Latest / DDReg Knowledge Capsule / AI in Regulatory Affairs: What It Can and Can’t Do in 2026

Transcript

AI in Regulatory Affairs: What It Can and Can’t Do in 2026 · DDReg Knowledge Capsule

• 0:00
  Speaker 1: Welcome to drag Knowledge capsule.
• 0:01
  Speaker 2: Thank you.
• 0:02
  Speaker 1: So, right now, there is a pharmaceutical company somewhere in
• 0:06
  Speaker 1: the world and they're basically trusting an algorithm to read
• 0:09
  Speaker 1: a ten thousand page clinical trial report.
• 0:12
  Speaker 2: Oh absolutely, it's happening right this second.
• 0:15
  Speaker 1: Right, And if that algorithm works exactly as intended, they
• 0:18
  Speaker 1: are going to save what like six months of agonizing manual.
• 0:21
  Speaker 2: Labor baisly six months.
• 0:23
  Speaker 1: But and this is the big but if it hallucinates
• 0:26
  Speaker 1: a single crucial data point, their drug gets rejected, the
• 0:30
  Speaker 1: launch is delayed, and it costs them millions of dollars.
• 0:33
  Speaker 2: Millions not to mention the delayed patient access, which is the.
• 0:36
  Speaker 1: Real cost exactly. So today we are finding out exactly
• 0:39
  Speaker 1: where that line is drawn. We're taking on this mission
• 0:41
  Speaker 1: to kind of separate the extreme hype from the actual
• 0:47
  Speaker 1: on the ground reality of using artificial intelligence and regulatory affairs,
• 0:51
  Speaker 1: specifically now in the year twenty twenty six.
• 0:53
  Speaker 2: Yeah, And the sources today they come directly from ddreg's
• 0:57
  Speaker 2: industry insights, which outline quite starkly what AI can and
• 1:01
  Speaker 2: absolutely cannot do right now.
• 1:02
  Speaker 1: Which we really need because the hype is just everywhere
• 1:05
  Speaker 1: it is.
• 1:05
  Speaker 2: But I think we've largely moved past that initial shock
• 1:08
  Speaker 2: and awe of generative models, right like, we are deep
• 1:12
  Speaker 2: into the implementation phase. Now. The real question for twenty
• 1:16
  Speaker 2: twenty six isn't whether AI is going to change the
• 1:19
  Speaker 2: regulatory landscape. I mean it already has.
• 1:21
  Speaker 1: Yeah, that ship has sailed, right.
• 1:23
  Speaker 2: So the critical analysis now is really about identifying the
• 1:26
  Speaker 2: exact fault lines, like where does AI create genuine efficiency
• 1:30
  Speaker 2: versus where does a blind reliance on it introduce existential
• 1:34
  Speaker 2: risks to a regulatory submission.
• 1:36
  Speaker 1: And I want to say right up front to you listening,
• 1:38
  Speaker 1: whether you are deep in the pharmaceutical industry trying to
• 1:41
  Speaker 1: get some complex biologic to market, or you're just I
• 1:45
  Speaker 1: don't know, a knowledge worker trying to figure out how
• 1:47
  Speaker 1: to integrate AI into your own daily workflow without completely
• 1:50
  Speaker 1: breaking things. This breakhand is going to give you a
• 1:52
  Speaker 1: highly practical reality check.
• 1:54
  Speaker 2: It really is applicable across the board.
• 1:55
  Speaker 1: Yeah, okay, let's unpack this because reading through these sources,
• 1:58
  Speaker 1: it feels like AI in twenty two twenty six is
• 2:00
  Speaker 1: less like this omniscient robot overlord we were all promised,
• 2:04
  Speaker 1: and more like a well, like a brilliant, incredibly fast
• 2:07
  Speaker 1: but just highly overconfident intern.
• 2:10
  Speaker 2: That is. I love that that intern analogy maps perfectly
• 2:14
  Speaker 2: onto the underlying architecture of large language.
• 2:16
  Speaker 1: Models, because if you don't check their work, you're gonna
• 2:20
  Speaker 1: end up in.
• 2:20
  Speaker 2: A lot of trouble exactly. I mean, if you ask
• 2:22
  Speaker 2: an overconfident human intern to synthesize one hundred pages of
• 2:26
  Speaker 2: really dense reading by tomorrow morning, they will do it.
• 2:29
  Speaker 1: They'll stay up all night, drink a bunch of coffee
• 2:32
  Speaker 1: and get it done right.
• 2:33
  Speaker 2: They will find the patterns and summarize the text. But
• 2:36
  Speaker 2: if you ask that same intern to, say, negotiate a
• 2:39
  Speaker 2: complex legal pathway with a federal health agency using undocumented
• 2:45
  Speaker 2: historical precedence, oh boy, yeah, they are going to confidently
• 2:49
  Speaker 2: steer your entire operation right off a cliff.
• 2:51
  Speaker 1: Because they lack the contextual judgment to know what they
• 2:54
  Speaker 1: don't know. They're just guessing based on the vibes of.
• 2:57
  Speaker 2: The room, excisely.
• 2:58
  Speaker 1: So let's look at the sweet spot first. What are
• 3:00
  Speaker 1: the specific tasks that we should actually be handing over
• 3:03
  Speaker 1: to this intern right now, Because according to the DDRAG insights,
• 3:07
  Speaker 1: the reality of what actually works is incredibly specific.
• 3:12
  Speaker 2: It is it hinges entirely on what large language models
• 3:15
  Speaker 2: and natural language processing actually do well at a purely
• 3:19
  Speaker 2: mechanical level, which is pattern recognition and semantic synthesis within
• 3:24
  Speaker 2: very narrow, well defined parameters.
• 3:26
  Speaker 1: Okay, so give me an example in the regulatory space.
• 3:28
  Speaker 2: Sure, let's look at literature review and evidence synthesis. To
• 3:32
  Speaker 2: prepare things like PBRs, those periodic benefit risk evaluation reports,
• 3:38
  Speaker 2: or even clinical overviews, regulatory teams are forced to scan massive,
• 3:43
  Speaker 2: almost incomprehensible volumes of published literature. We are talking mountains
• 3:48
  Speaker 2: of medical journals, adverse event reports, trial data.
• 3:52
  Speaker 1: But wait, how is the AI actually reading that much
• 3:55
  Speaker 1: without losing the plot? Because you know, anyone who's used
• 3:58
  Speaker 1: an LM knows they have context windows. They tend to
• 4:01
  Speaker 1: just forget what was at the beginning of a document
• 4:03
  Speaker 1: by the time they reached the end.
• 4:04
  Speaker 2: Yeah, the goldfish memory problem. But this is where the
• 4:06
  Speaker 2: enterprise deployment of AI differs from like consumer chatbots, these
• 4:10
  Speaker 2: teams are using vector databases and retrieval augmented generation.
• 4:14
  Speaker 1: Okay, that sounds intense, What does that actually mean for
• 4:16
  Speaker 1: the user?
• 4:17
  Speaker 2: Basically, instead of just dumping ten thousand pages into a
• 4:20
  Speaker 2: prompt and hoping for the best the system mathematically indexes
• 4:24
  Speaker 2: the entire ocean of text first.
• 4:27
  Speaker 1: Ah, so when a.
• 4:28
  Speaker 2: Regulatory scientist needs to extract, say, all mentions of a
• 4:32
  Speaker 2: specific cardiac anomaly across five years of literature, the AI
• 4:36
  Speaker 2: isn't just doing a simple keyword search, right.
• 4:38
  Speaker 1: It's not just hitting ctrlff exactly.
• 4:41
  Speaker 2: It understands the semantic relationship between terms, and it retrieves
• 4:45
  Speaker 2: the relevant paragraphs instantly. It synthesizes that specific data significantly
• 4:50
  Speaker 2: faster than a human team ever could.
• 4:52
  Speaker 1: Okay, so that handles the massive data synthesis part. Now,
• 4:56
  Speaker 1: the sources also heavily emphasize regulatory intelligence, monitor, drink and pushback. Here,
• 5:02
  Speaker 1: tracking FDA versus EMAY guidelines isn't exactly a new problem, right,
• 5:05
  Speaker 1: Like regulatory professionals do that every Tuesday.
• 5:08
  Speaker 2: They do, but it is grueling.
• 5:10
  Speaker 1: So the real trick seems to be how an AI
• 5:12
  Speaker 1: model handles the semantic conflicts between those agencies, especially when
• 5:16
  Speaker 1: they use the exact same terminology to mean slightly different things.
• 5:19
  Speaker 2: Yes, what's fascinating here is that that exact friction point
• 5:24
  Speaker 2: is why human teams struggle to track ten or twenty
• 5:26
  Speaker 2: international markets simultaneously.
• 5:28
  Speaker 1: Because it just gets jumbled in your head exactly.
• 5:31
  Speaker 2: A human gets fatigued trying to remember that, you know,
• 5:34
  Speaker 2: the FDA's definition of a specific reporting threshold differs subtly
• 5:39
  Speaker 2: from the PMDA in Japan.
• 5:40
  Speaker 1: Right.
• 5:41
  Speaker 2: But AI powered surveillance tools excel here because they handle
• 5:44
  Speaker 2: discrete compartmentalization perfectly. They can monitor agency updates, new guidance publications,
• 5:50
  Speaker 2: and legislative changes simultaneously across dozens of markets without crowd
• 5:55
  Speaker 2: contaminating the definitions in their memory.
• 5:57
  Speaker 1: And they don't need to sleep.
• 5:58
  Speaker 2: They don't sleep no, And they provide this wide net
• 6:01
  Speaker 2: monitoring system that prevents missed deadlines.
• 6:03
  Speaker 1: That's huge. Now, another area the source highlights is gap
• 6:06
  Speaker 1: analysis against regulatory templates and pharmaceutical submissions. Are I mean
• 6:11
  Speaker 1: they are incredibly.
• 6:12
  Speaker 2: Structured, very rigid.
• 6:13
  Speaker 1: Yes, you have ICHCTD modules and common technical document along
• 6:17
  Speaker 1: with these strict FDA guidance checklists and EMA assessment structures.
• 6:22
  Speaker 2: And this is just a mechanical dream for AI. Some
• 6:25
  Speaker 2: of the most effective tools today are trained specifically to
• 6:28
  Speaker 2: compare a draft document against these standard ICH structures, just
• 6:32
  Speaker 2: matching them up right. Because the reference framework is so rigid,
• 6:36
  Speaker 2: the AI can instantly flag missing sections inconsistent terminology or
• 6:40
  Speaker 2: formatting deviations. It is simply comparing data set A against
• 6:45
  Speaker 2: template B and highlighting the anomalies.
• 6:48
  Speaker 1: Okay. The insights also mentioned translation and cross referencing, like
• 6:52
  Speaker 1: accelerating the production of document cross references and hyperlinking these
• 6:56
  Speaker 1: massive CTD modules so the reviewers can actually, you know.
• 7:00
  Speaker 2: Navigate them, which is a huge pain point.
• 7:02
  Speaker 1: I bet, But I have to be honest here. Looking
• 7:05
  Speaker 1: at this list triage, formatting, tracking, hyperlinking, it sounds like
• 7:09
  Speaker 1: AI is basically just an advanced search engine and a
• 7:12
  Speaker 1: hyperactive spell checker.
• 7:13
  Speaker 2: That's one way to look at it.
• 7:14
  Speaker 1: Sure, I mean, is this really moving the needle for
• 7:16
  Speaker 1: these massive pharmaceutical companies. It doesn't sound like the grand
• 7:19
  Speaker 1: technological revolution we were promised.
• 7:21
  Speaker 2: Well, while these tasks might sound purely administrative to an outsider,
• 7:25
  Speaker 2: in the context of global regulatory compliance, they are the
• 7:28
  Speaker 2: foundational bedrock. Okay, how so we aren't just talking about
• 7:32
  Speaker 2: saving an assistant a few hours of typing. We are
• 7:34
  Speaker 2: dealing with an environment where the sheer volume of data
• 7:37
  Speaker 2: across dozens of international jurisdictions makes human error practically statistically.
• 7:43
  Speaker 1: Inevitable because there's just too much of it.
• 7:45
  Speaker 2: Exactly and missing a single check boss on an FDA submission,
• 7:50
  Speaker 2: or failing to hyperlink a critical toxicity report in a
• 7:54
  Speaker 2: CTD module. That doesn't just result in a polite email
• 7:57
  Speaker 2: asking for a revision.
• 7:58
  Speaker 1: No, it results in a refusal.
• 7:59
  Speaker 2: To file exactly, a refusal defile, which is a multi
• 8:03
  Speaker 2: month delay in getting a life saving drug to market.
• 8:06
  Speaker 2: The financial cost is measured in the millions, and the
• 8:08
  Speaker 2: human cost is measured in delayed patient access. So AI
• 8:12
  Speaker 2: fundamentally reduces the risk of global regulatory surprises. It provides
• 8:17
  Speaker 2: a level of comprehensive, tireless vigilance that a human team
• 8:21
  Speaker 2: simply cannot sustain without burning out. The value isn't always
• 8:25
  Speaker 2: in doing a complex task. Often the highest value is
• 8:28
  Speaker 2: executing a high volume, tedious task with mechanical.
• 8:31
  Speaker 1: Flawlessness that reframes it perfectly. The value is in the
• 8:35
  Speaker 1: flawless execution of volume. Yes, but then produce us right
• 8:38
  Speaker 1: to the edge of the cliff. I think, because if
• 8:41
  Speaker 1: AI is so incredibly good at synthesizing data and organizing
• 8:45
  Speaker 1: these massive regulatory documents, why our experienced regulatory professionals so
• 8:50
  Speaker 1: deeply skeptical about letting it write the actual submissions?
• 8:54
  Speaker 2: Ah, that is the million dollar question.
• 8:56
  Speaker 1: Like, we trust it to organize five thousand pages, but
• 8:58
  Speaker 1: we won't let it write the three pagetrategic summary. Why
• 9:01
  Speaker 1: the massive.
• 9:01
  Speaker 2: Disconnect Because organizing data and exercising contextual judgment rely on
• 9:06
  Speaker 2: two entirely different computational realities. And this is exactly where
• 9:10
  Speaker 2: we enter the danger zone.
• 9:11
  Speaker 1: You walk me through it.
• 9:12
  Speaker 2: When you move beyond well defined templates and into tasks
• 9:16
  Speaker 2: that require interpretation, negotiation, and strategy, AI consistently falls apart.
• 9:22
  Speaker 2: The DDRIG insights are very explicit about what is essentially
• 9:26
  Speaker 2: off limits.
• 9:26
  Speaker 1: For AI right now and what's at the top of that.
• 9:28
  Speaker 2: List, starting with regulatory strategy, formulation and determining novel product pathways.
• 9:34
  Speaker 1: Let's break down why it fails there. I mean, when
• 9:36
  Speaker 1: you are bringing a completely new innovative drug to market,
• 9:42
  Speaker 1: maybe a new gene therapy or something, there isn't a
• 9:44
  Speaker 1: historical checklist to follow.
• 9:46
  Speaker 2: Now you're charting new territory.
• 9:48
  Speaker 1: You have to coordinate with multiple agencies, interpret unwritten precedents,
• 9:52
  Speaker 1: and essentially negotiate the pathway with the health authorities right exactly.
• 9:55
  Speaker 2: And the mechanism behind an AI's failure here is rooted
• 9:59
  Speaker 2: in what large language models actually are at their core,
• 10:02
  Speaker 2: they are probabilistic next word predictors.
• 10:04
  Speaker 1: They're just guessing the next word right.
• 10:07
  Speaker 2: They look at vast amounts of training data and guess
• 10:09
  Speaker 2: the most mathematically likely sequence of words. They are not
• 10:13
  Speaker 2: strategic negotiators. They cannot read the room. They don't understand
• 10:16
  Speaker 2: the shifting political dynamics within a regulatory agency, which is huge, huge,
• 10:22
  Speaker 2: And they cannot anticipate how an agency might interpret a novel,
• 10:26
  Speaker 2: unprecedented piece of clinical data.
• 10:29
  Speaker 1: It's the same reason it fails at risk benefit assessment narratives,
• 10:32
  Speaker 1: which the source also flags like you are weighing the
• 10:36
  Speaker 1: potential side effects of a drug against its therapeutic benefits,
• 10:39
  Speaker 1: often dealing with conflicting or incomplete clinical.
• 10:42
  Speaker 2: Data, very messy data usually Yeah.
• 10:45
  Speaker 1: And an LLM doesn't understand human biology. It doesn't know
• 10:49
  Speaker 1: what a patient actually is. It just mapps the statistical
• 10:52
  Speaker 1: proximities of words in a document, and that.
• 10:54
  Speaker 2: Maps directly onto another major failure point highlighted in the sources,
• 10:58
  Speaker 2: which is MEDDRC. Yeah. Med DEAA is the highly specific
• 11:03
  Speaker 2: medical terminology used to classify adverse event reports globally. If
• 11:08
  Speaker 2: a patient reports a side effect in a clinical trial.
• 11:10
  Speaker 2: It must be coded correctly using this taxonomy, and the
• 11:13
  Speaker 2: source notes that using AI for complex MEDDA coding carries
• 11:16
  Speaker 2: a moderate to high risk.
• 11:17
  Speaker 1: I can definitely see why if an LM is just
• 11:20
  Speaker 1: predicting words based on training data, well, myocardial infarction and
• 11:23
  Speaker 1: severe heartburn might share adjacent textual space in some broad
• 11:26
  Speaker 1: Internet data set, But clinically confusing a heart attack with
• 11:29
  Speaker 1: heartburn is catastrophic.
• 11:31
  Speaker 2: Beyond catastrophic, and the stakes go beyond just technical inaccuracy.
• 11:35
  Speaker 2: The legal liability remains strictly with the MH, the marketing
• 11:39
  Speaker 2: authorization holder, so.
• 11:40
  Speaker 1: The pharmaceutical companies on.
• 11:41
  Speaker 2: The hook entirely liable. An algorithm cannot go to court.
• 11:45
  Speaker 2: If the AI miscodes a serious adverse event signal and
• 11:49
  Speaker 2: the company submits that to the health authorities, the company
• 11:52
  Speaker 2: is on the hook for masking a safety signal. The
• 11:55
  Speaker 2: authorities do not care if a machine made the error, which.
• 11:59
  Speaker 1: Leads us directly into what the source describes as this
• 12:02
  Speaker 1: fascinating time cost paradox. We implement AI because we think, oh,
• 12:06
  Speaker 1: it's going to save us all this time and money.
• 12:08
  Speaker 2: That's the sales pitch anyway, right.
• 12:10
  Speaker 1: But the insights reveal that if you don't validate these
• 12:12
  Speaker 1: tools properly, they actually create significantly more work than they save.
• 12:17
  Speaker 2: It is a profoundly unintuitive outcome, and it is catching
• 12:20
  Speaker 2: a lot of companies off guard in twenty twenty six.
• 12:22
  Speaker 1: How does it happen?
• 12:23
  Speaker 2: Well, a regulatory team will implement an AI tool without
• 12:27
  Speaker 2: a clear oversight framework, expecting it to do the heavy
• 12:30
  Speaker 2: lifting of, say, drafting a complex clinical overview. The AI
• 12:34
  Speaker 2: generates a fifty page document in seconds. That it's great, right, Oh,
• 12:37
  Speaker 2: it looks perfect. The grammar is flawless, the formatting perfectly
• 12:42
  Speaker 2: matches the ICCH guidelines. The tone sounds highly authoritative, But
• 12:46
  Speaker 2: beneath that polished surface, it is hiding subtle factual errors,
• 12:51
  Speaker 2: or citing outdated guidelines from twenty twenty two, or drawing
• 12:55
  Speaker 2: logically inconsistent conclusions about the drug safety profile.
• 12:58
  Speaker 1: Okay, here is where using AI for these high level
• 13:02
  Speaker 1: judgments is literally like using a bulldozer to weed a garden. Yes,
• 13:06
  Speaker 1: I mean yes, it's incredibly powerful, it moves incredibly fast,
• 13:09
  Speaker 1: and the job will technically get doneically, But the bulldozer
• 13:12
  Speaker 1: doesn't actually know what a weed is, and it doesn't
• 13:13
  Speaker 1: know what a flower is it just knows how to
• 13:15
  Speaker 1: push dirt. It's going to completely destroy the flowers in
• 13:18
  Speaker 1: the process. Plausible, but factually wrong is the absolute most
• 13:22
  Speaker 1: dangerous kind of wrong in regulatory affairs.
• 13:24
  Speaker 2: That is spot on. The insidious nature of a hallucination
• 13:28
  Speaker 2: is its confidence. If that interurn we talked about hands
• 13:32
  Speaker 2: you a document riddled with spelling errors and weird formatting,
• 13:36
  Speaker 2: your guard immediately goes up, right.
• 13:37
  Speaker 1: Oh, yeah, you're double checking everything they wrote.
• 13:39
  Speaker 2: You are going to check every single fact. But if
• 13:42
  Speaker 2: they hand you a beautifully bound, perfectly formatted report that
• 13:45
  Speaker 2: reads like it was written by a senior scientist, you
• 13:48
  Speaker 2: might just skim it and sign off.
• 13:50
  Speaker 1: And suddenly the subject matter experts are spending forty hours
• 13:54
  Speaker 1: meticulously reviewing every single line of an AI generated document
• 13:58
  Speaker 1: just defind the hallucine citations, which.
• 14:01
  Speaker 2: Takes way more time than just writing the document from
• 14:03
  Speaker 2: scratch exactly.
• 14:05
  Speaker 1: Plus the company has to develop new standard operating procedures
• 14:08
  Speaker 1: just to manage the AI. They have to create validation
• 14:11
  Speaker 1: documentation for GXP relevant processes. All of that is unbudgeted time.
• 14:17
  Speaker 2: And regulatory agencies are acutely aware of this dynamic. The
• 14:20
  Speaker 2: FDA's AML Action Plan now explicitly requires that any AI
• 14:25
  Speaker 2: assisted tools used in the preparation of regulatory submissions must
• 14:29
  Speaker 2: be fully disclosed and validated appropriately.
• 14:31
  Speaker 1: No more flying under the radar, exactly.
• 14:34
  Speaker 2: You can no longer use an algorithm in the shadows.
• 14:37
  Speaker 2: If you haven't formalized your AI governance framework, or if
• 14:40
  Speaker 2: you can't prove how your model arrived at a specific
• 14:43
  Speaker 2: data synthesis, you are going to face intense halting scrutiny
• 14:47
  Speaker 2: from the FDA.
• 14:48
  Speaker 1: They want to see the blueprints for the bulldozer, and
• 14:50
  Speaker 1: they want to know exactly who is holding the keys.
• 14:53
  Speaker 2: That's a great way to put it, which brings us to.
• 14:54
  Speaker 1: The practical solution. Recognizing these severe time, cost paradoxes and
• 15:00
  Speaker 1: compliance risks, how are organizations like DDRAG actually containing this technology?
• 15:06
  Speaker 1: How do you put guardrails around the bulldozer so it
• 15:08
  Speaker 1: only pushes the dirt you actually wanted to push.
• 15:10
  Speaker 2: Well, the industry has gravitated toward a very strict, practical,
• 15:14
  Speaker 2: three tier task classification framework. This is how the most
• 15:18
  Speaker 2: effective and compliant teams in twenty twenty six are structuring
• 15:22
  Speaker 2: their entire operational workflow. It is essentially a triage system
• 15:26
  Speaker 2: for cognitive labor.
• 15:27
  Speaker 1: Okay, So if the FDA is going to penalize us
• 15:29
  Speaker 1: for operating blindly, let's break down these tiers. What does
• 15:31
  Speaker 1: Tier one look like? In practice?
• 15:32
  Speaker 2: Tier one is classified as AI led human.
• 15:35
  Speaker 1: Reviewed, AI led human reviewed.
• 15:38
  Speaker 2: Right, this is the realm of literature triage scanning for
• 15:41
  Speaker 2: adverse events, regulatory tracking across multiple international markets, and heavy
• 15:46
  Speaker 2: document cross referencing. The AI is firmly in the driver's
• 15:50
  Speaker 2: seat here doing the high volume heavy lifting.
• 15:53
  Speaker 1: And what's the human doing?
• 15:55
  Speaker 2: The human is in the passenger seat, monitoring the dashboard,
• 15:58
  Speaker 2: checking the mirrors, and ensuring the vehicle stays on the road.
• 16:01
  Speaker 1: But I have to challenge that a bit. If we
• 16:03
  Speaker 1: are asking a human to review an AI's summary of
• 16:07
  Speaker 1: ten thousand pages of clinical literature, aren't they basically having
• 16:11
  Speaker 1: to read all ten thousand pages anyway, just to verify
• 16:15
  Speaker 1: if the AIS summary is accurate. It seems that way, right, Yeah, Like,
• 16:18
  Speaker 1: how is that actually saving time?
• 16:19
  Speaker 2: Ah? Because a robust Tier one system doesn't just hand
• 16:22
  Speaker 2: a human a summary and say trust me. The software
• 16:25
  Speaker 2: provides source linking and confidence scoring.
• 16:27
  Speaker 1: Oh.
• 16:27
  Speaker 2: Interesting, Yeah, When the AI extracts a specific adverse event metric,
• 16:32
  Speaker 2: it directly hyperlinks to the exact source paragraph in the
• 16:35
  Speaker 2: original document.
• 16:36
  Speaker 1: Okay, that changes things exactly.
• 16:38
  Speaker 2: The human reviewer is executing targeted spot checking and verifying
• 16:42
  Speaker 2: anomalies rather than reading sequentially from page one to page
• 16:45
  Speaker 2: ten thousand. It is the difference between searching for a
• 16:49
  Speaker 2: needle in a haystack by hand versus using a metal
• 16:52
  Speaker 2: detector to find the exact quadrant and then using your hands.
• 16:55
  Speaker 1: That makes perfect sense. The human is verifying the specific extraction,
• 16:59
  Speaker 1: not read doing the search. So that's tier one. What
• 17:02
  Speaker 1: about Tier two? There has to be a middle ground
• 17:05
  Speaker 1: where humans in AI are actually collaborating on the drafting process.
• 17:09
  Speaker 2: Tier two is exactly that mental ground. It is classified
• 17:12
  Speaker 2: as AI assisted human authored Okay. This involves generating the
• 17:15
  Speaker 2: initial drafts of highly standard sections like quality modules or
• 17:19
  Speaker 2: formatting a summary of product characteristics the SMPC. It also
• 17:23
  Speaker 2: covers translation.
• 17:24
  Speaker 1: Review, so the dynamic shifts here.
• 17:26
  Speaker 2: It does. In this dynamic, the human regains control of
• 17:29
  Speaker 2: the steering wheel. The AI acts as an advanced copilot
• 17:33
  Speaker 2: or navigation system, suggesting standard phrasing, ensuring terminology aligns with
• 17:37
  Speaker 2: the latest EMAI guidelines, and helping construct the standardized elements.
• 17:41
  Speaker 2: But the human is doing the actual authoring. The AI
• 17:44
  Speaker 2: is just providing real time structural support.
• 17:47
  Speaker 1: Got it, Humans drive AI navigates. And then we reach
• 17:50
  Speaker 1: Tier three, which seems to be the most critical boundary
• 17:52
  Speaker 1: line in this entire framework.
• 17:54
  Speaker 2: Absolutely, Tier three is human led AI unsupported.
• 17:57
  Speaker 1: AI unsupported meaning the algorithm is completely locked out of.
• 18:00
  Speaker 2: The room, entirely locked out. This tier is strictly reserved
• 18:03
  Speaker 2: for regulatory strategy, direct agency interactions, risk benefit determinations, and
• 18:09
  Speaker 2: novel pathway formulation.
• 18:11
  Speaker 1: No robots allowed none.
• 18:12
  Speaker 2: AI has no place in these tasks because they require
• 18:15
  Speaker 2: the exact cognitive abilities that large language models intrinsically lack.
• 18:19
  Speaker 2: Contextual judgment, human empathy, nuanced and negotiation skills, and most importantly,
• 18:24
  Speaker 2: legal and ethical accountability.
• 18:26
  Speaker 1: You know, thinking about this, shouldn't you and I and
• 18:29
  Speaker 1: really everyone listening be applying this exact three tier framework
• 18:33
  Speaker 1: to all of our knowledge work.
• 18:34
  Speaker 2: I think so.
• 18:34
  Speaker 1: Yeah, Like, beyond just pharmaceutical regulations, taking a hard look
• 18:38
  Speaker 1: at your own job and defining exactly what tasks are
• 18:42
  Speaker 1: Tier three AI unsupported seems like an absolute superpower in
• 18:46
  Speaker 1: twenty twenty six. If you know what your tier three is,
• 18:49
  Speaker 1: you know exactly what makes you irreplaceable as a human
• 18:52
  Speaker 1: being in your organization.
• 18:53
  Speaker 2: That is profound, and it's absolutely true. The professionals who
• 18:56
  Speaker 2: are thriving in this current landscape are the ones who
• 19:00
  Speaker 2: draw that Tier three boundary line in thick permanent marker.
• 19:03
  Speaker 2: And this tiered approach is precisely how DDREG structures its
• 19:07
  Speaker 2: own software ecosystems. Tools like Vitalic LC for life cycle
• 19:11
  Speaker 2: management and Vitallic Regenttele for regulatory intelligence are built entirely
• 19:15
  Speaker 2: on this philosophy.
• 19:16
  Speaker 1: So they bake the framework right into the software exactly.
• 19:19
  Speaker 2: They treat AI purely as an efficiency mechanism embedded within
• 19:22
  Speaker 2: an expert led operation. The technology is leveraged as a
• 19:26
  Speaker 2: highly capable tool, never as a substitute for human expertise,
• 19:30
  Speaker 2: which makes.
• 19:31
  Speaker 1: Sense when you look at Vitallic REGENTTEL. It isn't deciding
• 19:34
  Speaker 1: if a pharmaceutical product is safe for human consumption. It
• 19:38
  Speaker 1: is simply ensuring that the human expert who is deciding
• 19:41
  Speaker 1: if the drug is safe has every single updated regulation
• 19:44
  Speaker 1: from one hundred and twenty different global health agencies organized
• 19:47
  Speaker 1: perfectly in front of them before they make the call.
• 19:49
  Speaker 2: Precisely, the human expert retains total uncompromised ownership of the
• 19:54
  Speaker 2: judgment dependent decisions. When a regulatory team applies this three
• 19:58
  Speaker 2: tier classification rigorous, the results are highly predictable. You achieve
• 20:03
  Speaker 2: massive scale and efficiency in tier one, measured and careful
• 20:06
  Speaker 2: productivity gains in tier two, and absolute compliance and security
• 20:10
  Speaker 2: in tier three.
• 20:11
  Speaker 1: So, bringing all of this together, the reality of AI
• 20:14
  Speaker 1: in twenty twenty six is that it is neither a
• 20:17
  Speaker 1: magic bullet that is going to seamlessly do all our
• 20:19
  Speaker 1: work while we sleep on a beach, nor is it
• 20:21
  Speaker 1: a useless gimmick. Its true value is strictly, rigorously bounded
• 20:25
  Speaker 1: by its technical architecture. Well said, and the winners right now,
• 20:29
  Speaker 1: whether you are running a global regulatory affairs department for
• 20:32
  Speaker 1: a massive pharma company or you're just a solo knowledge
• 20:35
  Speaker 1: worker managing your own projects, are the ones who know
• 20:37
  Speaker 1: exactly where that boundary line is drawn. You have to
• 20:40
  Speaker 1: know when to let the intern organize the filing cabinets,
• 20:43
  Speaker 1: and you have to know exactly when to kick them
• 20:45
  Speaker 1: out of the boardroom so you can make the hard
• 20:47
  Speaker 1: strategic decisions exactly.
• 20:50
  Speaker 2: Though, this landscape does raise one deeply provocative question, and
• 20:53
  Speaker 2: it is something really worth considering. When we look at
• 20:56
  Speaker 2: the FDA's current disclosure requirements. Well, if regulatory agencies are
• 21:02
  Speaker 2: increasingly demanding full disclosure and stringent validation of AI tools
• 21:06
  Speaker 2: used by pharmaceutical companies today, what happens a few years
• 21:10
  Speaker 2: from now when those regulatory agencies themselves are heavily deploying
• 21:14
  Speaker 2: AI to review the AI generated submissions they receive Oh wow?
• 21:18
  Speaker 2: Right when you reach a point where two distinct algorithms
• 21:21
  Speaker 2: are essentially negotiating a complex drug approval across a digital
• 21:25
  Speaker 2: table from one another, where exactly does the human liability sit.
• 21:29
  Speaker 1: Algorithm to algorithm negotiation? That is? I mean, just picture
• 21:33
  Speaker 1: the over confident intern from the pharma company debating the
• 21:37
  Speaker 1: overconfident intern from the FDA while all the human executives
• 21:40
  Speaker 1: just cross their fingers and hope nobody hallucinates a critical
• 21:43
  Speaker 1: safety profile.
• 21:44
  Speaker 2: It's a brave new world.
• 21:45
  Speaker 1: That is a wild, slightly terrifying thought to leave on.
• 21:48
  Speaker 1: Thank you for joining us on this deep dive into
• 21:50
  Speaker 1: the realities of twenty twenty six. As you go back
• 21:53
  Speaker 1: to your day, I highly encourage you to take five
• 21:56
  Speaker 1: minutes look at your own to do list and figure
• 21:58
  Speaker 1: out exactly what belongs in your Tier three guard your
• 22:01
  Speaker 1: human lad AI unsupported tasks closely because that is where
• 22:05
  Speaker 1: your real value is until next time.