0:07

Welcome, everyone, to the AI

0:09

and Business Podcast. I'm Matthew

0:11

D'Amelio, Senior Editor here at

0:13

Emerge Technology Research. Today's

0:16

guest is Lax Pernankill, a

0:18

Principal in Life Sciences Consulting

0:20

at Deloitte. Lax joins

0:22

Emerge CEO and Head of

0:24

Research Daniel Fajela on today's

0:26

show to discuss AI in Pharmaceutical

0:28

Supply Chain Manufacturing, showcasing a

0:30

host of AI use cases

0:32

and applications across manufacturing and operational

0:35

lines of business. Later,

0:37

the two surmise the near future

0:39

of AI adoption in Pharmaceutical Supply

0:41

Chain Management and where business leaders

0:44

should focus their investments and attention.

0:47

Today's episode is sponsored by Deloitte,

0:49

and without further ado, here's their

0:51

conversation. So

0:58

Lax, welcome to the program. Thank you. Yeah, glad to

1:00

be able to dive in with you here today. We're

1:03

talking about a part of the life sciences world that you

1:05

are awfully up close and personal with,

1:07

which is supply chain manufacturing. I mentioned one

1:09

of our previous intros for the series. So

1:12

much of the AI focus in this area

1:14

is like clinical trials, drug development, but there's

1:16

all these other big operations in life sciences.

1:18

So because some of it will even be

1:20

new for our audience, I'd love to kick

1:22

off with what you're seeing boots on the

1:24

ground in terms of trends

1:27

and challenges within supply chain manufacturing that

1:29

that's making AI relevant making data relevant

1:31

today. Thanks. Well, first of all, thank

1:33

you for having me on. And of

1:35

course, I'll say this. So in my

1:38

two decades of being in

1:40

technical operations, manufacturing supply chain

1:42

quality and other aspects of

1:45

pharma and tech operations, I've

1:47

never seen a time where my

1:49

clients have been so surprised with

1:52

the pace at which AI

1:54

and data has taken over the

1:57

conversation and the focus.

2:00

of organizations, almost every single client

2:02

of ours is very indexed on

2:04

trying to think through what

2:06

does AI mean for our business, particularly

2:08

in operations. And so operations is not

2:10

far behind at all in that journey.

2:12

I'll say a couple of things. So

2:15

first, from a trend perspective, operations

2:17

has always been a game of

2:19

using data and facts in making

2:22

decisions, whether it's in

2:24

supply chains, whether it's in manufacturing,

2:26

whether it's in quality, because these

2:28

are highly engineered products and processes,

2:31

there has always been a need

2:33

for using that data to make

2:35

those decisions. What the new emerging

2:37

trends around AI has done is

2:40

put a shining light on data

2:42

that's not used. My thesis

2:44

advisor or my PhD would

2:46

often say the best

2:48

way to produce any variance

2:51

or any error in a process is

2:53

not to measure it. So in the

2:55

industry, it's always been the norm that

2:57

in the past and a long time before,

3:00

you don't measure if you can't explain why

3:02

the data is the way that it is.

3:04

But the industry has now come around that,

3:06

the regulators have come around that and there

3:08

is this mega trend in the industry in

3:10

operations on doing more with

3:13

data. The other tidbit I have, which

3:15

is along with this trend that you

3:17

see is that a typical manufacturing floor

3:20

creates about a terabyte of data every day. And

3:23

barely 5% of that is used

3:25

in operations. And so pharmaceutical

3:28

operations and our clients have

3:30

out changed that. The second trend you will

3:32

see is that there's a resurgence of sort

3:36

of industry product technologies in

3:39

pharmaceutical operations. And this emergence

3:41

of these technologies are creating

3:44

new sources of streaming data right from

3:46

the production floor, right from the supply

3:48

chains, where data is actively coming in,

3:51

whether it's in cell engine therapies, the

3:53

data of where your particular therapy is

3:55

in the out heading to the treatment

3:57

center or in a production.

4:00

where the manufacturing yield of the manufacturing process

4:02

is why available on our dashboard that we

4:04

need to see and make sure that we're

4:06

taking the process to get the most out

4:08

of the manufacturing process. The data

4:11

streaming coming from the production floor has

4:14

exponentially increased. That increased because of

4:16

the new and emerging technologies that

4:19

are being deployed in manufacturing and

4:21

supply chain. The third

4:23

thing is regulators have become

4:25

more and more attuned

4:27

to leveraging these novel

4:30

technologies, including AI. So

4:32

all of these three trends, the

4:34

regulators, availability of data, the availability

4:36

of new technologies to capture the

4:39

data and do something with it,

4:41

has created a substrate for pharmaceutical

4:43

companies to bring AI to life.

4:45

And bring AI not only for

4:48

a rear view, backward looking analytics,

4:50

but a forward looking, take

4:53

active decisions as the process is happening,

4:55

as the supply chain is executing to

4:57

delivering products to patients. So hopefully that

4:59

gives you a sense for what's happening

5:01

in the market. It does,

5:03

and I'd like to actually paint the picture. You're talking

5:05

about a terabyte of data a day, which is a

5:07

pretty overwhelming amount. I think I can

5:09

only imagine the number of pieces of equipment.

5:12

I can imagine in certain applications there's more computer

5:14

vision being used, whether it's detecting errors in the

5:16

machines themselves or the product or what have you,

5:18

and then all kinds of other movements and sensors,

5:21

et cetera. Give us an idea of what the

5:23

pie chart or some of the most important big

5:25

streams of data are and which of those are

5:27

really new? Which of those are growing? Because clearly

5:29

the expansion of this data is part of what

5:32

we need to wrangle to get value, and we're

5:34

gonna get into that. But talk about

5:36

what makes up that big pie of

5:38

a terabyte here. Yeah, yeah, that's great.

5:40

That's a great question. So generally speaking,

5:43

there's a lot of data that gets

5:45

generated in what are called L0, L1

5:47

systems. So the typical manufacturing architecture has

5:50

got from level zero to level five,

5:52

at level five, enterprise planning systems and

5:55

level zero are sensors on the production

5:57

floor. And about the L0, L1 layer. is

6:00

where 70 to 80% of

6:02

the data gets generated. These are

6:04

pressure data, sensor data, recipe data,

6:07

equipment data, temperature of the environment, samples

6:09

that people take. There's a lot of

6:12

data that gets generated at that L0,

6:14

L1 level. And then

6:16

the rest of 20% is split between metadata associated

6:19

with all the other systems that are

6:21

trying to control the manufacturing process and

6:24

the ensuing business processes. So

6:26

that's typically what we see is there's a lot of that

6:28

data on the L0, L1 layer closest

6:30

to the production set is not used

6:32

as much because these are high-volume data

6:34

sets and also these are regulated manufacturing

6:36

processes that we don't tend to change

6:38

as much. What we are seeing is

6:40

a credible change in how that's

6:43

being transformed. There's an example of a

6:45

client of ours built in a manufacturing

6:47

AI tool that on the production floor,

6:49

as things change, they can actually immediately

6:52

take an action and what

6:54

is the next best action they can take to bring that

6:56

almost like a self-driving car, bring the car

6:59

back between the two lanes and deliver the

7:01

product at the end of the manufacturing process.

7:03

I would say a lot of it is

7:05

on the lowest level of data generation. Got

7:08

it. And clearly leveraging that data

7:10

is part of how we bring to

7:12

bear the value of artificial intelligence. There's

7:15

probably innumerable applications beyond

7:17

even the vision

7:20

in manufacturing and other things that I might be

7:22

thinking about. You're deep in this game. You're looking

7:24

at which of these are adding

7:26

value today. You're looking at which of these

7:29

applications of AI could make

7:31

the biggest impact in the next couple of

7:33

years in terms of really driving efficiencies and

7:35

effectiveness within operations. What are some use cases

7:37

in this domain that for you are

7:39

really worth talking about that leaders should understand and that

7:41

you can help explain? Yeah, so

7:44

I'll break them into kind of

7:46

broadly with the emergence of

7:48

generative AI. I'll talk specifically about generative

7:50

AI and where generative AI applications are

7:52

in pharmaceuticals and that type. But let

7:54

me start with what was the journey

7:57

in the last half decade or so

7:59

around bringing. large machine learning and AI

8:01

into the production floor as a starting point

8:03

for the use cases and then we'll step

8:05

over to the narrative AI. On the machine

8:07

learning model, the

8:10

use case and applications spread

8:12

between trying to get control

8:14

of your manufacturing process, trying to

8:17

get control of inventory, trying to

8:19

get control of distribution

8:21

and logistics of your products

8:23

in the marketplace. So each

8:25

of those domains have

8:27

a significant set of applications of using

8:29

AI. As an example, we talked

8:32

at length about manufacturing process data

8:35

and then controlling the process. Another

8:37

example on the production floor is

8:39

almost always inadvertently there are deviations

8:41

that happen or nonconformances

8:43

that happen in manufacturing. Most

8:46

of them, 80 to 90% of them are

8:48

benign nonconformances. Somebody forgot to

8:50

put a signature, somebody forgot

8:53

to change their temperature from

8:55

35 to 37. These

8:58

really don't have an impact on the

9:00

product because the process is usually robust.

9:02

But a lot of these deviations are

9:04

written up because of the regulated process,

9:06

then somebody has to adjudicate whether this

9:08

is a risk to the product and

9:10

to the patient. So using

9:12

AI to very quickly scan

9:14

these deviations and adjudicating what

9:16

and then only giving that

9:19

5% of significant critical deviations

9:21

to a human to look at is a

9:23

talk you say that almost 89%

9:26

of my clients are trying to deploy

9:28

right now actively in the marketplace. And

9:30

these are all just legacy, I'll call

9:32

it legacy, but machine learning type AI.

9:35

In the generative AI space, which

9:38

is picking up significantly right now,

9:40

there are assistive applications that clients

9:42

are thinking about and

9:45

generative applications. So a typical pharmaceutical

9:47

company will need to submit what's

9:49

called annual report on that product

9:51

that kind of looks back at the

9:53

last year and say here are all the changes

9:55

that happened in the manufacturing process, the product and

9:57

so on and so forth and they compile a

9:59

lot. of data and create a report out of

10:01

it. Pull your use case for

10:04

gender to AI that can pick up data

10:06

sets and create a standard report that

10:08

then somebody, a human, can go and edit

10:10

in mnds. That's a use case that a

10:12

lot of times thinking about a

10:14

typical pharmacist comes

10:16

to the point where you have a million documents

10:19

that are SOPs and manuals and all

10:21

of these things that people need to

10:24

understand. A large language model can read

10:26

all those documents and become an SOP

10:28

assistant for manufacturing and operations in operators

10:31

on the production floor. There

10:33

are many of these gender to AI use

10:35

cases that are coming to bear. Supplier management

10:38

is another gender to AI use case where

10:40

suppliers being able to track and see which

10:43

suppliers are performing well and then creating and

10:45

then communicating with those suppliers. If you're seeing

10:47

these trends, what are you going to do

10:49

in terms of adjusting your performance with another

10:51

use case? There are a plethora of use

10:54

cases. In fact, we put out

10:56

a few hardware on what those use cases

10:58

are in the public domain as well. Yeah,

11:01

I'm sure that some of that's going to be tied into

11:03

our broader show notes and the bigger sort of picture we're

11:05

painting here in Lifeside with all of your experts. I

11:08

want to dive into a few of these that you mentioned

11:10

if that's all right just to paint a mental picture for

11:12

the listeners who are sort of tuned in and sort of

11:14

saying, okay man, what does this look like in real life?

11:16

So one example you talked

11:18

about was in the manufacturing process, we've got

11:20

temperatures and pressures and combinations of chemicals and

11:22

a thousand things that happen, all of which

11:25

need to be measured and managed because we

11:27

cannot have a negative impact on patients. And

11:30

we can have AI sort of sipped through those

11:32

variants. We've got a lot of streams of data

11:34

here. I imagine some of this stuff, I don't

11:36

know if all of this is coming from amazingly

11:38

fancy new machinery. I imagine some of it is coming out

11:40

in a pretty, you know, gunky old

11:42

school format that somebody has to export

11:45

and harmonize and then load into something. Yeah,

11:47

it's a great point. I have no idea

11:49

what this looks like in practice, but it

11:51

seems like kluge is the word

11:53

that comes to mind. I mean, just from

11:55

what I know about manufacturing outside of Lifeside,

11:57

kluge is pretty much the modus operandi. Life's

12:00

eye might be a little bit sharper, but you know, individual

12:02

piece of equipment that are exactly the same, just

12:06

temperature settings need to be different for this one than this

12:08

one. Stuff is crazy in the

12:10

world of physical stuff. Let me ask this,

12:12

you know, when it comes to finding that

12:14

variance, my guess is what has to happen

12:16

is we need to have enough history of

12:18

what that variance is and enough humans really

12:21

identify, hey, when these temperatures at

12:23

this point or these pressures at this

12:25

point or this combination that happens here

12:27

doesn't have these things involved and

12:29

this other thing happens later in the process, then

12:32

this has actually become a problem.

12:34

We need to figure out when these

12:36

combinations overlap enough where it would be a

12:38

challenge and that's got to be our training

12:41

information for us to be able to, let's

12:43

say to your point, we can flag the

12:46

green, yellow, red, right? You've got 5% red, we've

12:48

got 10% yellow, we've got the rest of it

12:50

is pretty clean even if there's a little bit

12:52

of variance. Is this the right way to understand

12:54

it? I want the leaders at home to kind

12:56

of think about what it might look like to

12:58

train such a system. That's great. That's

13:00

a great question. So that's actually two parts

13:02

to the question you raised. One is what

13:04

I call readiness of the data and the

13:07

other is the completeness of the data to

13:09

train and to use it in AI. The

13:12

readiness of the data in of itself is

13:14

a massive, massive undertaking. A lot of the

13:16

times what we see is that our clients

13:18

have invested in systems and data and

13:21

I'm not necessarily, for

13:23

good reasons, not necessarily spent the time

13:26

to kind of clean, have master data

13:28

that's clean that can actually be used

13:30

in a usable format to use

13:33

your words of not be clued as you

13:35

use it in your decision

13:37

making. So now there are data

13:39

sets that are regulated that they need

13:42

to be clean so they have invested

13:44

those but those that are not

13:46

required, why bother investing in cleaning those?

13:48

So a lot of the times many

13:51

of these programs either win

13:53

or lose on the back of

13:55

getting that front end cleaned up

13:57

really well and there are some.

14:00

very cool examples of where we went

14:02

in with some unique ways of cleaning

14:04

that with very little human intervention using

14:07

graph methodologies and so on and so forth

14:09

that kind of drives that cleaning very quickly.

14:11

So that's one part of the mountain, you

14:13

know, digging out of the mountain. The

14:16

second part of it is completeness of

14:18

data. And that's where you

14:20

kind of mentioned, you know, do you have

14:22

enough radiation in your training data for the

14:24

AI models to detect when is it good,

14:27

when is it bad. And this is where

14:29

it's an art as much as it is

14:31

a science because when you train models,

14:33

you will want to make sure that

14:36

the data that you have is an

14:38

accurate representation of the reality that there

14:40

is. But also this is why we

14:42

look at AI as a human and

14:44

machine interface with an assistance in the

14:46

middle, because many

14:49

manufacturing processes change over time.

14:52

So every time a manufacturing process and

14:54

operations and supply chains change, you're

14:57

now in a new regime that the

14:59

old AI model will need to continue

15:01

to learn. So how do we build

15:03

training data sets and

15:05

train AI models to

15:07

recognize these regime changes

15:10

is another interesting topic that we've come up

15:12

with some unique ways of solving.

15:15

And we've used new anomaly

15:17

detection algorithms, we've used trend detection algorithms

15:19

to kind of solve for that and

15:22

which is really cool. So yeah,

15:24

these two are probably the biggest

15:26

challenges in seeing the value from AI. Yeah,

15:29

and I want to get a little bit into

15:31

that. I mean, setting the data table and making

15:34

data infrastructure come to life, you

15:36

know, not an easy game. And then, you know,

15:38

being in this space for the last 10 years,

15:40

there was a long time in AI, which I'm

15:42

sure, you know, you recall well, where

15:45

we're really talking about kind of bandaid surface

15:47

level stuff, if it got beyond a POC,

15:49

it was kind of hanging out in one

15:52

specific corner, one specific workflow. We're now seeing

15:54

enough AI fluency from many

15:56

good experiments, many failed experiments and enough

15:58

excitement around the world. around really being

16:00

able to see how powerful this next wave

16:03

of AI is for people to be open

16:05

to the data infrastructure question. What

16:07

do you think, because this is gonna apply to the

16:09

next use case I'll talk about on the supplier side,

16:11

but what do you think leaders are gonna have to

16:13

understand about making those kind of

16:16

undergirded investments? Because there's gonna be some systems

16:18

that cannot stay the way they are if

16:20

we wanna do forecasting, if we wanna make

16:22

smart decisions, if we wanna be compliant. How

16:24

do some people think about that investment? Because

16:26

you're articulating some really important ideas and probably

16:28

haven't reached every leader who needs to hear them.

16:31

So there's one part of that

16:33

which basically says the more AI

16:36

is at the fingertips of the

16:38

frontline colleagues that

16:40

are manufacturing, that are supplying, that

16:43

are performing logistics activities, the

16:45

more it becomes helping those

16:48

individuals in the front lines, the

16:50

more value that AI creates. A

16:53

lot of the times AI helps in correcting

16:55

for what might be a wrong

16:58

step that a particular operator might

17:00

be taking, or an adjustment or

17:02

immediate turn to what's called a

17:04

golden path or operations, trying to

17:07

get back to the golden path.

17:09

So pushing the AI decisions, AI

17:11

adoption down to the front lines

17:14

is one way to scale this and make this

17:17

beyond the one POC or the two

17:19

POCs and trying to deploy AI at

17:21

the front lines. The second part of

17:23

that is we

17:26

always ask our clients to think

17:28

about value as a starting

17:30

point. So the applications or the

17:32

use cases you're reference, if there is no

17:35

value to be had from those use cases,

17:37

don't even go down that path even if

17:39

it's cool, right? And value is measured in

17:42

only a handsome of ways, whether

17:44

you're driving up more capacity, more

17:46

revenue, more product being

17:48

made, you're driving cost down, you're

17:51

turning your assets faster, your

17:53

asset efficiencies higher, or

17:55

you're distinctly driving some

17:58

other non-tangible, sometimes tangible. and

18:00

it will benefit around sustainability or quality or efficacy

18:02

or whatever it is. So it needs to tie

18:04

back to some things that the

18:06

operations leadership team cares

18:08

about as part of their objectives and goals.

18:10

So starting with that value and then driving

18:12

down to the use cases is another way

18:15

to ensure that you're delivering value back to

18:17

the patient. Absolutely, I mean, beginning with the

18:19

end in mind that there's a

18:21

fear and FOMO are very bad guides for where

18:23

to apply emerging technology, but value is a really

18:25

good place to start. So hard to disagree with

18:27

you there. When you mentioned supply chain, just to

18:30

touch on this before a little bit of parting

18:32

advice for the leaders here, another great use case.

18:34

I can imagine, you know, you're talking about supplier

18:36

performance. Talk a little bit about,

18:38

you know, because there's a lot of players, I mean,

18:40

there's people that are supplying raw

18:42

chemicals here. There are people that are handling part

18:45

of the manufacturing process and giving us goods that

18:47

maybe aren't finished yet or whatnot. And there's complexity

18:49

here that some of the listeners and including myself

18:51

might not be familiar with. What are we measuring

18:53

for those folks where AI can

18:55

help us really keep expectations in

18:57

line? So the supplier of score carding

19:00

and using AI to perform supplier score

19:02

carding, supplier performance management is another emerging

19:04

use case. To

19:06

give you a little bit of flavor for this, there

19:09

are 170 points that typical FOMO

19:11

companies use, data points, if you

19:13

will, for evaluating how a supplier

19:15

is doing. And a lot of the times, 89% of

19:17

that is manual, sort

19:21

of hard charged, sifting through a lot

19:23

of, you know, past performance, being all,

19:25

you know, documents, you know, all of

19:27

that stuff, come at those data points.

19:30

And being able to automate and have,

19:32

you know, large language models

19:35

kind of sit on top of all

19:37

of these data sets, whether it's certificates

19:39

of analysis, supplier purchase orders, invoices, whatever

19:41

it is, and supply communications, and being

19:44

able to kind of come up with,

19:46

how is the supplier performing? Have they

19:48

met these six or seven distinguished criteria,

19:52

qualitative and quantitative in

19:54

terms of performance is going to be the

19:56

art of the future. This supplier performance also

19:58

extends to, you know, We did an

20:01

example with a client where we

20:03

applied AI on all of their

20:05

certificates of analysis, COFAs as they're called, for

20:08

all of their chemicals that they get. And

20:10

when we ran that, we could

20:12

easily find issues where

20:15

some COFAs was approved, but actually should

20:17

not have been approved because

20:19

that was out of spec as it

20:21

was received, as received from the vendor. So

20:23

some of these things can be caught using

20:25

AI, and that helps you grow and have

20:27

meaningful conversations with suppliers to make sure that

20:29

your supply is not only the supply that

20:31

you paid for, the product that you paid

20:34

for, but also it doesn't have a downstream

20:36

impact on products that you are making using

20:38

those raw materials. The second thing

20:40

I'd say is there's an amazing amount

20:42

of information in the public

20:44

domain on like what is happening in

20:46

the marketplace for a variety of these supplied

20:48

products, right? And so being able to mine

20:51

that information from the public domain and then

20:53

generate insights is another sort of supplier

20:56

performance and insights

20:58

are used that are kind of

21:00

thinking about as well. Got it.

21:02

Is there information about pricing, timing,

21:05

whatever for what's happening with... Or even something

21:07

has happened, like for example, the earthquake that

21:09

happened in Japan, what is the issue of

21:12

all the supply issues that might come down

21:14

the pipe to me maybe six months down

21:16

the line because of this earthquake

21:18

that shut down a supplier that was a third

21:20

tier supplier for my... I think tier

21:22

supplier for my first year. Being

21:25

able to sort of field

21:27

and forecast and search at scale the external data

21:29

that could help us know maybe we want to

21:31

steer clear from these two suppliers or maybe we

21:33

still think they're safe, something like that. Exactly. We've

21:35

had on folks like Signal and Meltwater and some

21:37

of the players that do like this broader media

21:39

monitoring, but you're obviously talking about in a very

21:42

narrow context of how is this going to affect

21:44

my supply chain, which I think is useful for

21:46

the audience to know. I know we're going to

21:48

be coming up on time in a second here,

21:50

Lex, but I want to get your

21:52

vantage point. You're seeing a lot of people now almost

21:54

certainly, at least in

21:56

a significant way, start their AI journey. Certainly

21:58

their Gen AI journey. journey in the

22:01

supply chain manufacturing portion of life sciences.

22:04

What sort of parting advice do you have for leaders that are

22:06

looking at the different use cases? They maybe

22:08

have a good sense of where the value is

22:10

in their business, but they're not exactly sure where

22:12

AI would fit in. Where should they begin their

22:15

thought process of making this a really high ROI

22:17

endeavor into a new technology? Yeah,

22:19

I think some of the most

22:22

common starting points that I have

22:24

seen in the last couple of

22:26

years in both AI, broadly engine

22:29

AI specifically, have been in manual

22:32

data intensive parts of the

22:34

value chain. Whether it is,

22:37

as I mentioned before, mining

22:39

for information on complaints, mining

22:41

for information on non-contamancers to

22:43

reduce the human effort involved

22:46

in removing them, or

22:48

getting more, squeezing more juice from

22:50

their manufacturing assets by using the data

22:52

that is coming from the manufacturing assets

22:54

to make the decisions. Those

22:57

are two common standard starting points for

22:59

many of our pharmaceutical clients. On

23:01

the supply chain side, the other common

23:04

starting point is inventory, deploying AI on

23:06

managing inventory in their network, because there

23:08

are so many moves that happen in

23:10

the manufacturing network, as well as in

23:12

the outbound distribution network. Optimizing

23:15

for that inventory goes a long way because these

23:17

are really, really, really expensive products. Those

23:20

three are the most common starting points

23:22

that net a lot of return for

23:24

our clients, because either direct to bottom

23:27

line, direct to top line impacts of

23:29

the executed interest. But then there are,

23:32

we've compiled 45 plus 50, nearly 50

23:34

use cases where they can go and

23:38

deploy these AI use cases. Each

23:41

of them have returned tight with

23:43

them, the scale will differ between

23:45

clients. But those three would be my starting point for a

23:47

lot of my clients. That's cool. So

23:50

in terms of early conversations, early investigations of the quality

23:52

of our data, the potential impact of these technologies, we've

23:54

now got some low hanging fruit, hopefully for the listeners

23:56

tuned in around where they might be able to find

23:58

them. So I appreciate you being as practical as you've

24:00

been and I had a lot of fun learning with

24:02

you here at LAP So thanks so much for being

24:05

with me Before

24:14

we draw a close to today's

24:17

episode some highlighted points I think

24:19

will go a long way for

24:21

our listeners, especially where they came

24:23

from our guest First

24:25

AI is helping shed light

24:27

on untapped data sources in

24:29

pharmaceutical manufacturing workflows and beyond

24:32

Operations are now prioritizing the use

24:34

of data to make decisions rather

24:37

than relying solely on experience and

24:39

intuition as regulators and industry

24:41

leaders encourage more data driven decision

24:43

making LACS

24:45

emphasizes that data needs

24:47

to be ready for training a

24:50

system to flag potential issues They're

24:52

in including identifying when combinations of

24:54

variables overlap and become problematic LACS

24:57

also highlights the importance of clean

24:59

and complete Data

25:02

for AI models to detect

25:04

variations in manufacturing processes operations

25:06

and supply chains Leaders

25:08

must understand the importance of data

25:11

infrastructure to support AI adoption and

25:13

make them and make informed

25:15

decisions Leaders must understand

25:17

the importance of data infrastructure to

25:20

support AI adoption and make informed

25:22

decisions Deploying AI at the frontlines

25:24

and starting with value are key

25:27

to scaling AI and delivering tangible

25:29

benefits to operations leadership teams AI

25:32

can help automate supplier performance management

25:35

by analyzing large data sets and

25:37

identifying issues enabling meaningful

25:40

conversations with suppliers And

25:43

on behalf of Daniel Fagella our CEO

25:45

and head of research in the entire

25:47

team here at Emerge Technology Research Thanks

25:50

so much for joining us today, and we'll catch you

25:52

next time on the AI in fitness podcast Thank

26:00

you.