Listen to Audio
Listen Anywhere
June 2026
In this episode, listen to highlights from our most recent Artificial Intelligence-focused episodes, featuring discussions with speakers from the FDA, Sanofi, and the Open BioPharma Research and Training Institute on the potential for AI to reshape the drug lifecycle, how companies are leveraging artificial intelligence to drive quality transformation and accelerate workforce readiness, and more.
1
00:00:00,240 --> 00:00:10,080
Welcome to the ISPE podcast, shaping the future of pharma, where ISPE supports you on your journey, fueling innovation, sharing insights, thought
2
00:00:10,080 --> 00:00:14,240
leadership, and empowering a global community to reimagine what's possible.
3
00:00:14,240 --> 00:00:15,205
You
4
00:00:15,205 --> 00:00:22,565
shared a a use case called Plai, p l a I, at the ISPE conference.
5
00:00:22,565 --> 00:00:27,765
And I understand this tool is used at scale, at Sanofi.
6
00:00:27,765 --> 00:00:33,750
Can you share how the tool was taken from pilot to at scale?
7
00:00:33,750 --> 00:00:34,230
The
8
00:00:34,230 --> 00:00:43,545
initial pilot focused on one specific use case, where we were, trying to automate deviations and complaint trending while providing root cause
9
00:00:43,545 --> 00:00:44,585
suggestions.
10
00:00:44,585 --> 00:00:54,585
Traditionally, it's a very manual, process where the quality team spend hours looking for patterns across different types of events, and they spend time brainstorming what the potential
11
00:00:54,585 --> 00:00:55,625
root causes could be.
12
00:00:55,625 --> 00:01:05,460
But within a few months, we could see concrete results with play with faster identification of trending issues, and AI could generate
13
00:01:05,460 --> 00:01:10,580
potential root cause hypotheses that allowed the investigators to to explore further.
14
00:01:10,580 --> 00:01:20,075
And what it also enabled us to, bring this tool to scale was our, governance frame framework, which we call
15
00:01:20,075 --> 00:01:20,715
RAISE.
16
00:01:20,715 --> 00:01:30,270
So RAISE stands for responsible AI of at Sanofi Enterprise, and it really provides, the guardrails for how we can, develop, validate,
17
00:01:30,270 --> 00:01:33,790
deploy AI tools across our organization.
18
00:01:33,790 --> 00:01:34,590
Please
19
00:01:34,590 --> 00:01:38,590
tell me about Open Biopharma Research and Training Institute.
20
00:01:38,590 --> 00:01:43,005
How did it get started, and what is its scope of services today?
21
00:01:43,005 --> 00:01:43,485
We
22
00:01:43,485 --> 00:01:47,485
like to do that bridge between, you academia and industries since there's often a gap between them.
23
00:01:47,485 --> 00:01:56,580
We wanna make sure that, you know, there are folks who are coming out of these academic institutions who are right, ready to work, ready to and and not just ready to work, like, ready to work at a at
24
00:01:56,580 --> 00:01:57,380
a certain level.
25
00:01:57,380 --> 00:02:03,860
One of the goals of our of our nonprofit, you know, because Open Biopharma Research and Training Institute is a nonprofit training institute.
26
00:02:03,860 --> 00:02:07,695
And we also have, you know, we have a we have a core lab services.
27
00:02:07,695 --> 00:02:14,975
We have our, you know, our polypractical production space, and we have our our training services that we we offer, and we also host as a venue as well.
28
00:02:14,975 --> 00:02:23,280
So we have often a host leading training services and training come training from different organizations who are have industry ready training.
29
00:02:23,280 --> 00:02:31,360
And though we often bring do we often work deals with them to have a a channel for all of our, apprentices and interns to be able to access that training.
30
00:02:31,360 --> 00:02:34,835
So this way, then when they come out, third, that's how we say when they're industry ready.
31
00:02:34,835 --> 00:02:42,355
They're they're able to do all of these other things and actually try out these different elements, work on the core lab services, and really get a real well rounded level of experience to understand
32
00:02:42,355 --> 00:02:44,035
where they best fit.
33
00:02:44,035 --> 00:02:48,035
Because oftentimes, it can be difficult when you're in college trying to figure out exactly what you wanna do.
34
00:02:48,035 --> 00:02:51,930
And then that never really matches with whatever industry is about to do because it always changes.
35
00:02:51,930 --> 00:02:55,370
You know, two years, five years, that's difficult for these things to work.
36
00:02:55,370 --> 00:02:57,530
So that's Open BioPharma in a nutshell.
37
00:02:57,530 --> 00:03:02,755
And, yeah, we're we're happy to be here, and it's a great it's a great organization.
38
00:03:02,755 --> 00:03:03,075
How
39
00:03:03,075 --> 00:03:12,035
do you see the integration of advanced AI techniques transforming drug development, regulatory submissions, and also manufacturing oversight?
40
00:03:12,035 --> 00:03:12,355
I
41
00:03:12,355 --> 00:03:22,060
think there are many different ways that AI could be used, You know, some of which would be actively regulated by FDA, some of it which, you know, is within your own, you know, industry's own
42
00:03:22,060 --> 00:03:24,940
control systems to be able to moderate and regulate.
43
00:03:24,940 --> 00:03:32,855
Given the right boundary conditions and ask the right questions, the AI can analyze the data far quicker than a human being can.
44
00:03:32,855 --> 00:03:33,815
You know?
45
00:03:33,815 --> 00:03:36,615
But ultimately, you know, a human being has to check it.
46
00:03:36,615 --> 00:03:45,940
A human being has to make sure the boundary conditions are set correctly, that the parameters are set correctly, and that, you know, whatever comes out makes sense within the context of what you're
47
00:03:45,940 --> 00:03:46,500
using.
48
00:03:46,500 --> 00:03:49,380
So I think there there's a there are great opportunities.
49
00:03:49,380 --> 00:03:53,300
With opportunities, there are risks, and you have to control for those risks.
50
00:03:53,300 --> 00:04:03,175
But I don't think that we should shy away from those opportunities in the same way that we haven't shied away from adding or using newer technologies
51
00:04:03,175 --> 00:04:07,415
and, for example, manufacturing or molecule selection that we have in the past.
52
00:04:07,415 --> 00:04:09,655
We're to use process models.
53
00:04:09,655 --> 00:04:12,695
We're used computational modeling in order to do molecule selection.
54
00:04:12,695 --> 00:04:16,400
These I think of this as, like, the next logical step.
55
00:04:16,400 --> 00:04:16,640
I've
56
00:04:16,640 --> 00:04:26,560
been with, ISPE for maybe thirty years now, or even more, and it's amazing how initiatives and groups
57
00:04:26,560 --> 00:04:35,585
such as yours can really drive and influence the direction of regulatory, acceptance of new technologies.
58
00:04:35,585 --> 00:04:45,330
But anyways, what do you see as the biggest challenges and the biggest opportunities facing pharmaceutical manufacturing in this
59
00:04:45,330 --> 00:04:47,810
age of AI powered technologies?
60
00:04:47,810 --> 00:04:48,370
On
61
00:04:48,370 --> 00:04:51,570
the challenge side, I think the biggest hurdle is trust.
62
00:04:51,570 --> 00:04:54,130
Patient safety is nonnegotiable in our industry.
63
00:04:54,130 --> 00:05:00,815
So when we talk about bringing AI into GXP environments, there's natural skepticism there.
64
00:05:00,815 --> 00:05:05,855
But we think about how do we validate these systems, how do we explain what the algorithm is doing.
65
00:05:05,855 --> 00:05:12,430
When a regulator asks, You know, we need to be able to explain, and how do we ensure that AI doesn't drift over time?
66
00:05:12,430 --> 00:05:15,150
That's one of the biggest challenges on the challenge side.
67
00:05:15,150 --> 00:05:16,990
There's also the cultural piece.
68
00:05:16,990 --> 00:05:26,785
We've built our quality systems on deterministic rule based thinking for decades, and now AI is introducing probabilistic decision making that can feel
69
00:05:26,785 --> 00:05:29,265
uncomfortable to some some of the users.
70
00:05:29,265 --> 00:05:34,225
Predictive quality is definitely, something that's part of the future.
71
00:05:34,225 --> 00:05:39,390
Instead of reacting to deviations after they happen, What if we could actually see them coming?
72
00:05:39,390 --> 00:05:48,030
AI lets us analyze vast amounts of process data, and it allows us to spot patterns that humans simply can't see as easily.
73
00:05:48,030 --> 00:05:57,845
So if we can use AI to streamline processes and make smarter decisions faster, that means getting, like, receiving medicines to patients sooner,
74
00:05:57,845 --> 00:05:58,965
and that's a real win.
75
00:05:58,965 --> 00:05:59,285
Those
76
00:05:59,285 --> 00:06:01,125
opportunities are really exciting.
77
00:06:01,125 --> 00:06:09,680
I guess that with with any new technology, there's there's always the question relation to guardrails that protect the patient without impeding that innovation.
78
00:06:09,680 --> 00:06:12,160
Samantha, maybe you touched on a little bit there.
79
00:06:12,160 --> 00:06:22,000
What do you think are the opportunities to update regulatory frameworks to keep pace with the rapid AI innovations that we're seeing, in the pharmaceutical sector?
80
00:06:22,000 --> 00:06:22,535
We
81
00:06:22,535 --> 00:06:24,695
always have to look for opportunities for convergence.
82
00:06:24,695 --> 00:06:33,735
We have to have conversations with our fellow regulators across the globe to make sure that there's, you know, as much of a singular voice as possible.
83
00:06:33,735 --> 00:06:39,960
You know, there are the points, you know, points to consider or just general guidelines that we published with EMA.
84
00:06:39,960 --> 00:06:46,040
We have to look across not just within, you know, within the scope of what the FDA regulates.
85
00:06:46,040 --> 00:06:46,840
You know?
86
00:06:46,840 --> 00:06:56,415
We we have to use our regulatory partners to make sure that we are giving the the industry a consistent message on how to integrate, this new
87
00:06:56,415 --> 00:06:57,135
technology.
88
00:06:57,135 --> 00:06:57,535
Tell
89
00:06:57,535 --> 00:07:03,695
me how your use of AI, is improving how you are training your students.
90
00:07:03,695 --> 00:07:05,900
So
91
00:07:05,980 --> 00:07:14,860
when it comes to AI, we started working with it very kind of oh, and not really early on per se because, you know, AI has been around since nineteen forties, technically speaking.
92
00:07:14,860 --> 00:07:15,180
Right?
93
00:07:15,180 --> 00:07:21,395
But what we're talking about when we use AI is, generally speaking, we're using, we started using generative AI.
94
00:07:21,395 --> 00:07:30,995
So what that means in our case is that in 2000 in 2023, all of our we started with our, apprentices, and we went, okay.
95
00:07:30,995 --> 00:07:36,810
How do we make sure to give them the readiness they need to work in the in the in the future for the future workforce.
96
00:07:36,810 --> 00:07:37,290
Right?
97
00:07:37,290 --> 00:07:40,570
And so what we did was we started identifying.
98
00:07:40,570 --> 00:07:48,410
We saw there was a there was a paper from the federal from the Fed that explained what the strong suits and weeks and weaknesses of are these are these tools.
99
00:07:48,410 --> 00:07:48,890
And we said, okay.
100
00:07:48,890 --> 00:07:50,335
Well, it seems we're weak.
101
00:07:50,335 --> 00:07:52,095
Research is a bit of a weak point.
102
00:07:52,095 --> 00:07:52,655
Why?
103
00:07:52,655 --> 00:07:59,855
And we knew that research is a very critical element, these students are learning, and these apprentices are learning how to really conduct appropriate research.
104
00:07:59,855 --> 00:08:01,695
And so we wanted to really test that.
105
00:08:01,695 --> 00:08:11,570
So we started doing what's called a parallel project where everything that they would do, they would have the they would train what, what we'll call an AI assistant to do at the same time.
106
00:08:11,570 --> 00:08:15,295
And so they would iterate this forth back and forth, and they've been doing this for years at point.
107
00:08:15,295 --> 00:08:17,775
So every cohort that comes on does the same thing.
108
00:08:17,775 --> 00:08:22,335
And so everything that they're doing, when they do it, they do it in parallel with the actual AI assistant.
109
00:08:22,335 --> 00:08:23,615
They create the assistant.
110
00:08:23,615 --> 00:08:24,735
They build the elements in there.
111
00:08:24,735 --> 00:08:32,920
And so that this way, what they're doing creates a little bit of that that tacit record kind of gets a little bit more under better understood.
112
00:08:32,920 --> 00:08:42,915
Because it also means that when, inevitably, the apprentice or intern leaves because we are an institute that takes in apprentices and interns on in six month, year, two
113
00:08:42,915 --> 00:08:44,355
year, they differ.
114
00:08:44,355 --> 00:08:46,995
And so a project, you know, might suffer from a gap.
115
00:08:46,995 --> 00:08:49,635
And so this allows the next person to kind of pop up
116
00:08:49,635 --> 00:08:50,275
and go, oh, look.
117
00:08:50,275 --> 00:08:50,435
Here's
118
00:08:50,435 --> 00:08:50,995
the assistant.
119
00:08:50,995 --> 00:08:52,595
Let's let's help them kinda carry along.
120
00:08:52,595 --> 00:08:54,195
It ends up kind of carrying these gaps forward.
121
00:08:54,195 --> 00:09:00,280
And so it ended up being a really useful tool to kind of carry along this tribal knowledge that normally kind of can get a little bit lost along the way.
122
00:09:00,280 --> 00:09:08,360
It also meant that in this way, they were also testing their own critical thinking skills because that was one of the elements that they found out early on is that it was bad.
123
00:09:08,360 --> 00:09:18,095
Early generative AI in in in 2023 was, you know, maybe at best, had to hit about a 10% ratio in terms of the
124
00:09:18,095 --> 00:09:21,855
success rate to actually get the link and the name and all the rest of materials right.
125
00:09:21,855 --> 00:09:24,140
Now maybe it's about 40%.
126
00:09:24,140 --> 00:09:25,020
It's getting better.
127
00:09:25,020 --> 00:09:26,220
And this is off the shelf tool.
128
00:09:26,220 --> 00:09:28,780
So when you really hone something, and that's what they learned.
129
00:09:28,780 --> 00:09:31,180
When they really honed something, they managed to make it better.
130
00:09:31,180 --> 00:09:33,020
But we're not just using it there.
131
00:09:33,020 --> 00:09:39,075
So that's what we started working with it there to make sure that all of our staff have a really good understanding as well as all of our apprentices.
132
00:09:39,075 --> 00:09:47,075
Anyone who really walks through our doors, we wanna make sure that they're a little bit better off, you know, understanding what the AI landscape looks like and how to integrate it into their
133
00:09:47,075 --> 00:09:47,875
workflows.
134
00:09:47,875 --> 00:09:48,195
What
135
00:09:48,195 --> 00:09:51,475
allowed this deployment to take place successfully?
136
00:09:51,475 --> 00:09:53,050
But,
137
00:09:53,050 --> 00:10:01,130
honestly, I think the secret sauce was treating this as a quality initiative that happened to use AI and not an AI initiative that happened to touch quality.
138
00:10:01,130 --> 00:10:08,905
So we the quality teams were actively involved in, deploying the the QA platform for play, for example.
139
00:10:08,905 --> 00:10:10,105
Several benefits.
140
00:10:10,105 --> 00:10:14,185
First of all, there's the, of course, the speed and the efficiency piece.
141
00:10:14,185 --> 00:10:22,230
So investigations, for example, which used to take hours or even days with manual tasks, that's actually,
142
00:10:23,590 --> 00:10:27,590
significantly decreased in terms of the the amount of time it takes.
143
00:10:27,590 --> 00:10:31,750
And beyond the speed, it's also the insights for quality.
144
00:10:31,750 --> 00:10:40,015
So AI can help us spot patterns and connect to the dots for, you know, thousands of data points, more quickly than a human can.
145
00:10:40,015 --> 00:10:42,495
And thirdly, there's the the consistency piece.
146
00:10:42,495 --> 00:10:51,220
So, human analysis can depend, highly on who's doing the review, their experience level, and even what time of day it is.
147
00:10:51,220 --> 00:10:57,380
The play tool brings a level of standardization to how we approach trending and root cause analysis.
148
00:10:57,380 --> 00:11:06,955
From your perspective, how should companies prepare for inspections where AI supported decision making is integral to GMP compliance?
149
00:11:06,955 --> 00:11:14,475
You know, are other new expectations for human oversight, for failure mode analysis, or or for real time monitoring?
150
00:11:14,475 --> 00:11:15,680
Even
151
00:11:15,680 --> 00:11:22,880
though we use the terminology of AI decision making, we have to remember that, by definition, AI is not making a decision.
152
00:11:22,880 --> 00:11:23,920
It's providing output.
153
00:11:23,920 --> 00:11:25,280
Human beings make decisions.
154
00:11:25,280 --> 00:11:35,055
And so, you know, when we look at our regulations, you know, we're we're never if software had an output and that output was used without
155
00:11:35,055 --> 00:11:37,615
human intervention, we would still call it output.
156
00:11:37,615 --> 00:11:44,735
Just because AI behaves more human like, we kind of start using this this nomenclature of decision making, but it's still output.
157
00:11:44,735 --> 00:11:48,470
And so, you know, ultimately, AI is a tool.
158
00:11:48,470 --> 00:11:51,990
It provides output on which decisions by people are made.
159
00:11:51,990 --> 00:12:01,695
And, you know, whether or not, you know, the output that is given by the AI is used without any additional human intervention or human in the
160
00:12:01,695 --> 00:12:02,255
loop.
161
00:12:02,255 --> 00:12:03,295
That is a decision.
162
00:12:03,295 --> 00:12:07,375
That's the decision, not the output that the that was given by the AI.
163
00:12:07,375 --> 00:12:10,335
And so I think that distinction needs to be clearly made.
164
00:12:10,335 --> 00:12:16,140
And if that distinction is clearly made, then then the thought process about inspections becomes easier actually.
165
00:12:16,140 --> 00:12:19,420
Because then you're still thinking about human beings.
166
00:12:19,420 --> 00:12:21,020
You're still thinking about record keeping.
167
00:12:21,020 --> 00:12:30,955
You're still thinking thinking is our is the output being given by this very advanced software still appropriate to maintain the quality of the product that comes out at the other
168
00:12:30,955 --> 00:12:31,435
end?
169
00:12:31,435 --> 00:12:41,320
The quality unit is ultimately responsible for the end product and for making sure that along the way, all the parts of the processes are operating the
170
00:12:41,320 --> 00:12:47,080
way they should be, whether it's validation, whether it's, you know, the specifications, whether it's the output, etcetera.
171
00:12:47,080 --> 00:12:56,755
And, you know, so I think, you know, once we frame the the use of the AI in the appropriate way, like, AI, yes, AI decision
172
00:12:56,755 --> 00:13:01,395
making, quote, unquote, it seems like it's making a decision, but it's really AI output.
173
00:13:01,395 --> 00:13:06,515
It's AI output and human decision making on what to use with that out how to use that output.
174
00:13:06,515 --> 00:13:07,170
In
175
00:13:07,170 --> 00:13:16,610
your presentation at the ISPE Facilities of the Future Conference, you talked about quality maturity index versus
176
00:13:16,930 --> 00:13:18,690
quality risk exposure.
177
00:13:18,690 --> 00:13:21,075
What exactly are they?
178
00:13:21,075 --> 00:13:23,475
And how are they complementary?
179
00:13:23,475 --> 00:13:29,235
And how is AI giving you better insights into these measures?
180
00:13:29,235 --> 00:13:30,035
QMI
181
00:13:30,035 --> 00:13:38,710
and QRE are essentially two metrics, that help us understand the complete picture of quality across our global network.
182
00:13:38,710 --> 00:13:48,695
So QMI or the Quality Maturity Index is essentially a digital analytics tool, within our platform, and it provides, dynamic
183
00:13:48,695 --> 00:13:57,975
data driven, essentially, KPIs that we consider to be signals for our quality maturity, at the site level.
184
00:13:57,975 --> 00:13:58,455
Many
185
00:13:58,455 --> 00:14:08,430
times, a deviation gets resolved, quote, unquote, by saying, well, it was operator error, and
186
00:14:08,430 --> 00:14:10,590
we retrained the operator.
187
00:14:10,590 --> 00:14:11,310
Yes.
188
00:14:11,310 --> 00:14:12,030
Okay?
189
00:14:12,030 --> 00:14:21,615
Now, so the training was inadequate to begin with, and we're gonna retrain the operator on a training system that was
190
00:14:21,615 --> 00:14:23,295
inadequate to begin with.
191
00:14:23,295 --> 00:14:25,215
That's the logic that I see here.
192
00:14:25,215 --> 00:14:32,850
But would your AI be able to kind of cry foul and say, wait a minute.
193
00:14:32,850 --> 00:14:37,330
The human error was either, okay.
194
00:14:37,330 --> 00:14:47,305
The individual really screwed up or the training was not good enough or the SOP is confusing or there's
195
00:14:47,305 --> 00:14:52,425
some other root cause in the manufacturing system that needs to be addressed.
196
00:14:52,425 --> 00:14:52,665
Yeah.
197
00:14:52,665 --> 00:14:56,185
Do you see AI getting us to that place?
198
00:14:56,185 --> 00:14:57,010
Oh,
199
00:14:57,010 --> 00:14:57,730
I mean, absolutely.
200
00:14:57,730 --> 00:14:59,490
That's exactly before.
201
00:14:59,490 --> 00:15:08,930
So so one of the things one of the first projects, you know, that because on the side, my brother and I, you know, we we we do consulting for generative AI.
202
00:15:08,930 --> 00:15:14,075
And one of the things we did was one of the first projects we did was on deviations, was around deviations.
203
00:15:14,075 --> 00:15:18,155
And the thing is, you know, this is something that we see commonly now.
204
00:15:18,155 --> 00:15:26,880
All of the big players all have AI built into looking at their deviations because it is just it's it's a great way to summarize a lot of elements and reduce a lot of the the the hours.
205
00:15:26,880 --> 00:15:29,440
It's a pretty low hanging fruit, generally.
206
00:15:29,440 --> 00:15:34,880
But the thing is that part that you touched on was is not something I hear a lot about.
207
00:15:34,880 --> 00:15:36,415
They always hear about, hey.
208
00:15:36,415 --> 00:15:39,135
I'm moving it, but they're like, ah, but we don't wanna solve the problem, though.
209
00:15:39,135 --> 00:15:39,295
That
210
00:15:39,295 --> 00:15:40,255
that part is a
211
00:15:40,255 --> 00:15:43,295
bit absent, which is unfortunate because I think it's such a missed opportunity.
212
00:15:43,295 --> 00:15:44,495
And I think you're right.
213
00:15:44,495 --> 00:15:53,510
I and I think this is one of the areas where, you know, we're going to see our when our deviations come out come into play, and we see that element that says, hey.
214
00:15:53,510 --> 00:15:54,790
This is a human error.
215
00:15:54,790 --> 00:15:59,350
We happen to have the training record that's very transparent, that's very that's very traceable test.
216
00:15:59,350 --> 00:15:59,830
Oh, look.
217
00:15:59,830 --> 00:16:04,790
Looks like they're we see right here that they're missing this specific pipetting skill.
218
00:16:04,790 --> 00:16:09,645
What we see now why this did not work, well, because they're not doing very well in this.
219
00:16:09,645 --> 00:16:15,565
Let's go back and retrain them on this specific area so we don't have to waste time with all the rest of this because this is the real weak point.
220
00:16:15,565 --> 00:16:19,120
Focus on that, and then we'll probably see a lot better results.
221
00:16:19,120 --> 00:16:23,680
Now I say probably because we haven't instituted that yet, but we will see that happen in the future.
222
00:16:23,680 --> 00:16:23,840
I'm
223
00:16:23,840 --> 00:16:24,400
certain of it.
224
00:16:24,400 --> 00:16:34,205
How is AI helping you move complaints and deviations from being a reactive process to a predictive process.
225
00:16:34,205 --> 00:16:35,645
It's
226
00:16:35,965 --> 00:16:41,245
more difficult for humans to to see the patterns and to, do this in an in a manual way.
227
00:16:41,245 --> 00:16:51,100
With play, it can analyze thousands of historical deviations and complaints and identify early warning signals that proceed major
228
00:16:51,100 --> 00:16:52,220
quality events.
229
00:16:52,220 --> 00:17:00,620
So maybe, for example, we'll see an uptick in a certain type of minor deviation that historically has led to more serious issues
230
00:17:01,835 --> 00:17:05,195
or a particular combination of fact of factors.
231
00:17:05,195 --> 00:17:09,675
For example, equipment, product line, shifting of time.
232
00:17:09,675 --> 00:17:12,315
So that correlates with problems down the road.
233
00:17:12,315 --> 00:17:16,280
So humans simply can't process that volume of data to spot the patterns.
234
00:17:16,280 --> 00:17:18,920
So AI is really allowing us to do that.
235
00:17:18,920 --> 00:17:19,480
Thinking
236
00:17:19,480 --> 00:17:29,480
about the regulatory filing, what are the most common challenges that reviewers might encounter when evaluating AI driven evidence or models in in
237
00:17:29,480 --> 00:17:30,520
those CMC submissions?
238
00:17:30,520 --> 00:17:31,215
As
239
00:17:31,215 --> 00:17:34,655
with anything challenging, it's, you know, something is new.
240
00:17:34,655 --> 00:17:38,815
There's always challenges of, you know, what what do we need to look at?
241
00:17:38,815 --> 00:17:40,495
How deeply do we need to look?
242
00:17:40,495 --> 00:17:50,450
But when we're looking at the credibility framework and the risk framework, we I think we need to make sure that products are well validated, that they're fit for use, that they're correct for
243
00:17:50,450 --> 00:17:55,010
the context of use, that the risk is appropriate for that part of the system.
244
00:17:55,010 --> 00:17:55,490
A
245
00:17:55,490 --> 00:18:04,175
number of companies, including CAI, where I sit on the board of directors, have created AI powered apps for,
246
00:18:04,655 --> 00:18:06,575
deviation report writing.
247
00:18:06,575 --> 00:18:13,855
How is Sanofi using AI for such tasks, and what benefits are you seeing?
248
00:18:13,855 --> 00:18:14,175
RA
249
00:18:14,175 --> 00:18:23,910
tools can help to draft the initial report sections and can suggest investigation pathways based on similar historical cases, and it can even help
250
00:18:23,910 --> 00:18:28,390
ensure that we're covering all the regulatory requirements for the investigation.
251
00:18:28,390 --> 00:18:37,265
But we're also applying this to regulatory submissions, quality assessments, and even internal quality reviews.
252
00:18:37,265 --> 00:18:47,260
So the AI tools that we're using can pull relevant data from multiple systems, identify key trends, and generate draft content that our quality professionals
253
00:18:47,260 --> 00:18:49,660
can then refine and and validate.
254
00:18:49,660 --> 00:18:50,220
How
255
00:18:50,300 --> 00:18:59,305
is the FDA addressing concerns about things like bias, transparency, and explainability in the models that are used for regulatory decision making.
256
00:18:59,305 --> 00:18:59,785
The
257
00:18:59,785 --> 00:19:08,425
training and their skill sets are still essential because, as with anything, if you're not trained well, you don't know what you don't know.
258
00:19:08,425 --> 00:19:17,370
While we use tools, some tools, we have I think, you know, everyone knows that Elsa is the tool that's used, that that FDA has,
259
00:19:18,090 --> 00:19:28,025
built, our AI, chatbot or, you know you know, and there are plug ins that can that can be utilized with Elsa to help with certain parts of
260
00:19:28,025 --> 00:19:31,065
the review or certain parts of looking at policy documents.
261
00:19:31,065 --> 00:19:40,540
We have a number of RAG libraries with documents where, you know, you can focus your inquiries so that there isn't, like, kind of this hallucination from all of the
262
00:19:40,540 --> 00:19:41,180
Internet.
263
00:19:41,180 --> 00:19:44,460
Although I'm assured that, Elsa is blocked off from the Internet.
264
00:19:44,460 --> 00:19:47,500
So if we upload a document, it's not going out into the wide world.
265
00:19:47,500 --> 00:19:50,060
So I wanna make sure that people do understand that as well.
266
00:19:50,060 --> 00:19:57,305
But it was trained on the Internet, so there is information that could potentially cause hallucinations.
267
00:19:57,305 --> 00:20:02,265
And so we have other tools where we can use it, where we can focus it on a RAG library.
268
00:20:02,265 --> 00:20:05,305
Like, this is our RAG library of quality policy.
269
00:20:05,305 --> 00:20:09,070
We want to make a new call policy statement about x.
270
00:20:09,070 --> 00:20:18,955
We can point to that regulatory that RAG library so that the output that we get is based on the information that's there and not and won't contain
271
00:20:18,955 --> 00:20:25,515
any of the noise that possibly opinion pieces on the Internet about how certain things should be regulated get seeked their way through.
272
00:20:25,515 --> 00:20:29,115
But that output still needs to be checked by a person.
273
00:20:29,115 --> 00:20:32,555
And, ultimately, that person who is doing the work is the one that's responsible.
274
00:20:32,555 --> 00:20:34,040
And so we have to say, okay.
275
00:20:34,040 --> 00:20:43,880
Do you still have the knowledge and skills to be able to look at this statement that comes out or this analysis that comes out if you're looking at data,
276
00:20:43,960 --> 00:20:47,240
graph, or whatever that comes out, table that comes out?
277
00:20:47,240 --> 00:20:52,725
Do you still have the skills and knowledge to look at it and say, yes.
278
00:20:52,725 --> 00:20:55,205
This makes sense, and this is what I want.
279
00:20:55,205 --> 00:20:55,685
Or, no.
280
00:20:55,685 --> 00:20:57,685
This actually doesn't make sense.
281
00:20:57,685 --> 00:21:02,005
And, you know, perhaps we need to change it or maybe I have to change my prompt.
282
00:21:02,005 --> 00:21:05,250
And so there's a lot of prompt engineering going on.
283
00:21:05,250 --> 00:21:14,690
You know, we have work groups to work on prompt engineering to make sure that the output is appropriately worded or formatted in the, appropriate way for the work that we're doing.
284
00:21:14,690 --> 00:21:21,865
There are different tools for different assessors, for different stages of assessment that are currently being developed and tested.
285
00:21:21,865 --> 00:21:23,705
Some of them have been deployed.
286
00:21:23,705 --> 00:21:25,785
Some of them, you know, are still being tested.
287
00:21:25,785 --> 00:21:35,610
And so I think that continuing to validate in the same way that we we want you to validate to make sure that the outputs that we could potentially use are
288
00:21:35,610 --> 00:21:37,850
still appropriate for the work that we do.
289
00:21:37,850 --> 00:21:47,695
But it comes down to whoever is the one who's who's utilizing that AI still needs to have that background, that historical knowledge, the training
290
00:21:47,695 --> 00:21:51,135
in order to make sure that the output is appropriate for the work.
