Listen to Audio
Listen Anywhere
April 2026
In this episode, Miguelina Matthews, Head of Quality Intelligence, Advocacy and Pharmacopoeia Affairs at Sanofi, joins the podcast to share how the company is leveraging artificial intelligence to shift pharmaceutical quality management from reactive investigations to predictive, data-driven insights. Learn more about the deployment of Sanofi's "Plai" tool, emphasizing the importance of robust governance, effective change management, and human-led decision-making in scaling AI technologies across global operations.
1
00:00:00,000 --> 00:00:13,000
Hello, and welcome to the ISPE podcast, Shaping the Future of Pharma.
2
00:00:13,000 --> 00:00:18,000
I'm Bob Chew, your host, and today we have another episode where we'll be sharing the
3
00:00:18,000 --> 00:00:25,000
latest insights and thought leadership on manufacturing, technology, supply chains,
4
00:00:25,000 --> 00:00:29,000
and regulatory trends impacting the pharmaceutical industry.
5
00:00:29,000 --> 00:00:34,000
You will hear directly from the innovators, experts, and professionals driving progress
6
00:00:34,000 --> 00:00:36,000
and shaping the future.
7
00:00:36,000 --> 00:00:42,000
Thank you again for joining us, and now let's dive into this episode.
8
00:00:42,000 --> 00:00:48,000
Our topic today is the use of AI quality transformation at Sanofi.
9
00:00:48,000 --> 00:00:53,000
To share more about this topic, I would like to welcome Miguelina Matthews, who is the
10
00:00:53,000 --> 00:01:00,000
head of Quality Intelligence, Advocacy, and Pharmacopeia Affairs at Sanofi.
11
00:01:00,000 --> 00:01:06,000
Miguelina gave a presentation at the recent ISPE Facilities of the Future Conference
12
00:01:06,000 --> 00:01:09,000
on AI transformation at Sanofi.
13
00:01:09,000 --> 00:01:13,000
Welcome to this podcast, Shaping the Future of Pharma.
14
00:01:13,000 --> 00:01:20,000
First, before we dive into how Sanofi is using AI, would you give us a bit of insight into
15
00:01:20,000 --> 00:01:24,000
your role at Sanofi?
16
00:01:24,000 --> 00:01:26,000
First of all, thank you so much for having me on the show.
17
00:01:26,000 --> 00:01:28,000
I'm excited to be here.
18
00:01:28,000 --> 00:01:32,000
So I'm the head of Quality Intelligence, Advocacy, and Pharmacopeia Affairs at Sanofi, and I
19
00:01:32,000 --> 00:01:37,000
know that's a mouthful, but it's a role that sits right at the intersection of regulatory
20
00:01:37,000 --> 00:01:40,000
science and quality strategy.
21
00:01:40,000 --> 00:01:46,000
And increasingly these days, it's also focusing on digital innovation as well.
22
00:01:46,000 --> 00:01:49,000
So on the intelligence side of things, my team essentially keeps our fingers on the
23
00:01:49,000 --> 00:01:52,000
pulse of what's coming down the pipeline.
24
00:01:52,000 --> 00:01:59,000
We're constantly monitoring emerging regulation requirements and industry trends that could
25
00:01:59,000 --> 00:02:03,000
impact our GXP processes across the enterprise.
26
00:02:03,000 --> 00:02:06,000
It's like being an early warning system for quality.
27
00:02:06,000 --> 00:02:10,000
We want to make sure Sanofi is always ahead of the curve.
28
00:02:10,000 --> 00:02:14,000
For the advocacy piece, we get to be more proactive.
29
00:02:14,000 --> 00:02:18,000
We work to influence both industry standards and the health authorities on topics that
30
00:02:18,000 --> 00:02:20,000
really matter to us.
31
00:02:20,000 --> 00:02:25,000
And one area that's particularly exciting right now is how we can responsibly integrate
32
00:02:25,000 --> 00:02:29,000
artificial intelligence into our GXP processes.
33
00:02:29,000 --> 00:02:34,000
It's fascinating work because we're literally helping to shape the future of how the industry
34
00:02:34,000 --> 00:02:36,000
operates.
35
00:02:36,000 --> 00:02:39,000
And the combination really keeps things dynamic for us.
36
00:02:39,000 --> 00:02:44,000
One day we're analyzing guidances, and then the next we're presenting at an industry conference
37
00:02:44,000 --> 00:02:47,000
about AI and applications and quality systems.
38
00:02:47,000 --> 00:02:52,000
It's the kind of role where you never stop learning.
39
00:02:52,000 --> 00:02:59,000
You know, I've been with ISPE for maybe 30 years now, or even more, and it's amazing
40
00:02:59,000 --> 00:03:07,000
how initiatives and groups such as yours can really drive and influence the direction of
41
00:03:07,000 --> 00:03:13,000
regulatory acceptance of new technologies.
42
00:03:13,000 --> 00:03:19,000
But anyways, what do you see as the biggest challenges and the biggest opportunities facing
43
00:03:19,000 --> 00:03:25,000
pharmaceutical manufacturing in this age of AI-powered technologies?
44
00:03:25,000 --> 00:03:29,000
That's a really good question, and it's something I've been thinking about a lot in the past
45
00:03:29,000 --> 00:03:34,000
couple years, especially now that AI is really booming everywhere.
46
00:03:34,000 --> 00:03:39,000
On the challenge side, I think the biggest hurdle is trust.
47
00:03:39,000 --> 00:03:42,000
Patient safety is non-negotiable in our industry.
48
00:03:42,000 --> 00:03:49,000
So when we talk about bringing AI into GXP environments, there's natural skepticism there.
49
00:03:49,000 --> 00:03:52,000
So we think about how do we validate these systems?
50
00:03:52,000 --> 00:03:56,000
How do we explain what the algorithm is doing?
51
00:03:56,000 --> 00:04:00,000
When a regulator asks, we need to be able to explain.
52
00:04:00,000 --> 00:04:03,000
And how do we ensure that AI doesn't drift over time?
53
00:04:03,000 --> 00:04:09,000
So that's one of the biggest challenges on the challenge side.
54
00:04:09,000 --> 00:04:11,000
There's also the cultural piece.
55
00:04:11,000 --> 00:04:17,000
We've built our quality systems on deterministic, rule-based thinking for decades.
56
00:04:17,000 --> 00:04:23,000
And now AI is introducing probabilistic decision making that can feel uncomfortable to some
57
00:04:23,000 --> 00:04:24,000
of the users.
58
00:04:24,000 --> 00:04:29,000
So getting people to understand and embrace the shift while maintaining our rigorous standards
59
00:04:29,000 --> 00:04:32,000
is a real challenge.
60
00:04:32,000 --> 00:04:36,000
And then additionally, the regulatory landscape is still catching up.
61
00:04:36,000 --> 00:04:39,000
We're seeing drafts from different health authorities.
62
00:04:39,000 --> 00:04:44,000
And we're all figuring this out together, the industry, the health authorities, and
63
00:04:44,000 --> 00:04:45,000
everyone.
64
00:04:45,000 --> 00:04:49,000
But there are real opportunities, and they're enormous.
65
00:04:49,000 --> 00:04:54,000
Predictive quality is definitely something that's part of the future.
66
00:04:54,000 --> 00:04:58,000
Instead of reacting to deviations after they happen, what if we could actually see them
67
00:04:58,000 --> 00:04:59,000
coming?
68
00:05:00,000 --> 00:05:06,000
AI lets us analyze vast amounts of process data, and it allows us to spot patterns that
69
00:05:06,000 --> 00:05:11,000
humans simply can't see as easily.
70
00:05:11,000 --> 00:05:18,000
There's also the efficiency piece, document review, batch record analysis, deviation trending.
71
00:05:18,000 --> 00:05:21,000
This takes hours for humans to do manually.
72
00:05:21,000 --> 00:05:28,000
So AI can compress this time dramatically, freeing up our quality professionals for higher
73
00:05:28,000 --> 00:05:30,000
value activities.
74
00:05:30,000 --> 00:05:36,000
And most importantly, there's real potential to speed products to patients.
75
00:05:36,000 --> 00:05:41,000
So if we can use AI to streamline processes and make smarter decisions faster, that means
76
00:05:41,000 --> 00:05:46,000
getting life-saving medicines to patients sooner.
77
00:05:46,000 --> 00:05:48,000
And that's a real win.
78
00:05:49,000 --> 00:05:59,000
Well, you shared a use case called PLAY, P-L-A-I, at the ISPE conference.
79
00:05:59,000 --> 00:06:05,000
And I understand this tool is used at scale at Sanofi.
80
00:06:05,000 --> 00:06:10,000
Can you share how the tool was taken from pilot to at scale?
81
00:06:10,000 --> 00:06:11,000
Absolutely.
82
00:06:11,000 --> 00:06:17,000
So PLAY is a great example of how we approached scaling AI at Sanofi.
83
00:06:17,000 --> 00:06:20,000
So what we did is we started small and smart.
84
00:06:20,000 --> 00:06:26,000
The initial pilot focused on one specific use case, where we were trying to automate
85
00:06:26,000 --> 00:06:32,000
deviations and complaint trending, while providing root cause suggestions.
86
00:06:32,000 --> 00:06:37,000
So this was traditionally, I think it's the same case in many other companies, where traditionally
87
00:06:37,000 --> 00:06:43,000
it's a very manual process, where the quality teams spend hours looking for patterns across
88
00:06:44,000 --> 00:06:48,000
different types of events, and they spend time brainstorming what the potential root
89
00:06:48,000 --> 00:06:50,000
causes could be.
90
00:06:50,000 --> 00:06:56,000
But within a few months, we could see concrete results with PLAY, with faster identification
91
00:06:56,000 --> 00:06:59,000
of trending issues.
92
00:06:59,000 --> 00:07:06,000
And AI could generate potential root cause hypotheses that allowed the investigators
93
00:07:06,000 --> 00:07:10,000
to explore further.
94
00:07:10,000 --> 00:07:14,000
But the difference is, the real difference is that we didn't just focus on the technology.
95
00:07:14,000 --> 00:07:18,000
From the start, we built change management into the process.
96
00:07:18,000 --> 00:07:26,000
We involved end users to help design and identify issues with the tool.
97
00:07:26,000 --> 00:07:31,000
And we were transparent with what the tool could and couldn't do.
98
00:07:31,000 --> 00:07:39,000
And what also enabled us to bring this tool to scale was our governance framework, which
99
00:07:39,000 --> 00:07:41,000
we call RAISE.
100
00:07:41,000 --> 00:07:48,000
So RAISE stands for Responsible AI at Sanofi Enterprise, and it really provides the guardrails
101
00:07:48,000 --> 00:07:55,000
for how we can develop, validate, and deploy AI tools across our organization.
102
00:07:55,000 --> 00:08:00,000
So this governance ensures we're addressing risk, ethics, and compliance systematically
103
00:08:00,000 --> 00:08:04,000
with any new tool.
104
00:08:04,000 --> 00:08:10,000
So today, PLAY is used across multiple manufacturing sites globally.
105
00:08:10,000 --> 00:08:15,000
And it's an available toolkit, not only at the site level, but at the global level.
106
00:08:15,000 --> 00:08:23,000
And really, overall, the lesson was to start focused, prove value, govern responsibly,
107
00:08:23,000 --> 00:08:25,000
and then scale systematically.
108
00:08:26,000 --> 00:08:35,000
Well, that sounds to be kind of aligned with a talk I heard recently from an FDA person
109
00:08:35,000 --> 00:08:38,000
on this same podcast.
110
00:08:38,000 --> 00:08:44,000
You just work within your existing quality system framework, but you understand what
111
00:08:44,000 --> 00:08:51,000
the technology can do, and you build it up one step at a time.
112
00:08:51,000 --> 00:08:57,000
So what allowed this deployment to take place successfully?
113
00:08:57,000 --> 00:09:02,000
So in my opinion, I think there are really four success factors that allowed us to make
114
00:09:02,000 --> 00:09:04,000
this work for us.
115
00:09:04,000 --> 00:09:11,000
So first of all, we had strong support from our leadership from the beginning.
116
00:09:11,000 --> 00:09:15,000
This wasn't just something that was driven by IT or one specific department.
117
00:09:15,000 --> 00:09:21,000
It was really all the leaders throughout the company that were supporting this effort and
118
00:09:21,000 --> 00:09:25,000
could see the potential value.
119
00:09:25,000 --> 00:09:31,000
Second of all, as I mentioned, that RAISE governance framework was absolutely essential.
120
00:09:31,000 --> 00:09:32,000
We needed to make sure.
121
00:09:32,000 --> 00:09:36,000
We're a highly regulated industry, and our patients are in the forefront.
122
00:09:36,000 --> 00:09:41,000
So we needed to ensure that we had a framework and guardrails around the tools that we were
123
00:09:41,000 --> 00:09:43,000
deploying.
124
00:09:43,000 --> 00:09:48,000
Third, we really focused on the user experience.
125
00:09:48,000 --> 00:09:54,000
So getting the input from the pilot sites and the end users was absolutely essential.
126
00:09:54,000 --> 00:09:59,000
And we also built some change management into the process to ensure that the users were
127
00:09:59,000 --> 00:10:05,000
embarked and understood the tools that they were using.
128
00:10:05,000 --> 00:10:14,000
And finally, we were able to provide the value and the ROI concretely and quickly.
129
00:10:14,000 --> 00:10:18,000
But honestly, I think the secret sauce was treating this as a quality initiative that
130
00:10:18,000 --> 00:10:22,000
happened to use AI and not an AI initiative that happened to touch quality.
131
00:10:22,000 --> 00:10:29,000
So the quality teams were actively involved in deploying the QA platform for Play, for
132
00:10:29,000 --> 00:10:33,000
example.
133
00:10:33,000 --> 00:10:41,000
So what benefits are you now seeing through the use of this Play, PL-AI?
134
00:10:41,000 --> 00:10:44,000
So we're seeing several benefits.
135
00:10:44,000 --> 00:10:49,000
First of all, there's, of course, the speed and the efficiency piece.
136
00:10:49,000 --> 00:10:57,000
So investigations, for example, which used to take hours or even days with manual tasks,
137
00:10:57,000 --> 00:11:05,000
looking for patterns, that's actually significantly decreased in terms of the amount of time it
138
00:11:05,000 --> 00:11:09,000
takes.
139
00:11:09,000 --> 00:11:13,000
And beyond the speed, it's also the insights for quality.
140
00:11:13,000 --> 00:11:20,000
So AI can help us spot patterns and connect the dots for thousands of data points more
141
00:11:20,000 --> 00:11:24,000
quickly than a human can.
142
00:11:24,000 --> 00:11:26,000
And thirdly, there's the consistency piece.
143
00:11:26,000 --> 00:11:33,000
So human analysis can depend highly on who's doing the review, their experience level,
144
00:11:33,000 --> 00:11:35,000
and even what time of day it is.
145
00:11:36,000 --> 00:11:41,000
The Play tool brings a level of standardization to how we approach trending and root cause
146
00:11:41,000 --> 00:11:43,000
analysis.
147
00:11:43,000 --> 00:11:46,000
So that's a huge benefit as well.
148
00:11:46,000 --> 00:11:49,000
And it's elevated the role of the quality professionals, in my opinion.
149
00:11:49,000 --> 00:11:57,000
So instead of spending time on the mechanical work and the manual work, the investigators
150
00:11:57,000 --> 00:12:03,000
can really focus on the critical piece of assessing the outcome.
151
00:12:03,000 --> 00:12:09,000
So overall, we're identifying quality signals faster, investigating more thoroughly and
152
00:12:09,000 --> 00:12:13,000
consistently, and ultimately addressing potential issues earlier.
153
00:12:13,000 --> 00:12:16,000
Well, that's really great.
154
00:12:16,000 --> 00:12:21,000
That sounds like a really good buildup and deployment.
155
00:12:21,000 --> 00:12:28,000
Now in your presentation at the ISPE Facilities of the Future Conference, you talked about
156
00:12:29,000 --> 00:12:35,000
quality maturity index versus quality risk exposure.
157
00:12:35,000 --> 00:12:37,000
What exactly are they?
158
00:12:37,000 --> 00:12:39,000
And how are they complementary?
159
00:12:39,000 --> 00:12:44,000
And how is AI giving you better insights into these measures?
160
00:12:44,000 --> 00:12:47,000
Oh, thank you for the question.
161
00:12:47,000 --> 00:12:55,000
So these are, QMI and QRE are essentially two metrics that help us understand the complete
162
00:12:55,000 --> 00:12:58,000
picture of quality across our global network.
163
00:12:59,000 --> 00:13:05,000
So QMI, or the Quality Maturity Index, is essentially a digital analytics tool within
164
00:13:05,000 --> 00:13:06,000
our platform.
165
00:13:06,000 --> 00:13:15,000
And it provides dynamic, data-driven, essentially KPIs that we consider to be signals for our
166
00:13:15,000 --> 00:13:18,000
quality maturity at the site level.
167
00:13:18,000 --> 00:13:27,000
So in practical terms, QMI will enable us to evaluate specific areas requiring attention.
168
00:13:27,000 --> 00:13:39,000
So for example, it takes into account deviations, repeat deviations, Kappas, and it also allows
169
00:13:39,000 --> 00:13:43,000
us to benchmark across different sites.
170
00:13:43,000 --> 00:13:48,000
So it's essentially a way of monitoring KPIs.
171
00:13:48,000 --> 00:13:56,000
The difference between QMI and QRE is that the quality risk exposure, it's also a data
172
00:13:56,000 --> 00:14:03,000
analytics tool platform, but it actually takes into account additional factors, mainly some
173
00:14:03,000 --> 00:14:08,000
external inspection trends, audit outcomes.
174
00:14:08,000 --> 00:14:11,000
So it adds an additional layer.
175
00:14:11,000 --> 00:14:18,000
It also includes, considers the type of processes that occur at a given site.
176
00:14:18,000 --> 00:14:25,000
So with the quality risk exposure, again, it just takes into account additional factors
177
00:14:25,000 --> 00:14:32,000
that will then allow us to identify the potential risk at a given site.
178
00:14:32,000 --> 00:14:42,000
So in short, QMI scores the site's quality maturity from zero to 100, while the QRE considers
179
00:14:42,000 --> 00:14:44,000
external factors.
180
00:14:44,000 --> 00:14:55,000
And together, what makes them really powerful is that it allows us to identify where we're
181
00:14:55,000 --> 00:15:04,000
most at risk and to prioritize actions accordingly.
182
00:15:04,000 --> 00:15:18,000
So if I understand correctly, QMI is focused on true quality impacting items, whereas QRE
183
00:15:18,000 --> 00:15:25,000
layers on regulatory, which is supposed to just be quality impacting, but we know that's
184
00:15:25,000 --> 00:15:26,000
not always the case.
185
00:15:26,000 --> 00:15:29,000
Do I have that right?
186
00:15:29,000 --> 00:15:30,000
I think so, yes.
187
00:15:30,000 --> 00:15:32,000
So you're right with QMI.
188
00:15:32,000 --> 00:15:38,000
It's essentially the quality indicators that traditionally are tracked, the deviations
189
00:15:38,000 --> 00:15:39,000
and Kappas.
190
00:15:39,000 --> 00:15:47,000
But the QRE takes into account, for example, external trends, and I'm just making one up.
191
00:15:47,000 --> 00:15:52,000
So maybe we're seeing a trend in visual inspection observations.
192
00:15:52,000 --> 00:15:57,000
And we know that certain sites perform visual inspection activities.
193
00:15:57,000 --> 00:16:02,000
And if you put those two together, and let's say that site is having a lot of deviations
194
00:16:02,000 --> 00:16:06,000
at the same time, that could increase the quality risk exposure.
195
00:16:06,000 --> 00:16:07,000
Okay.
196
00:16:07,000 --> 00:16:11,000
I hope that's a little clearer.
197
00:16:11,000 --> 00:16:20,000
So how is AI helping you move complaints and deviations from being a reactive process to
198
00:16:20,000 --> 00:16:22,000
a predictive process?
199
00:16:22,000 --> 00:16:25,000
Thank you.
200
00:16:25,000 --> 00:16:27,000
That's also a really good question.
201
00:16:27,000 --> 00:16:32,000
So traditionally, the biopharma approach to deviations and complaints has been, as you
202
00:16:32,000 --> 00:16:35,000
mentioned, inherently reactive.
203
00:16:35,000 --> 00:16:40,000
Something goes wrong, we investigate, we find the most probable root cause, and then we
204
00:16:40,000 --> 00:16:42,000
implement corrective actions.
205
00:16:42,000 --> 00:16:49,000
And it's a solid process, but we're always responding after the fact, after the deviation
206
00:16:49,000 --> 00:16:53,000
has occurred or after the complaint has been filed.
207
00:16:53,000 --> 00:16:59,000
So I feel that AI is really flipping this model on its head.
208
00:16:59,000 --> 00:17:03,000
First of all, the AI allows us to have pattern recognition in place.
209
00:17:03,000 --> 00:17:08,000
As I mentioned before, it's a little, it's more difficult for humans to see the patterns
210
00:17:08,000 --> 00:17:12,000
and to do this in a manual way.
211
00:17:12,000 --> 00:17:18,000
With Play, it can analyze thousands of historical deviations and complaints and identify early
212
00:17:18,000 --> 00:17:23,000
warning signals that precede major quality events.
213
00:17:23,000 --> 00:17:28,000
So maybe, for example, we'll see an uptick in a certain type of minor deviation that
214
00:17:28,000 --> 00:17:35,000
historically has led to more serious issues, or a particular combination of factors.
215
00:17:35,000 --> 00:17:40,000
For example, equipment, product line, shifting of time.
216
00:17:40,000 --> 00:17:43,000
So that correlates with problems down the road.
217
00:17:43,000 --> 00:17:47,000
So humans simply can't process that volume of data to spot the patterns.
218
00:17:47,000 --> 00:17:51,000
So AI is really allowing us to do that.
219
00:17:51,000 --> 00:17:59,000
Second, we're using AI to do what we call, or what I call, a weak signal detection.
220
00:17:59,000 --> 00:18:03,000
So not every complaint or deviation is created equal.
221
00:18:03,000 --> 00:18:09,000
Some are isolated incidents, but others are the first indicator of a systematic issue.
222
00:18:09,000 --> 00:18:13,000
So AI helps us distinguish between noise and signal much faster.
223
00:18:13,000 --> 00:18:18,000
So we can investigate and intervene before a trend fully develops.
224
00:18:18,000 --> 00:18:22,000
And thirdly, there's the root cause prediction capability.
225
00:18:22,000 --> 00:18:28,000
When a deviation occurs, instead of starting from scratch, the AI tool can say, you know,
226
00:18:28,000 --> 00:18:34,000
based on similar events across the network, here are the most likely root causes to investigate
227
00:18:34,000 --> 00:18:35,000
first.
228
00:18:35,000 --> 00:18:44,000
That accelerates our investigation and gets us to the corrective action faster.
229
00:18:44,000 --> 00:18:53,000
So 60 years ago, Edward Deming took his statistical quality methods into manufacturing, and in
230
00:18:53,000 --> 00:18:57,000
Japan first, because they were the ones that would listen.
231
00:18:57,000 --> 00:19:00,000
And there were control charts.
232
00:19:00,000 --> 00:19:10,000
I'm sure, does Sanofi use control charts, and is this AI basically allowing you to automatically
233
00:19:10,000 --> 00:19:19,000
interpret drifts and trends and such that the human eye used to be doing with the visual
234
00:19:19,000 --> 00:19:20,000
control charts?
235
00:19:20,000 --> 00:19:22,000
Is that what we're talking about?
236
00:19:22,000 --> 00:19:23,000
Yes.
237
00:19:24,000 --> 00:19:31,000
Play is essentially a dashboard, and it does exactly what you just described.
238
00:19:31,000 --> 00:19:36,000
And what's kind of cool about it is that you can visualize this on your computer, and it's
239
00:19:36,000 --> 00:19:44,000
also available as an app that any Sanofian can download and look at at any point in time.
240
00:19:44,000 --> 00:19:45,000
Okay.
241
00:19:45,000 --> 00:19:53,000
A number of companies, including CAI, where I sit on the board of directors, have created
242
00:19:53,000 --> 00:19:58,000
AI-powered apps for deviation report writing.
243
00:19:58,000 --> 00:20:07,000
How is Sanofi using AI for such tasks, and what benefits are you seeing?
244
00:20:07,000 --> 00:20:12,000
So we're using AI across several reporting functions.
245
00:20:12,000 --> 00:20:17,000
So, the creation and completion of the investigation reports is definitely a big one.
246
00:20:17,000 --> 00:20:23,000
So the investigation and reporting can be incredibly time-intensive, especially for
247
00:20:23,000 --> 00:20:25,000
really complicated investigations.
248
00:20:25,000 --> 00:20:32,000
So RA tools can help to draft the initial report sections and can suggest investigation
249
00:20:32,000 --> 00:20:37,000
pathways based on similar historical cases, and it can even help ensure that we're covering
250
00:20:37,000 --> 00:20:43,000
all the regulatory requirements for the investigation.
251
00:20:43,000 --> 00:20:49,000
But we're also applying this to regulatory submissions, quality assessments, and even
252
00:20:49,000 --> 00:20:52,000
internal quality reviews.
253
00:20:52,000 --> 00:20:57,000
So the AI tools that we're using can pull relevant data from multiple systems, identify
254
00:20:57,000 --> 00:21:03,000
key trends, and generate draft content that our quality professionals can then refine
255
00:21:03,000 --> 00:21:06,000
and validate.
256
00:21:06,000 --> 00:21:08,000
So the benefits are really compelling.
257
00:21:08,000 --> 00:21:14,000
There's the obvious time savings, but we're talking about reducing report writing, you
258
00:21:14,000 --> 00:21:18,000
know, 40 to 50 percent in many cases.
259
00:21:18,000 --> 00:21:19,000
But it's not just about speed.
260
00:21:19,000 --> 00:21:24,000
The consistency piece, which I mentioned earlier, is also really huge.
261
00:21:24,000 --> 00:21:31,000
It enables us to follow standard formats, making sure that all the required elements
262
00:21:31,000 --> 00:21:37,000
are covered, and that we maintain a consistent tone and approach across different authors
263
00:21:37,000 --> 00:21:39,000
and sites.
264
00:21:39,000 --> 00:21:45,000
And that's particularly valuable for regulatory submissions, where consistency really matters.
265
00:21:45,000 --> 00:21:48,000
And then overall, there's the quality improvement aspect.
266
00:21:48,000 --> 00:21:52,000
AI can suggest additional data points to include.
267
00:21:52,000 --> 00:21:59,000
It can flag potential gaps in the investigation, and it can recommend similar cases to reference.
268
00:21:59,000 --> 00:22:06,000
So it's having a really experienced quality professional looking over your shoulder is
269
00:22:06,000 --> 00:22:09,000
basically what the tool can feel like.
270
00:22:09,000 --> 00:22:12,000
It's making sure you're not missing anything.
271
00:22:12,000 --> 00:22:18,000
But the key, I just want to add, the key is that we're not replacing human judgment.
272
00:22:18,000 --> 00:22:20,000
The tools are really augmenting it.
273
00:22:20,000 --> 00:22:25,000
So the AI handles the heavy lifting and the data gathering, but it's the human in the
274
00:22:25,000 --> 00:22:30,000
end that's making the decisions.
275
00:22:30,000 --> 00:22:38,000
AI is a statistical tool, and the industry relies on statistics, everything from drug
276
00:22:38,000 --> 00:22:42,000
development and clinical trials.
277
00:22:42,000 --> 00:22:52,000
Do we think that AI could, given sufficient data, look at a deviation and assess whether
278
00:22:52,000 --> 00:23:01,000
it was really just operator error or really just an incorrect laboratory result, and instead
279
00:23:01,000 --> 00:23:04,000
point to the true root cause?
280
00:23:04,000 --> 00:23:11,000
Or if it was operator error, be able to say, you know, you tried retraining in the past
281
00:23:11,000 --> 00:23:13,000
and it didn't work.
282
00:23:13,000 --> 00:23:18,000
Consider process improvement or at least a different approach to operator training.
283
00:23:18,000 --> 00:23:19,000
So said AI.
284
00:23:19,000 --> 00:23:20,000
What do you think?
285
00:23:21,000 --> 00:23:22,000
Yes.
286
00:23:22,000 --> 00:23:28,000
So in the case of play, it can look across thousands of similar deviations and corresponding
287
00:23:28,000 --> 00:23:32,000
kappas and say exactly what you just said, Bob.
288
00:23:32,000 --> 00:23:37,000
It can say, wait a minute, you've attributed this type of event to operator error five
289
00:23:37,000 --> 00:23:42,000
times in the past, in the past two years, and you've done retraining each time, but
290
00:23:42,000 --> 00:23:43,000
the pattern keeps recurring.
291
00:23:43,000 --> 00:23:49,000
So it does alert the investigator that, hmm, maybe it's time to investigate further and
292
00:23:49,000 --> 00:23:55,000
this isn't really, the kappa wasn't effective or it wasn't really the right root cause.
293
00:23:55,000 --> 00:24:03,000
So it can flag that maybe that root cause was not appropriate the last time.
301
00:24:25,000 --> 00:24:32,000
Are you able to discuss additional applications of AI beyond what you discussed in your presentation
302
00:24:32,000 --> 00:24:35,000
at Facilities of the Future conference?
303
00:24:35,000 --> 00:24:41,000
So at Sanofi, we have a few AI initiatives underway across the enterprise and it's a
304
00:24:41,000 --> 00:24:45,000
really dynamic space for us.
305
00:24:45,000 --> 00:24:50,000
But for today, I think I'd rather keep the focus on play and the quality intelligence
306
00:24:50,000 --> 00:24:52,000
applications we discussed.
307
00:24:52,000 --> 00:24:57,000
There's so much depth there and honestly, these are the use cases where we've seen the
308
00:24:57,000 --> 00:25:05,000
most mature implementation and measurable impact.
309
00:25:05,000 --> 00:25:12,000
What I can say is that we are using AI broadly and we, like I said earlier, we start with
310
00:25:12,000 --> 00:25:18,000
a clear business problem, prove the value in a pilot, engage our stakeholders, leverage
311
00:25:18,000 --> 00:25:24,000
our raised governance framework, and then scale systematically so that the consistency
312
00:25:24,000 --> 00:25:26,000
has been key to our success.
313
00:25:26,000 --> 00:25:33,000
Well, this has been a great story, but let's look at some of the other presentations from
314
00:25:33,000 --> 00:25:36,000
Facilities of the Future conference.
315
00:25:36,000 --> 00:25:43,000
Which ones do you recall as being especially memorable, exciting, impactful to you?
316
00:25:43,000 --> 00:25:44,000
Yes.
317
00:25:44,000 --> 00:25:48,000
So there was one presentation that really stood out to me.
318
00:25:48,000 --> 00:25:54,000
It was titled Realizing the Potential of Digital Twins in ATMPs.
319
00:25:54,000 --> 00:26:00,000
And what I found fascinating was how they articulated the unique complexity of ATMP
320
00:26:00,000 --> 00:26:01,000
manufacturing.
321
00:26:01,000 --> 00:26:07,000
You know, the patient-specific execution, the really small batch sizes, how compartmentalized
322
00:26:07,000 --> 00:26:13,000
the manufacturing can be with frequent changeovers and manual steps.
323
00:26:13,000 --> 00:26:20,000
But they explained how digital twins could really enhance and provide benefits to that
324
00:26:20,000 --> 00:26:22,000
type of manufacturing.
325
00:26:22,000 --> 00:26:30,000
So they shared how equipment optimization through leveraging digital twins.
326
00:26:30,000 --> 00:26:32,000
And that was really eye-opening to me.
327
00:26:32,000 --> 00:26:40,000
They showed how using average growth rates to estimate equipment needs can be grossly
328
00:26:40,000 --> 00:26:41,000
underestimated.
329
00:26:41,000 --> 00:26:49,000
But what really resonated with me was their progression from holistic models to digital
330
00:26:49,000 --> 00:26:50,000
shadows.
331
00:26:50,000 --> 00:26:51,000
They didn't just theorize.
332
00:26:51,000 --> 00:26:54,000
They actually showed measurable results.
333
00:26:54,000 --> 00:26:59,000
And I also appreciated their candor about the regulatory considerations.
334
00:26:59,000 --> 00:27:06,000
They were clear that while digital twins for supply chain, scheduling, and facility design
335
00:27:06,000 --> 00:27:12,000
face minimal regulatory hurdles, applying them to manufacturing process
338
00:27:18,000 --> 00:27:23,000
But I felt that their presentation connected really well with Clay in terms of using digital
339
00:27:23,000 --> 00:27:26,000
tools for predicting quality.
340
00:27:26,000 --> 00:27:35,000
Well, I'll give you my personal view of digital twins, both equipment digital twins,
341
00:27:35,000 --> 00:27:43,000
which by the way, aircraft industry has used for a long time to model jet engines and what
342
00:27:43,000 --> 00:27:49,000
happens if a fan blade blows and, you know, modeling all of that, that's been used for
343
00:27:49,000 --> 00:27:50,000
a long time.
344
00:27:50,000 --> 00:27:53,000
And also, process digital twins.
345
00:27:53,000 --> 00:28:00,000
I could imagine during startup, you train the equipment digital twin at the factory
346
00:28:00,000 --> 00:28:02,000
acceptance test.
347
00:28:02,000 --> 00:28:08,000
You then disassemble the machine, bring it to the site, reassemble it, plug in the digital
348
00:28:08,000 --> 00:28:18,000
twin, run the machine a little bit, and the digital twin says, this whatever needs a little
349
00:28:18,000 --> 00:28:23,000
adjustment and this isn't connected properly over here.
350
00:28:23,000 --> 00:28:30,000
And then you fix that, and then the digital twin says, okay, this equipment is now working
351
00:28:30,000 --> 00:28:36,000
as it was at factory, and of course, you did all this testing at the factory, and it helps
352
00:28:36,000 --> 00:28:41,000
you get through the startup and setting to work phase much faster.
353
00:28:41,000 --> 00:28:48,000
Maybe you can even have it run through the automation and give you an instant qualification
354
00:28:48,000 --> 00:28:50,000
decision.
355
00:28:50,000 --> 00:28:58,000
And then transferring processes, again, using a process digital twin to say that, yes, this
356
00:28:58,000 --> 00:29:04,000
process is producing product the way that it was at wherever you're transferring from,
357
00:29:04,000 --> 00:29:08,000
from the development site or a sister plant.
358
00:29:08,000 --> 00:29:17,000
And then you maintain those twins going down in time, and they can identify when the equipment
359
00:29:17,000 --> 00:29:20,000
needs adjustment or the process is drifting.
360
00:29:20,000 --> 00:29:22,000
What do you think of that?
361
00:29:22,000 --> 00:29:24,000
I think it's absolutely fascinating.
362
00:29:24,000 --> 00:29:33,000
And I think the other thing that digital twins allow for is for operator training, right?
363
00:29:33,000 --> 00:29:38,000
Operators can practice and learn on a twin rather than the actual equipment or especially
364
00:29:38,000 --> 00:29:43,000
if product is a limiting factor.
365
00:29:43,000 --> 00:29:48,000
It's a great way to leverage digital twins, and I really found their presentation fascinating
366
00:29:48,000 --> 00:29:53,000
and interesting, and the future is coming.
367
00:29:53,000 --> 00:29:59,000
So where do you see industry and its adoption of AI in 5 to 10 years from now?
368
00:29:59,000 --> 00:30:07,000
So looking ahead for the next 5 to 10 years, this is what I personally see is coming.
369
00:30:07,000 --> 00:30:12,000
In my view, in the next 5 years or so, we'll see AI becoming more regularly embedded in
370
00:30:12,000 --> 00:30:15,000
our quality and manufacturing operations.
371
00:30:15,000 --> 00:30:18,000
We're already starting to see it now.
372
00:30:18,000 --> 00:30:23,000
The tools and approaches we're piloting today will become standard practice.
373
00:30:23,000 --> 00:30:28,000
Predictive quality intelligence, AI-assisted investigations, automated report writing,
374
00:30:28,000 --> 00:30:32,000
I think these won't be considered innovative anymore.
375
00:30:32,000 --> 00:30:38,000
I think in the next 5 years, they'll be basically standard use.
376
00:30:38,000 --> 00:30:42,000
And for companies maybe that aren't using them, they may start to feel a little behind
377
00:30:42,000 --> 00:30:47,000
if they haven't already started leveraging this.
378
00:30:47,000 --> 00:30:52,000
I also think that we'll see much tighter integration across systems.
379
00:30:52,000 --> 00:30:56,000
Right now, many of our AI applications are still somewhat siloed.
380
00:30:56,000 --> 00:31:02,000
My prediction is that in 5 years, we'll have comprehensive digital ecosystems where AI
381
00:31:02,000 --> 00:31:08,000
is connecting insights across quality, manufacturing, supply chain, and regulatory, creating a more
382
00:31:08,000 --> 00:31:13,000
holistic view of product quality and patient impact.
383
00:31:13,000 --> 00:31:20,000
Additionally, the regulatory landscape should mature within the next 5 years.
384
00:31:20,000 --> 00:31:25,000
Right now, we've seen a lot of drafts from health authorities, but I expect that within
385
00:31:25,000 --> 00:31:32,000
the next 5 years, health authorities will have those frameworks and expectations formalized
386
00:31:32,000 --> 00:31:40,000
and in place, and we'll have a much clearer view of the health authorities' expectations.
387
00:31:40,000 --> 00:31:46,000
Hopefully, in my opinion, those expectations will be harmonized across the health authorities,
388
00:31:46,000 --> 00:31:51,000
but that may take a little longer.
394
00:32:13,000 --> 00:32:18,000
Well, Miguelina, this has been really fascinating.
395
00:32:18,000 --> 00:32:25,000
I compliment Sanofi on being what I'll call a pathfinder, not a trendsetter, but certainly
396
00:32:25,000 --> 00:32:28,000
an innovator.
397
00:32:28,000 --> 00:32:36,000
We talked about your PLAI, your play application, and how it's being used to analyze deviations
398
00:32:36,000 --> 00:32:44,000
and streamline report writing, but also to pull together a lot more data and analyze
399
00:32:44,000 --> 00:32:49,000
it statistically than a human could.
400
00:32:49,000 --> 00:32:57,000
The regulators ought to be head over heels about this kind of thing, where more data,
401
00:32:57,000 --> 00:33:07,000
more science and risk-based uses of that data is being done through these statistical tools,
402
00:33:07,000 --> 00:33:09,000
which is really what they are.
403
00:33:09,000 --> 00:33:17,000
It's not so much an artificial intelligence, it's a statistical tool.
404
00:33:17,000 --> 00:33:25,000
I think this is really great, and I compliment Sanofi both on its innovation and how it organizes
405
00:33:25,000 --> 00:33:33,000
itself and starts up and deploys these tools throughout the organization.
406
00:33:33,000 --> 00:33:37,000
I'd like to thank Miguelina for her time with us today.
407
00:33:37,000 --> 00:33:42,000
It's great to hear about innovative efforts at a major pharmaceutical company to adopt
408
00:33:42,000 --> 00:33:48,000
new technologies that promote efficient manufacturing and quality improvement.
409
00:33:48,000 --> 00:33:54,000
That brings us to the end of another episode of the ISPE podcast, Shaping the Future of
410
00:33:54,000 --> 00:33:55,000
Pharma.
411
00:33:55,000 --> 00:34:01,000
Please be sure to subscribe so that you don't miss future conversations with the innovators,
412
00:34:01,000 --> 00:34:06,000
experts, and changemakers driving our industry forward.
413
00:34:06,000 --> 00:34:12,000
On behalf of all of us at ISPE, thank you for listening, and we'll see you next time
414
00:34:12,000 --> 00:34:18,000
as we continue to explore the ideas, trends, and people shaping the future of pharma.