Juniper’s Bob Friday: Preparing for AI in Your Business

0:01 [Music]

0:03 hi i'm james mcguire

0:05 and today we're talking about getting

0:07 your organization ready for artificial

0:09 intelligence

0:10 which of course includes making sure

0:12 your ai system is fed

0:13 enough data enough quality data to

0:16 discuss that i'm joined by a real expert

0:17 in the field

0:18 with me is bob friday cto of juniper's

0:21 ai driven enterprise business

0:24 bob very good to have you with us today

0:25 yeah i know james thank you for having

0:27 me here

0:27 you know ai's you know a topic dear to

0:29 my heart so it's really great to be here

0:31 with you today

0:32 great you know there's a there's a

0:34 really interesting uh report about

0:36 the state of artificial intelligence it

0:38 seems like there's a there's a bit of a

0:39 gap between

0:40 some of some of the hopes and some of

0:41 the optimism for ai and some of the

0:43 reality of how it's actually getting

0:44 deployed

0:46 especially a report i think it's either

0:48 sponsored or put together by juniper

0:50 uh it'd be nice for me to do a little

0:51 screen share and show the stats for

0:53 which i think it's

0:54 let me do this kind of advanced

0:56 technology the

0:57 zoom screen sharing uh here we go let's

1:00 do share

1:02 yeah and if i if i look at this it's

1:04 interesting so we look at the figures

1:07 the survey found that 95 of 700

1:10 respondents

1:10 believed that their organizations would

1:12 benefit from embedding ai into their

1:14 daily

1:14 operations so you know everyone really

1:16 wants to do it on the other hand

1:18 this compares with only six percent of

1:21 c-level leaders

1:22 who have reported adoption of ai power

1:24 solutions across their organization so

1:26 there's a real disconnect between uh

1:29 you know the optimism the hope for ai

1:31 and and what people want it to do or

1:33 think it can do

1:34 and the reality of how it's deployed

1:36 what what's going on there

1:38 to create that gap yeah you know james

1:40 when you look at it

1:41 you know i used to be the mobility cto

1:43 at the at cisco right you know

1:45 it would really start you know what i

1:47 heard there is really from customers

1:49 there's a couple things those customers

1:50 really were looking to do one is they

1:52 really wanted to get the

1:54 they wanted to basically get their

1:55 systems to stop crashing uh but they

1:57 really wanted

1:58 end-to-end visibility you know from the

2:00 client to the

2:01 to the cloud and that's really when ai

2:04 ops started to become more critical

2:05 and what i really noticed was you know

2:08 and that's one reason i

2:09 i left cisco to start mist right is

2:11 because what was really happening was to

2:13 build an ai

2:14 ops solution you really needed really to

2:16 have kind of a real-time pipeline

2:19 to deal with the data right it was

2:20 basically an architectural change what

2:23 was happening

2:24 you know from what i did 20 years ago at

2:26 airspace we built systems that were

2:27 helping enterprises manage

2:29 network elements you know really the

2:31 paradigm shift is really now helping

2:33 enterprises

2:34 really manage the end-to-end user

2:36 experience and that really requires you

2:38 to change two things

2:40 technically it requires something in the

2:42 cloud that can actually process this

2:44 data real time

2:45 right interesting the other interesting

2:47 twist i found in this

2:48 is really it requires almost an

2:50 organizational change

2:52 right culture a cultural change i'm

2:53 assuming a cultural change in both the

2:55 customer and in the vendors right so so

2:57 i missed

2:58 what we ended up doing was i ended up

3:00 really tying my

3:01 data science team up to my customer

3:04 support team

3:05 you know because one of the challenges

3:07 is you know you know we have data

3:08 scientists who

3:09 you know they understand the math and

3:11 the science behind the

3:12 ml algorithms but they're not

3:14 necessarily network domain experts

3:16 so you look at the challenges here you

3:18 know you got the vendors who've got to

3:20 change their culture

3:21 you know we got to move to architectures

3:23 that can actually do

3:24 aiml and you got to organizationally

3:27 change on the vendor side

3:28 and then on the customer side you know

3:30 the customers have to basically start

3:33 changing their it teams right you know

3:35 they're going through a bunch of changes

3:36 on

3:37 the it side you know you look at it

3:39 teams you know we were asking it teams

3:41 to do a

3:43 almost an exponential change right how

3:45 they operate their networks

3:47 yeah because because of the level of

3:49 data in other words well the complexity

3:51 right i mean you look at what's

3:52 happening on that site you know their

3:53 networks are getting more complex

3:55 and we're asking it guys to basically go

3:57 from being cli

3:59 programmers to being python programmers

4:01 right you know so they're already making

4:03 that shift to becoming

4:05 python programmers and now we're asking

4:07 them to become

4:08 ai data scientist experts right so

4:10 that's a lot of pressure on our

4:12 you know our customers right now and

4:13 getting their it team skill sets

4:15 up to the next you know next stage you

4:18 may you mentioned ai

4:19 ops so this is really is is that a key

4:21 area of yours in particular especially

4:22 in terms of you know dealing with the

4:24 complexity and

4:25 data and the data in the it operations

4:27 is are you are you big at ai ops in

4:29 particular

4:30 yeah that i mean that is what we do here

4:32 right now and that's really

4:33 that's really the change right ai ops is

4:35 really that paradigm shift from

4:36 managing network elements to managing

4:39 client to cloud right

4:40 you know i really don't care what's

4:42 between the client the cloud whether

4:43 it's

4:44 juniper equipment or someone else's

4:45 equipment my mission in life is really

4:47 to help those

4:48 help them figure out why their devices

4:50 and customers are having problems

4:51 internet poor internet connectivity

4:55 let's talk about the the ai uh ai ops

4:58 market for a second

4:59 how do you think is it is it close to

5:00 mainstream or is it still

5:02 in its infancy in terms of adoption what

5:04 would you say

5:05 you know i think right now we're really

5:07 in the early early days of aiops

5:09 adoption right

5:10 you know you know you look at enterprise

5:12 businesses you know like we mentioned

5:14 before

5:14 i mean culturally they're trying to get

5:16 their it teams you know with the right

5:17 skill sets to really embrace ai

5:20 and they're really really trying to

5:21 understand you know what is what's the

5:23 difference between ai

5:25 and automation right and from my

5:27 perspective

5:28 you know ai is really just the the next

5:30 step in the evolution

5:32 of automation right what we're really

5:34 doing

5:35 is we're really building things

5:37 solutions on par

5:39 they can do things on par with humans

5:41 now right it's kind of like a

5:42 self-driving car

5:43 diagnosing cancers you know and if you

5:45 look really one of the inspirations that

5:47 missed right you know one of the reasons

5:48 i started this was

5:50 you know when i saw watson play jeopardy

5:52 i don't know

5:53 right yeah it was it was a big deal when

5:56 yeah absolutely

5:57 you know that that's when i kind of

5:58 realized hey you know if they can build

6:00 something that can play jeopardy

6:02 right we should be able to build

6:04 something that can basically answer

6:06 questions

6:06 and manage networks on par with

6:10 network domain experts right like you

6:12 know the technology is real

6:14 we've gone from marketing hype to that

6:15 there is real technology here that we

6:17 can actually build something that can do

6:18 something on par with a human

6:21 well it's true but in terms of the the

6:22 you know the weaning of jeopardy that's

6:24 that's spitting back out information

6:28 uh which of course is going to be and

6:29 handling information processing

6:31 information which is going to be one of

6:32 the cores of aiops

6:34 but it's also i think about those i.t

6:36 experts

6:37 they're making nuanced decisions uh

6:40 they're they're

6:41 they're actually having to forecast

6:43 things and it says we need

6:45 we need aiops to do more than just spit

6:47 out the information we need we need

6:48 prediction and forecasting

6:49 we also we'll almost need the system to

6:51 be giving us a little bit of advice am i

6:53 am i correct for that

6:55 well yeah and that's what we're starting

6:56 to see right i mean we're starting to

6:58 see

6:58 systems now that can actually predict

7:02 hardware failures or find rma failures

7:04 inside the network right

7:05 you know so the days of you know where

7:07 you used to have to argue with your

7:09 vendor about you know whether or not

7:10 you're having a hardware software

7:11 problem

7:12 those days are going away because these

7:14 more intelligent systems they

7:16 know whether or not it's a hardware

7:17 problem software problem you know so

7:19 my transition from what i did 20 years

7:21 ago at airspace you know where

7:23 basically a customer would have to you

7:24 know send all bunch of information to

7:26 convince me it was a hardware problem

7:28 we are now we're sending rmas out right

7:30 you know customers don't have to wait

7:31 for me we can basically tell them yes we

7:33 know that you have a hardware problem

7:35 please send it back or other use cases

7:38 like bad cables

7:39 you know we've gotten down the point now

7:41 we can actually find these needle in the

7:42 haystack

7:43 bad cable problems where we're 90

7:47 or better accurate where i t departments

7:49 now i can say hey i trust you

7:50 you know don't don't bother me or if you

7:53 find a bad cable issues a support ticket

7:55 and and move on so those are kind of

7:57 cases where

7:58 we're starting to see ai and ml really

8:01 take on the this next level of

8:03 automation right

8:04 helping helping offload i t task onto

8:07 these ai solutions

8:09 so so what what is the name of the ai

8:10 ops solutions that actually can find a

8:13 bad cable

8:14 amid all those cables in in the data

8:16 center is there a specific solution that

8:17 does that

8:18 well i mean so so juniper we call it

8:20 marvis uh exactly

8:22 if you dig under the hood and actually

8:24 look at it you'll see something called

8:25 logistic regression

8:26 so that's kind of the actual ml

8:27 algorithm that's down

8:29 in the uh under the hood there is

8:31 actually making the determination

8:32 whether or not you have a bad cable or

8:33 not

8:35 interesting let's talk about uh the data

8:38 challenges what what data challenges

8:40 do organizations need to address in

8:43 order to quickly inform ai systems

8:45 because of course they need to be fed

8:46 the right the right data how do

8:49 companies do that

8:50 yeah and to be honest i mean like i said

8:52 back when we started mist right

8:54 i mean one of the questions when we

8:55 started this was you know should we

8:57 actually build an access point

8:59 or not right you know and when i made

9:01 the decision

9:02 back then was i wanted to build the

9:04 access point not because i thought the

9:05 market needed another

9:06 wireless access point it's because i

9:09 really want to make sure i can get the

9:10 data i need to answer a question

9:13 when i look at most enterprises right

9:15 now the mistake they're making is

9:18 they tend to focus all their energy on

9:20 trying to build big data lakes

9:22 data warehouses without really starting

9:24 with the question

9:25 right and so usually the advice i give

9:28 people is like hey if you're going to

9:29 work on an ai

9:30 ops ai type of problem you know you

9:33 really start with the question you're

9:34 trying to answer

9:35 you know in our case it was we're trying

9:37 to answer the question of why are you

9:38 having poor internet connectivity

9:41 and so that really starts with making

9:42 sure you get the data from the edge of

9:44 the network you know if you

9:45 start with that then it starts to drive

9:47 to make sure you get the data you need

9:49 to answer the question

9:50 you're trying to answer so the idea is

9:52 make make sure you know your question

9:54 you're sort of

9:55 guiding leading question as opposed to

9:56 merely buying technology

9:59 right or trying to pile data in a day

10:01 like right you know you see a lot of

10:03 enterprises so the first thing they try

10:04 to do is they try to take

10:05 every piece of data they have and start

10:06 you know you know sticking into splunk

10:09 or sticking to some big data warehouse

10:10 and then

10:11 then they find out they're not using 95

10:14 of the data never gets used for much of

10:16 anything right

10:18 it is amazing that the small amount of

10:21 data that is actually mined for

10:22 competitive advantage i mean that's i

10:23 think it's sort of the story of our of

10:25 our time is that

10:26 we have an ocean of data but we're using

10:28 just a small portion of it to really

10:30 navigate right and i'm going to add you

10:34 know if you look at the question right

10:36 you know as we've got you know what

10:38 we're doing now is we're starting to

10:39 expand

10:40 across different data sources right you

10:41 know when we started it was basically

10:43 getting data from access points to

10:46 answer the connectivity question

10:47 now we're getting data from the clients

10:49 you know their visibility of the network

10:51 we're getting data from the switches

10:52 and the routers you know and as you

10:54 start to get more data

10:56 it lets you answer questions more

10:57 questions like when we get data from the

10:59 sd-wan

11:00 router it starts to let us answer

11:01 questions about applications and

11:03 services

11:04 and we start to be able to answer

11:05 questions with more granularity

11:07 and then you can start doing proactively

11:09 fix it once you have data from the

11:10 switches

11:11 you know if we see a misconfigured vlan

11:14 we can actually practically correct that

11:15 without

11:16 bothering the i.t teams right right

11:20 you know we we talked about this a

11:21 little bit beforehand the idea of a

11:23 common telemetry

11:24 standard so what why does there need to

11:26 be a common telemetry

11:28 standard to make obtaining data for ai

11:29 easier why is that important in your

11:31 view

11:32 well i think it's important i think

11:33 that's one friction point we see in the

11:35 adoption of ai

11:37 is are we really going to get to a

11:38 common language where every vendor uses

11:41 the same you know

11:42 format for time um i think that's

11:45 probably a little hard

11:46 i think what is probably required in the

11:48 industry is if we get everyone to kind

11:49 of agree to how to store data away

11:52 you know when we look at you know what

11:53 question or what problem we're trying to

11:55 save

11:55 you know you may have to get data from

11:57 your access vendor right and you may

11:59 have to get data from your

12:00 lan router vendor and you may have to

12:02 get some data from your application

12:04 vendor

12:04 you're really to answer an end-to-end

12:06 type of question

12:08 so i think what's more important in the

12:09 industry right now is you know can we at

12:11 least get people to agree

12:12 on how they're going to store the data

12:14 way in these cloud solutions

12:16 because if you look where we're headed

12:17 right you know the industry is moving to

12:19 kind of

12:20 cloud first direction right we're almost

12:23 you know all the data you need is being

12:25 stored away you know

12:26 your different vendors are throwing this

12:28 data for you on your behalf

12:29 in different clouds so it's more

12:31 important about getting the data stored

12:32 away in some sort of descriptive format

12:34 that ai and bi can actually access and

12:37 understand how the data is being stored

12:40 is is it realistic to think that there

12:42 would be an industry standard in terms

12:43 of that or is it

12:44 is it inevitable or is it is it a long

12:46 shot in europe

12:48 i think it's an inevitable i think we're

12:49 in the early days of ai and so i think

12:51 people are just now fully

12:52 starting to appreciate the fact and i

12:54 think a lot of the aai

12:57 efforts are being kind of focused on

12:58 very narrow efforts where the data is

13:01 you know coming from one source or one

13:03 vendor you know so i think as you know

13:05 as ai expands especially iops and

13:07 network and expands

13:08 across networking we'll start to see the

13:10 industry move to more

13:12 industry standards like uh you know on

13:14 the switch side we moved to open config

13:16 right that was kind of a standard

13:18 effort to get the switches managed i

13:20 think so i do think start to see the

13:21 industry move to kind of more

13:23 standard ways of getting telemetry data

13:26 from different devices

13:27 especially on streaming right you know

13:29 pulling and pushing

13:30 you know streaming data yeah streaming

13:33 data you know you look at ai

13:35 uh you know in the past you know 20

13:37 years ago was all snmp right we were

13:39 basically pulling all these devices to

13:41 get stats

13:41 and information out of our devices as we

13:44 move into more of a real

13:45 time ai environment it's more about i

13:48 want things to be streamed to me in real

13:50 time i don't want to go pull things i

13:52 want to have telemetry come at me

13:54 so it's probably more important that we

13:55 get that standardized

13:57 as opposed to actually trying to

13:59 describe each little

14:00 field in the in the in the schema right

14:04 let's let's look to the future for a

14:05 second i i think specifically the future

14:07 not of overall ai necessarily but first

14:10 the future of ai apps in particular

14:13 you know when you look at the definition

14:14 of ai apps it's artificial intelligence

14:16 for it operations and that's really the

14:18 commonly accepted one i think gartner

14:19 made

14:20 that up in 2017. um

14:23 but i think about this obviously going

14:25 to be an ai platform that runs

14:28 the entire enterprise from sales to it

14:32 to you know hr and it's all going to be

14:34 one platform so really the idea of ai

14:36 ops which

14:37 you know our ai for it operations

14:40 is going to be consumed into a larger ai

14:44 framework true false or what's your view

14:47 on that

14:48 well i mean i think there's a lot of

14:49 concepts here you know one of them is

14:51 this ai for me personally like i said ai

14:54 is really about

14:55 trying to do something on par with a

14:56 human you know whether it's

14:58 driving cars or detecting cancer and

15:00 medically

15:01 for networking it's more about can we

15:03 build something that does something on

15:04 par with network domain experts

15:06 you know and i think the future is you

15:08 know and my common things i think

15:10 you and i probably grew up i don't know

15:11 if you remember watching star trek when

15:13 you were a kid right i do

15:14 yep definitely you know in every almost

15:16 every technology that was in star trek

15:19 30 40 years ago has now separated

15:22 except for the transporter right the

15:23 transporters i'm not sure we're gonna

15:25 live to see that

15:26 yeah computer remember the computers

15:28 like computer tell me

15:29 oh definitely yeah no it's and it's here

15:31 yeah yeah so i think we are gonna see

15:33 that in our future where

15:35 you know network it teams are gonna have

15:37 an ai assistant on their team

15:40 right that's going to make it easier for

15:42 them to diagnose

15:43 and run and maintain these networks just

15:46 like we saw on star trek right

15:47 and this is what i call the transition

15:49 you know to conversational interfaces

15:52 you know you can you kind of remember

15:54 the industry you know

15:55 when you're younger it's all about cli's

15:57 right we're basically programming and

15:58 maintaining all these

15:59 switches and routers you know i mean cli

16:02 jockeys you know

16:03 and then we kind of move to the

16:05 dashboard the gui dashboards for running

16:07 and manage our networks

16:08 right i think we're now in that

16:09 transition to conversational interfaces

16:12 are just a much easier way for network i

16:14 t

16:15 teams to actually manage and extract

16:17 data

16:18 and visibility on other networks so i

16:20 think that's kind of the future of

16:22 what i call ai in networking is the

16:24 virtual system think star track then

16:25 computer

16:26 you know you're going to start to see

16:28 virtual systems become a key element

16:30 in most of these network i.t teams in

16:31 the future so that means that

16:34 that nlp will be really at the center of

16:37 it all so that the i.t

16:38 expert can speak to his or her machine

16:40 yeah nlp is a great example right now of

16:42 these conversations and that that's an

16:44 area that's changing as we speak

16:46 right now i mean i can tell you there

16:48 are things that we can do now with the

16:49 conversational interfaces that we

16:50 couldn't do two years ago right

16:52 you know so that technology is evolving

16:54 you know it's still evolving right now

16:56 right so so what about the future of ai

17:00 ops

17:00 specifically if you look at this you

17:03 know

17:04 they're expecting something like 15 to

17:05 20 percent annual growth in in revenues

17:08 from about 17 billion now at about 40

17:10 billion in the year 2026

17:12 based on i think at idc numbers what do

17:15 you see

17:16 we the future of ai up specifically

17:20 going forward

17:21 yeah so specifically you know in in the

17:23 short term i think right now we're gonna

17:25 see you know as i said data you know

17:29 data is gonna be the i think it's gonna

17:30 happen in phases and data is probably

17:32 the first phase

17:34 you know where enterprises are working

17:35 on making sure they can get the data

17:37 that they need to solve some aips

17:39 problem right

17:40 and as we success earlier right so you

17:43 know i make a barrel of wine every year

17:45 and you know names red wine or white

17:47 wine red wine

17:48 yeah okay yeah so but uh you know same

17:52 saying right

17:53 great wine starts with great grapes

17:54 great the great data

17:56 and so that's probably the first phase

17:58 of the journey that most people are

17:59 going to get on is get that down

18:01 and then you gotta really get that data

18:02 into a framework that you can actually

18:04 start deploying

18:05 ml to and that's kind of real-time

18:07 pipeline

18:08 you know so that's kind of the future i

18:10 see the journey

18:12 uh onto the adoption of ai ops into

18:14 ultimately

18:15 you know before we before i retire i

18:18 suspect we're going to start to see

18:20 these virtual ai assistants become a

18:22 mainstream of uh

18:23 the i.t teams of the future yeah it

18:26 doesn't seem like that

18:27 particular one is too far away actually

18:29 that could be in the next couple two

18:31 three years i would think

18:32 yeah before before i retire right

18:35 okay good uh bob i think you've said it

18:38 there's a lot more to say but i think

18:39 that's

18:39 that's we're good for today i greatly

18:42 appreciate you sharing expertise with us

18:44 thank you very much yeah great to be

18:46 here james