure was updated everywhere.

The advice is not new. What has changed is the cost of skipping it, because the reviewer’s cost of catching it has fallen.

There is a trade-off. Rigid consistency checks slow late-stage document work, and submissions are often assembled under deadline. A sensible compromise is to aim automated checks at the facts most likely to be compared: doses, sample sizes, adverse event counts and endpoint definitions. Stylistic variation can wait.

How the exposure differs by sponsor

The same tool changes different things for different kinds of applicant.

Large pharmaceutical companies usually have in-house regulatory operations, publishing systems and standard templates across programs. Their risk is scale. A company filing many submissions a year reuses text from one program in another, and old definitions travel with it. Their best control is a shared library of approved definitions, with program-level overrides that are recorded rather than improvised.

Small and mid-sized biotechs filing a first application often own the fewest of their own documents. A contract research organization wrote the study reports, a vendor publishes the submission, and consultants draft the summaries. The consistency check has to be written into those contracts. A vendor that publishes the dossier can run cross-document comparisons, but only if the sponsor asks for them and defines what they cover.

Generic drug applicants face the most direct effect. Their proposed labeling must match the reference listed drug’s labeling, apart from permitted differences, and the application includes a side-by-side comparison. FDA named faster label comparison as one of Elsa’s uses. A generic sponsor whose comparison explains every difference, such as a carved-out indication or a different inactive ingredient, gives the reviewer nothing to chase.

Device makers are in scope too, because Elsa was deployed across the agency rather than in the drug centers alone. A premarket notification argues that a device is substantially equivalent to a predicate device, usually through comparison tables. Those tables invite the same machine cross-check against test reports and labeling elsewhere in the file.

Manufacturers and contract manufacturers face a different exposure. FDA said Elsa could help identify high-priority inspection targets. Selection models of that kind work from the data the agency already holds, such as compliance history, product risk and reported problems. A site whose own filings and correspondence are consistent and current is less likely to stand out for the wrong reasons.

Measuring whether the checks work

A consistency program needs a small set of measures, or it decays into a checklist nobody reads. Useful ones include the number of cross-module discrepancies caught before filing. Another is the share resolved by fixing the document rather than by adding an explanation. A third is the number of review questions that trace back to an inconsistency. The last measure is the one leadership should watch. It shows whether the effort before filing is reducing the work during review.

How a sponsor corrects a machine misreading

Generative tools produce confident errors. A summary can merge two studies, invert a finding or cite a passage that does not exist. FDA described Elsa as an assistant to staff, and reviewers remain responsible for their conclusions. Even so, a sponsor may receive an information request, a formal question from the review team during the review, that appears to rest on a misreading. It will rarely be told whether a tool was involved.

The defense is an evidence trail that makes correction fast. That means three things:

• A submission index mapping every key claim to its location by module, section and table, maintained by the regulatory team through the whole review cycle.
• Response templates that quote the source passage instead of paraphrasing it, so the reviewer sees exactly what the submission says.
• Answers with citations a reviewer can verify in one step, rather than explanations that require hunting.

A sponsor that can say the figure appears in a named table of Module 2.7.4 and matches the study report resolves a misunderstanding in one round. A sponsor that has to reconstruct where a number came from loses days, and review clocks do not stop for that. Those days matter: review timelines drive launch dates, and launch dates drive revenue forecasts, manufacturing plans and investor expectations.

Two ways to answer the same question

Return to the same company. Midway through its review, it receives an information request asking why the rate of a particular side effect is reported as higher in its overall safety summary than in its main study. The question seems to assume the two figures describe the same group of patients. They do not: the summary pools several studies, including a longer one with more follow-up time, which naturally captures more events.

In the first version of this story, the team has no index. A medical writer spends a day tracing where each figure came from, a statistician spends another day confirming the pooling method, and the response, when it goes out, explains the difference in a long paragraph of paraphrase. The reviewer, reading carefully, has to find the right tables to confirm it. A follow-up question arrives. A week has gone.

In the second version, the regulatory lead opens the submission index, finds both figures and their source tables in minutes, and writes a short reply. It quotes the definition of the pooled population word for word. It names the table in the summary module where the pooled rate appears and the table in the study report where the single-study rate appears. It points to the section explaining the pooling method. The reviewer checks three references and closes the question. Nobody needed to know or care whether software played a part in the original question.

The second version is no cleverer than the first, only better organized, and the organizing was done months earlier, when there was no time pressure. The same pattern holds throughout: the work that pays off during review is almost always done before filing.

Review clocks and the cost of a second round

The time pressure is structural. Under the prescription drug user fee program, FDA commits to act on most standard applications for new molecular entities within ten months of the 60-day filing date. Priority applications get six months from that date.

Within that window, review teams send information requests as questions arise. For new molecular entities, the program adds a mid-cycle communication and a late-cycle meeting where the agency shares significant issues. A misreading that survives to the mid-cycle call has become a review issue, with its own place on the agenda.

The sponsor’s answers also change the clock. A major amendment submitted late in the cycle can extend the goal date by three months. A sponsor that responds to a misunderstanding by submitting large new analyses, rather than pointing to what is already in the file, risks turning a clarification into an extension.

The mechanics of a reply favor the organized sponsor. Responses travel as new sequences in the electronic submission, so a reply can point to a document already on file by its exact location instead of sending it again. A citation to an existing table adds nothing new for the reviewer to assess. A fresh reanalysis, however well meant, is new material that has to be reviewed in turn.

The organizing work therefore has a clear owner and a clear deliverable. The regulatory lead owns the index. The medical writing team owns the source quotations. Biostatistics confirms that each cited table is the current version. A response that passes through those three hands quickly is worth more than a long one that is perfectly worded.

Sponsors should also avoid the opposite error of assuming every odd question came from a tool. Speculating about AI in correspondence with a regulator helps nobody and can sour a working relationship. The goal is a record that works whoever, or whatever, did the reading.

The strongest objection

The best argument against everything above is that FDA reviewers were always thorough. Review divisions include clinicians, statisticians, pharmacologists and chemists who know where inconsistencies hide. FDA also said Elsa’s use was optional. On that view, a summarization tool changes little, and sponsors should not reorganize around it.

Much of that is right. Expert reviewers found inconsistencies long before 2025, and the controls described here were sound practice then. The argument does not depend on every reviewer using Elsa on every file. It depends on the cost of a search falling for anyone who chooses to run one. When finding a discrepancy drops from hours to seconds, more discrepancies get found, by humans and machines together.

The objection does mark the limit of the claim. Sponsors cannot see how review teams used the tool. The case for these controls rests on what they save in any review, with or without AI. Elsa raised the cost of skipping them. It did not create the need.

How the sponsor’s own AI use will be judged

In January 2025, FDA issued draft guidance on using AI to support regulatory decision-making for drugs and biological products. Draft guidance is FDA’s published thinking on a topic, offered for comment before it is finalized; it is not binding law, but companies treat it as a strong signal.

The guidance proposed a risk-based credibility assessment, a structured way of showing that an AI model can be trusted for a particular job. The steps are: define the question the model addresses, define its context of use (exactly what it is used for and how its output feeds decisions), assess model risk, then plan, carry out and document the work that establishes credibility for that context.

The guidance targets AI that produces information supporting decisions about safety, effectiveness or quality. It does not cover AI used only for operational efficiency, such as drafting assistance. It also leaves out AI used in drug discovery, before a candidate enters the regulated development path.

How the draft guidance scores model risk

The draft sets out seven steps. A sponsor defines the question of interest, defines the context of use and assesses model risk. It then develops a plan to establish credibility, executes the plan, documents the results with any deviations, and decides whether the model is adequate for that use.

Model risk combines two factors. Model influence is how much the AI’s output contributes to the decision, compared with other evidence. Decision consequence is how bad an incorrect decision would be. A model whose output is the sole basis for excluding patients at high risk of a serious side effect scores high on both. A model that flags records for a human to look at, with other checks behind it, scores lower.

The score drives the evidence. Higher-risk uses call for more extensive testing, more documentation and more engagement with FDA. The draft encourages sponsors to talk to the agency early about uses they are unsure of, through the existing meeting channels for development programs.

The draft also expects life cycle maintenance. Models can change as their inputs change, a problem engineers call drift. A model that classified data well at launch may degrade when a new site, a new data source or a new product version arrives. The sponsor needs a plan for monitoring performance and for deciding when changes require new credibility work.

That boundary matters, and it blurs easily in practice. A tool that drafts narrative text is an efficiency aid. A tool that codes adverse events (sorts free-text side effect reports into standard medical categories), selects patients from real-world data (information from health records, insurance claims and registries rather than controlled trials) or generates tables for the integrated safety summary is producing evidence.

Sponsors need an inventory that places each AI use on one side of that line, with an owner in quality assurance who signs off the classification and revisits it when the use changes. A drafting tool that someone later starts using to pull numbers into tables has crossed the line without anyone deciding it should, and the inventory should catch that.

Where AI enters the evidence chain

An inventory usually finds AI in more places than the AI strategy document lists. Common candidates include:

• Clinical operations. Screening health records to find eligible patients, or reading images and recordings to score an endpoint.
• Safety. Coding adverse event descriptions to standard medical terms, and triaging case reports in pharmacovigilance, the ongoing monitoring of side effects after approval.
• Real-world evidence. Extracting diagnoses or outcomes from unstructured clinical notes to build a study cohort.
• Nonclinical work. Analyzing pathology slides or other laboratory images.
• Manufacturing. Monitoring processes, predicting quality attributes or flagging batches for review.

Many of these arrive through vendors. A contract research organization’s coding platform or a safety vendor’s case-intake tool may use AI that the sponsor never selected. Accountability does not move with the work. ICH E6(R3), the updated international good clinical practice guideline finalized in January 2025, expects sponsors to oversee their service providers and to validate computerized systems in proportion to risk. The inventory therefore has to reach into vendor contracts and vendor questionnaires.

For uses on the evidence side, the record should show the model version, the data it was tested on, the error rates found and the human review applied. A regulator whose staff now work with a generative tool, and have seen its errors firsthand, is well placed to ask precise questions about how a sponsor caught its own tool’s mistakes.

What a validation file contains

A credible file is short and specific. It states the intended use in one paragraph. It describes the reference dataset, how it was labeled and by whom, and why it represents the data the model will see. It records acceptance criteria that were set before testing, not after.

It then reports performance overall and in the subgroups that matter, such as sites, populations or product strengths. It names the human checks applied to outputs and how often reviewers overruled the model. It identifies the exact model version and the change-control process that triggers retesting.

That last item is harder for generative models than for traditional ones. Vendors update and retire model versions on their own schedules. A sponsor that cannot pin a version, or at least learn when it changes, cannot show that the model it tested is the model it used. Contracts for AI services used on the evidence side should address versioning and notice of change directly.

Sponsors that answer with a validation record, instead of a policy statement, will spend less review time on process questions. The cost is real: validation takes data science time that would otherwise go to new uses. That cost is lower before filing than during a review.

What a slow answer costs

Submission quality can sound like a back-office concern. It is not.

Time is money. Every round of questions during review adds days or weeks. For a product with meaningful sales potential, delays translate directly into lost revenue and pushed launch plans.

Credibility compounds. A sponsor whose submissions are consistent and whose answers are quick and well cited builds a reputation with review teams. One whose numbers do not line up invites closer scrutiny on everything else.

AI governance is becoming a regulatory topic. The more a company relies on AI to produce evidence, the more it must be able to show how that AI was tested. Boards that ask about AI strategy should also ask about AI validation.

Responsibility is spread across functions. The three parts call on different functions: regulatory operations for the dossier, the regulatory affairs lead for the evidence trail, and quality with data science for the AI inventory. Without a senior sponsor connecting them, each can assume another has it covered.

How the cost differs by company

The arithmetic of delay depends on the sponsor. For a large company, one product’s slipped launch is a portfolio problem: painful, but absorbed. For a single-asset biotech, the approval date often anchors its financing plan, its commercial hiring and its manufacturing commitments. A three-month extension can force a new funding round on worse terms.

That asymmetry argues for spending effort in proportion to exposure. A small sponsor with one filing gains more from a rigorous consistency pass than a large one gains on any single program. It also usually has fewer people to do it, which is why the work has to be written into vendor contracts early.

There is also a cost to overdoing it. Validation work, consistency tooling and index maintenance all take scarce regulatory and data science time. Teams that try to validate every AI use to the highest standard will stall their own adoption. The risk-based structure of FDA’s draft is useful here. It gives a defensible reason to spend heavily on a few high-influence uses and lightly on the rest.

Records a skeptical reader can check

What links the three disciplines is that each produces records a skeptical reader can check quickly. That was always good practice. From mid-2025 it was also the fastest route through a review with software in the loop.

In practice that means four things. A single dictionary governs every key definition, and automated checks run before every filing. A living index lets anyone find the source of any claim in a minute. An AI inventory classifies every use, with a named owner and a validation file for those that produce evidence.

A sequence for the next filing

For a team with a filing six to twelve months away, the work can be staged.

1. First month. Name an owner for each key definition. Collect the program’s endpoint, population and analysis set definitions into one controlled list, and note where current documents disagree.
2. Second and third months. Choose the facts the automated checks will compare, such as doses, sample sizes, adverse event counts and endpoint definitions. Agree who runs the checks, at which drafting milestones, and who signs off exceptions.
3. Fourth month. Build the claim-to-source index template and assign its maintenance to the regulatory lead for the whole review cycle, including amendments.
4. In parallel. Complete the AI inventory, classify each use, and open validation files for anything on the evidence side. Add versioning and change-notice terms to the relevant vendor contracts at renewal.
5. Before filing. Run a mock review. A colleague outside the program, using an internal search or summarization tool, tries to find inconsistencies the way a reviewer might, and the team fixes what is found.

The mock review is the cheapest test of the whole system, and it can be repeated before each major amendment during review. If an internal reader with a search tool finds a discrepancy in an afternoon, so can anyone else.

The familiar failures:

• Checking everything or nothing. Trying to make every sentence consistent and missing deadlines, or skipping checks entirely under time pressure. Focus on doses, sample sizes, adverse event counts and endpoints.
• Letting the index go stale. Building a claim-to-source index for filing, then not updating it as amendments arrive during review.
• Paraphrasing in responses. Restating what the submission says rather than quoting and citing it.
• Blaming the robot. Suggesting to a reviewer that a question came from AI.
• Classifying once. Labeling an AI tool “efficiency only” at adoption and never revisiting as its use expands.
• Policy instead of proof. Answering questions about AI validation with a policy document rather than actual test results.

Where a filing team should look first

• Do we have one agreed definition for each endpoint, population and analysis set in this program, and who owns it?
• Do we run automated checks comparing numbers across modules before we file, and which facts do they cover?
• If a reviewer questioned a figure tomorrow, could we point to its exact source the same day, and would our response quote the submission or paraphrase it?
• Where do we use AI in preparing submissions, has each use been classified as an efficiency aid or an evidence producer, and who in quality assurance signs off those classifications?
• For each evidence-producing use, could we show a reviewer who knows how these tools fail the model version, test data, error rates and human review?

With hindsight

The tool drew scrutiny within weeks. In July 2025, CNN reported that several current and former FDA staff said Elsa could invent studies or misrepresent research. Agency leaders acknowledged that it could hallucinate and said its use was optional, and the same reporting said staff found it useful for routine tasks such as meeting summaries.

That mix supports the defensive part of the argument above: sponsors needed records that make a misreading easy to correct. A tool that is useful and occasionally wrong is exactly the kind that rewards a well-organized submission. The reviewer gets the benefit of speed, and the sponsor protects itself against the occasional error by making the truth easy to find.

The agency did not slow down. On December 1, 2025, FDA said all of its employees now had access to agentic AI capabilities, meaning systems that can plan and carry out multi-step tasks. It listed premarket reviews, review validation, postmarket surveillance, inspections and compliance among the intended uses. Use remained voluntary.

The same announcement said that, according to FDA’s internal data, more than 70 percent of staff had used Elsa voluntarily. FDA again said the tools ran in a high-security GovCloud environment and did not train on data submitted by regulated industry. It also launched an internal challenge for staff to build agentic AI solutions.

That moved the reader from a fast search assistant toward something that could run a sequence of checks across a file. If anything, it strengthened the case for consistent dossiers. A multi-step tool asked to validate a review can compare far more pairs of numbers than a person would ever choose to.

The policy direction on sponsors’ own AI use also held. In January 2026, FDA and the European Medicines Agency published ten shared principles for good AI practice across the medicines lifecycle, built around risk, context of use and documentation. The European Medicines Agency is the European Union’s counterpart to FDA, so this was the two largest drug regulators speaking with one voice.

By this point in 2026, the expectation that sponsors can show how AI-generated evidence was validated is shared by those two regulators. For a company filing in both the United States and Europe, that is helpful: one well-built validation approach can serve both audiences, rather than two separate efforts. It also gives boards a stable reference point when they ask management how the company’s AI is governed.

The European side had a head start of its own. The European Medicines Agency had published a reflection paper on AI across the medicinal product lifecycle in 2024, setting out a similar risk-based approach. The joint principles drew the two strands together.

Some parts of the argument remain hard to test. Sponsors still cannot see how individual reviewers use these tools, and public reporting has said more about Elsa’s limits than about its effect on any particular review. We are not aware of a public case in which a review outcome was traced to an AI error. That absence is weak evidence either way, because review files are confidential.

What can be said is that the direction of travel was one-way. The regulator’s tools grew more capable over 2025, its staff adopted them widely, and its expectations of sponsors’ own AI became more explicit. Sponsors did not need to know how any single review used the tool. They only needed to assume a capable, fast and occasionally mistaken reader.

Companies that classified their AI uses early and kept their dossiers internally consistent are better placed for the questions those principles imply. The broader point is that regulators adopting AI and regulators scrutinizing AI turned out to be two sides of the same development. Organizations that can show their work, clearly and quickly, benefit from both. Those that cannot will see reviewers find inconsistencies in their filings sooner and in greater number.