0:05

Welcome back to the Sales for

0:07

Admin's podcast. Today, we're catching up with

0:09

Momet O-run, longtime friend of the pod

0:11

and true expert in data and AI.

0:14

I'm going to tell you, a lot

0:16

has changed in the world of artificial

0:18

intelligence since our last chat. Momet's here

0:20

to break it down from

0:22

hallucination risks to the role

0:25

of data cloud in creating

0:27

trustworthy AI experiences. If you've ever

0:29

been wondering how to make your

0:31

data more meaningful and your AI

0:33

outputs more reliable, well. You are

0:35

in for a treat. So make

0:37

sure to follow the podcast so

0:39

you don't miss a single episode.

0:41

And with that, let's get Mamet

0:43

back on the podcast. So

0:46

Mamet, welcome back to the podcast.

0:48

It's wonderful to me, Mike,

0:50

Mike. I know. You just come

0:52

by with all these wisdoms and

0:54

knowledge that you have in the

0:56

world. Last time we were on,

0:58

and I'll link to that show,

1:00

we were talking about hallucination risks.

1:02

And it's been a year, and

1:04

boy, I tell you, a year

1:06

in AI time, everything's changed. So

1:08

what's new in your world? What

1:11

are you paying attention to in

1:13

terms of AI and agent force?

1:15

To be honest, one of

1:17

the interesting things about having

1:20

been around the while is while

1:22

the technologies are new,

1:24

our overall objective haven't really

1:27

changed. And one of the

1:29

things I've been really trying to...

1:31

look back to us what

1:33

were past challenges we overcame,

1:35

what were the parallels, and

1:38

what were some of the

1:40

best practices that people newer

1:42

to the field around day-day

1:44

integration, artificial

1:46

intelligence may not

1:48

know about, though we can

1:50

share this knowledge while absolutely

1:52

picking up new ways of doing

1:55

things also, because we

1:57

definitely have new tools under our

1:59

belt. a organized way to assess

2:01

what may cause hallucination risk

2:03

and mitigating it has been

2:05

a truly hot topic. I

2:07

have been visiting old friends,

2:09

making new friends as I've

2:11

been traveling across different sports

2:14

events as well. And the

2:16

good news is people are

2:18

excited about the potential. People

2:20

are also excited about having

2:22

tangible methods. They can take

2:24

back to their organization. I'm

2:26

looking forward to sharing some

2:28

of these with you today.

2:30

Yeah, I mean, I get

2:32

a whole lot of... It's

2:34

interesting, you look at some

2:36

of the stuff that's out

2:38

in the world, and the

2:40

spectrum for people looking at

2:42

what's going on with AI

2:44

goes all the way from

2:46

everything that it says is

2:48

right to nothing that it

2:50

says it's right, and somebody

2:52

falls somewhere in between there.

2:54

But I feel like, you

2:57

know, I did podcasts in

2:59

2024, early 24, I think

3:01

in 23, even talked about

3:03

hallucination. The one thing that

3:05

it kind of came back

3:07

to, it seems to have

3:09

gone away, because I think

3:11

more of the conversation is

3:13

around the quality of the

3:15

data and what we're feeding

3:17

agent force and getting your

3:19

data ready. Am I right?

3:21

So given you mentioned 2324,

3:23

a lot of stuff was

3:25

a long time ago. Yeah,

3:27

in the AI world, right?

3:29

Uh-huh. A lot of the

3:31

hallucination risk conversations that were

3:33

happening, and this was mostly

3:35

around Czechy PT, was because

3:37

the information that was available

3:40

was up to a particular

3:42

date. It was predominantly unstructured

3:44

data available on the internet.

3:46

So if something was published

3:48

past a certain date, it

3:50

was not going to show

3:52

up in answers. One of

3:54

the big changes I think

3:56

in the ecosystem is... In

3:58

the past, we talked about...

4:00

IT solutions, data warehouses, analytics,

4:02

which was separate than marketing

4:04

segmentation and engagement. And then

4:06

we had these, you know,

4:08

really interesting LLM and generative

4:10

AI for the past several

4:12

years, several months, it feels

4:14

like years. The focus is

4:16

the idea of a truly

4:18

enterprise scale data platform that

4:21

can power automation, that can

4:23

power analytics, that can power

4:25

analytics, that can power analytics,

4:27

that can power analytics, that

4:29

can power analytics, that can

4:31

power analytics, that can power

4:33

analytics, that can power analytics,

4:35

that can power analytics, that

4:37

can look at structured and

4:39

unstructured data in order to

4:41

provide complete, compliant, and contextual

4:43

information that can also power

4:45

AI. I know we bought

4:47

like storytelling. I had a

4:49

really interesting experience with my

4:51

father a couple of weeks

4:53

ago. Do you mind if

4:55

I tell that story? Oh,

4:57

please tell me. I love

4:59

a good story. She's a

5:01

90-year-old retired Burgadier General. It's

5:04

a notary engineer. 90 years

5:06

young, you mean? Oh man,

5:08

I still barely keep up

5:10

with him. See, that's what

5:12

I'm saying. And as a

5:14

military engineer, you're always given

5:16

a mission and you have

5:18

what you have, right? That

5:20

is the typical mindset. And

5:22

in every country and every

5:24

place people are talking about

5:26

artificial intelligence, what it may

5:28

mean, and he said, okay,

5:30

look, I think this is

5:32

your field. Help me understand

5:34

what is new versus what

5:36

he was working with. in

5:38

older computing days. And why

5:40

are people worried? Why are

5:42

people excited? So I sat

5:44

next to him. We brought

5:47

up CHATGPT and I asked

5:49

a series of three questions.

5:51

The first question was, I

5:53

said, tell me what you

5:55

know about, you know, my

5:57

dad's name, Sunday or a

5:59

retired engineer, a retired soldier,

6:01

not even rank, and it

6:03

gave... What rank he retired

6:05

at, what branch of the

6:07

military, where he went to

6:09

school? Simple question, limited. context,

6:11

I said what else do

6:13

you know about them? I

6:15

said him without the name

6:17

and I got information about

6:19

article fields written for magazines

6:21

in a couple of his

6:23

books, post-retirement he did poetry

6:25

which is a wonderful way

6:27

to retire and he's like

6:30

oh it's interesting how does

6:32

it know that I'm like

6:34

well can people find your

6:36

books in online storage? It's

6:38

like yes though it is

6:40

available information it can leverage

6:42

all of these as it

6:44

searches. It's like okay that

6:46

it makes sense. Then I

6:48

asked a question What do

6:50

you know about his son?

6:52

And Chechie Petee says, I

6:54

do not know who his

6:56

son is. And he's like,

6:58

so why doesn't it know

7:00

we are related? And I

7:02

said, because the fact that

7:04

you and I are related

7:06

would be in a government

7:08

database. It would not be

7:10

in public records. It's not

7:13

something that's on the internet.

7:15

And for him, this was

7:17

an obvious separation. So you

7:19

asked the question. This is

7:21

a long-winded way of getting

7:23

there, perhaps. What have changed?

7:25

What have changed? A year

7:27

ago, I could dump a

7:29

bunch of knowledge articles, or

7:31

perhaps a meeting transcript and

7:33

say summarize, or I could

7:35

use knowledge articles to power

7:37

chat bots. Now I can

7:39

look at what do I

7:41

know about a person in

7:43

their transactional context based on

7:45

their order history, based on

7:47

their case history, based on

7:49

their knowledge of the product.

7:51

and I can give much

7:53

more person-lised recommendations because the

7:56

AI platform idea as opposed

7:58

to an LLLM technology idea

8:00

is bringing together matching technology

8:02

where we used to think

8:04

about as duplicate management and

8:06

CRM, right? That mindset has

8:08

evolved and it is there

8:10

to provide contextual interactions. DataCloud

8:12

is not just powering the...

8:14

generative AI capabilities for Agent

8:16

Ford, it is also providing

8:18

the unified insights that can

8:20

even... be constrained to only

8:22

what a person is supposed

8:24

to know about where admins

8:26

and architects can control this

8:28

given the permission model and

8:30

capabilities of flow, which for

8:32

me is incredibly exciting because

8:34

that means we can deliver

8:37

more value, we can use

8:39

the technology we are already

8:41

deeply familiar with, and we

8:43

can show the true potential

8:45

of AI while minimizing risk

8:47

to our organizations. and minimizing

8:49

confusion for our end users.

8:51

That's a fabulous story. I

8:53

feel you're spot on. Just

8:55

the level of understanding why

8:57

and what we have available

8:59

to us is huge. In

9:01

the email you sent me,

9:03

I want to pull out

9:05

a sentence because we're talking

9:07

about data and we're talking

9:09

about a lot of things,

9:11

but I think I feel

9:13

this is a good foundation.

9:15

You said, part of this

9:17

helped them realize why historical

9:20

CRM data management techniques do

9:22

not scale versus benefits of

9:24

data cloud. To the uninitiated,

9:26

and I'm one of them,

9:28

so I'm asking this question

9:30

for me, can you give

9:32

me what you mean by

9:34

historical CRM data management techniques

9:36

and help me understand that

9:38

versus the benefits of data

9:40

cloud? So if I think

9:42

about What is in the

9:44

Salesports Admin Data Management Toolkit?

9:46

We talk about a distinct

9:48

set of areas. We expect

9:50

admins to do. They configure

9:52

objects, object fields with validation

9:54

rules and some data management

9:56

rules such as do you

9:58

want to default value or

10:00

not, if it's required or

10:03

not. We talk to them

10:05

about duplicate management rules. which

10:07

led the impression that all

10:09

the fociates are bad and

10:11

we talk about storage optimization

10:13

more around performance because in

10:15

every arc had a storage

10:17

limit, you wanted to think

10:19

about when you may want

10:21

to offload storage either for

10:23

cost savings or build like

10:25

skinny tables for large data

10:27

volume handling. Those were the

10:29

domains of data management we

10:31

got to, which was fairly

10:33

technical focused on mostly data

10:35

entry operations. Let's fast forward

10:37

to even two years ago,

10:39

if you have a sale

10:41

for Sierra Morgue with Experience

10:43

Club. You need to have

10:46

intentional duplicate records because the

10:48

records and end-user maintain their

10:50

information should be separated then

10:52

how that customer's information is

10:54

maintained by employees. You may

10:56

also have records maintained by

10:58

partners. using Experience Cloud that's

11:00

still about the same customer.

11:02

So already thinking that for

11:04

a customer they should have

11:06

one record is no longer

11:08

sufficient and acceptable because partners

11:10

need to have their view

11:12

of the information, customers want

11:14

to maintain their own perspective

11:16

of what they're called what's

11:18

their best contact information, and

11:20

companies want to be able

11:22

to have their internal view

11:24

as well, such as... in

11:26

a customer segment, customer risk,

11:29

so on and so forth.

11:31

But personalized engagement requires a

11:33

complete understanding of what's happening

11:35

with an organization where you

11:37

only act on information you're

11:39

allowed to see and you

11:41

act on insights that is

11:43

relevant to the outcomes you

11:45

want to achieve. So three

11:47

things I really, really like

11:49

that data cloud brought in.

11:51

to our solution kids is

11:53

first I can provide the

11:55

holistic understanding of the individual

11:57

or a business contact even

11:59

So I have multiple contact

12:01

or lead records in my

12:03

CRM, even in this simplest

12:06

of architectures. Let's talk

12:08

about a non-profit example. Let's

12:10

say that, you know, we're

12:12

talking to Sam Ms. And Sam

12:14

is a donor. Sam was a

12:16

board member. Sam worked for an

12:18

organization that gave us grants. That

12:20

is us interacting with Sam

12:23

Dehuman into a business context

12:25

and in a donor relationship.

12:27

We are going to want to

12:29

track these through different departments,

12:32

probably through different records. But

12:34

when we want to know what do

12:36

we know about the people we engage

12:38

with, how do we send them a

12:40

person lives? Thank you. This is

12:43

where DataCloud powers that

12:45

unification perspective. Does that example

12:47

make sense before I tie to

12:49

the AI specific examples that extends

12:52

this? Yeah, no, it does. I'm following

12:54

along. So let's say that we

12:56

are now in a data

12:58

model that we have accepted

13:00

we should maintain contextual transactions

13:03

in our business applications,

13:05

whether we have one or multiple

13:07

CRM works and of course other

13:09

systems. We first unify it

13:12

around individuals business context

13:14

and accounts for now

13:17

related transactions, related emails.

13:19

donation history from external

13:21

systems or cases regardless

13:23

of your industry can

13:26

come together in one umbrella. Now,

13:28

if I want to create a

13:30

personalized thank you message, we can

13:32

look at overall interaction history

13:34

and not just think that we

13:37

have seen someone for the first

13:39

time because they are using their

13:41

new email address in their new

13:43

corporate role, but they've been

13:45

a lifetime member. So, generative

13:47

AI solutions work better. When

13:49

interactions across a person's

13:51

contact points can be

13:54

made accessible, wooden compliance

13:56

rules of course, and agentic

13:58

solutions work better when

14:01

it can understand what are

14:03

all of the different type

14:05

of transactions that may be

14:07

associated to an individual or

14:09

an account, even when they

14:11

are distributed across multiple accounts

14:13

records, multiple contact records, even

14:15

multiple theorem orgs. You know,

14:17

you can see me, I

14:19

am pointing to things in

14:21

on the whiteboard in front

14:24

of me, but this is

14:26

something that used to take

14:28

organizations months is not years

14:30

to put on place. And

14:32

having done this now, like

14:34

for real, with a few

14:36

non-profits as part of my

14:38

pro bono work, I know

14:40

we can do assessment and

14:42

planning in a few days.

14:44

We can then onboard the

14:46

data and configure data clouds,

14:48

data unification capabilities in less

14:51

than a month. And that

14:53

includes identifying bad data that

14:55

is in the system in

14:57

A&A.com. They are still present,

14:59

whether you're August 3 years

15:01

old or 20 years old,

15:03

by filtering out irrelevant data,

15:05

by putting directly the standardizations

15:07

in place. These are all

15:09

part of a single umbrella

15:11

of capability, where as an

15:13

admin, you just worked with

15:16

the admin tools in the

15:18

past, and now many of

15:20

these transformation capabilities, configurable rules,

15:22

are accessible, fill under the

15:24

setup tree, fill under the

15:26

Salesforce tabs, that allows us

15:28

to be... more productive field

15:30

sports professionals and allows us

15:32

to decrease the total cost

15:34

of ownership as we support

15:36

our organizations. I mean, I've

15:38

always thought when I've asked

15:41

people a rhetorical question, what

15:43

is the most important thing

15:45

that your company owns? And

15:47

99% of the time when

15:49

I ask people that question,

15:51

they get it wrong because

15:53

they mention a patent or

15:55

a brand. or a

15:57

product that they produce. And I.

16:00

say no, it's your data. The

16:02

data that you have is the

16:04

most important thing for you to

16:07

take care of. And ironically, it's

16:09

also the most, least paid attention

16:12

to because we just throw things

16:14

in and we'll sort it and

16:16

figure it out later, right? Hurry

16:19

up, move on to the next

16:21

thing. And now, as you bring

16:23

up, the unification of all these

16:26

systems, or we've put all this

16:28

data, and the management or mismanagement

16:31

of it now is the vital

16:33

importance because now we can truly

16:35

link all of this information and

16:38

have AI sort through it and

16:40

give us the relevant information that

16:42

we need by just thinking through

16:45

a few more processes. I think

16:47

what's important and what you said

16:49

is AI is additive to what

16:52

we have had because I agree

16:54

data is the most important asset

16:57

and the fact that No organization

16:59

I've ever been a part of

17:01

or helped had perfect data is

17:04

something we just need to accept

17:06

but not live with. Right. I

17:08

have a friend that has a

17:11

small marketing agency. He probably has

17:13

200 people in his little CRM.

17:15

I promise you his data isn't

17:18

good. Even that, right? I mean,

17:20

nobody's got perfect data. So what

17:23

matters is, and this is what

17:25

we talked about last year is.

17:27

We can't assess data quality as

17:30

a technical concept. We can't just

17:32

look at what is in my

17:34

object. Is it good? Is it

17:37

not good? We always need to

17:39

look at data in context of

17:42

a business outcome. I think an

17:44

example I give often is how

17:46

much data you need to start

17:49

an opportunity is different than the

17:51

amount of data you need to

17:53

close on opportunity. What you want

17:56

to gather if you lost a

17:58

big opportunity is different than what?

18:00

you probably would ask people to

18:03

capture if you lost a small

18:05

opportunity. So these are all proportionate

18:08

to the business benefit, where I

18:10

don't think historically we did a

18:12

great job explaining as professionals, whether

18:15

we are admins, architects,

18:17

business, endless. But when it

18:19

comes to AI, because agentic AI

18:21

puts so much focus and emphasis

18:23

on use cases and the persona

18:26

we are empowering. If it is a

18:28

sales agent, we want to find

18:30

out what is the job a

18:32

sales agent is supposed to do,

18:34

what is the information they

18:37

need, what are the rules they

18:39

should follow, and whether you have

18:41

100 fields or 800

18:43

fields in your accounts

18:46

and opportunities objects, we

18:48

still need to look at what

18:50

data is reliable today.

18:52

Is that sufficient? If it is

18:54

not sufficient, we need to go

18:57

through some type of data improvement

18:59

process or when to look

19:01

at a different use case.

19:03

If we have sufficiently

19:05

reliable data, we need to look

19:08

at how do we ensure our

19:10

prompts both use data from

19:12

those fields that have

19:14

reliable data and sufficient

19:16

metadata? And also know

19:18

when a subset of records

19:20

don't have sufficient data

19:22

quality in those very same

19:25

fields. And then third, just

19:27

because it works today, we

19:29

shouldn't assume things are going

19:31

to be the same tomorrow

19:33

because processes are changing, configurations

19:36

are changing, people habits

19:38

are changing. So by monitoring

19:40

what's happening in our business

19:43

applications and catching deviations.

19:46

we can avoid unexpected bad

19:48

surprises also in the

19:50

flows. Honestly, these are things

19:52

with a time machine we should

19:54

have taught off and incorporated

19:56

into our automation flow into

19:59

our reports. the attention wasn't

20:01

there as much as it

20:03

is today, though people being

20:05

excited about AI, but we're

20:07

about to this nation risk,

20:09

is one of the best

20:11

things that happened to ensure

20:13

we can provide reliable data

20:15

for all types of decision-making

20:17

through Salesforce. Right. Well, I

20:19

mean, what do they say?

20:21

2020's hindsight. If you could

20:23

go back and know the

20:25

future. then you'd obviously plan

20:27

for it, but it also

20:29

creates opportunity for us to

20:31

be creative and corrective in

20:33

how we move forward, which

20:35

means that every solution you're

20:38

thinking of today moving forward

20:40

is going to look very

20:42

different than before Agent 4s.

20:44

You know, one of the

20:46

things I'm still noodling on,

20:48

and I'll probably spend a

20:50

few more years newtling, is

20:52

how do we make sure?

20:54

we can take better advantage

20:56

of unstructured data that is

20:58

the best majority of all

21:00

interactions. I remember being excited

21:02

about Einstein activity capture, which

21:04

was a few years ago,

21:06

and it's still an untapped

21:08

potential, but now we are

21:10

analyzing that data, we're incorporating

21:12

that data. The more we

21:14

can streamline the end-user experience

21:17

to capture information. know when

21:19

information may be missing in

21:21

complete, you know, potentially out

21:23

of date. So they can

21:25

improve it in a tactical

21:27

surgical way and then be

21:29

able to explain to them

21:31

why we're making certain recommendations

21:33

in AI assisted suggestions. I

21:35

think that's also going to

21:37

increase the confidence for agentic

21:39

experiences where humans are engaged

21:41

in a secondary level. Like

21:43

I know that. I would

21:45

like AI tools to give

21:47

me results I can believe

21:49

in when I'm directly engaging

21:51

first. Before I'm willing to

21:53

expose it to perhaps less

21:55

savvy or less aware of

21:58

my underlying processes and users,

22:00

I think a lot of

22:02

people are going to go

22:04

through the journey. So thinking

22:06

about process mapping, thinking about

22:08

testing strategies are also going

22:10

to be important considerations for

22:12

all of us. I mean,

22:14

that's the whole point is

22:16

to test, right? You want

22:18

something reliable. and to ask

22:20

why, I think the important

22:22

thing is people get things

22:24

wrong too. You know, we

22:26

sometimes look at some of

22:28

these technology solutions as infallible

22:30

as they're always perfect and

22:32

they're not. They're imperfect because

22:34

they're built by imperfect people.

22:37

But that doesn't mean that

22:39

you can't constantly iterate on

22:41

your solution. I remember long

22:43

time ago when I was

22:45

an admin, I feel like

22:47

it was Josh Burke or

22:49

it was another developer was

22:51

always like, you know, every

22:53

year I look at the

22:55

code I wrote for the

22:57

previous year and wonder, why

22:59

did I write it that

23:01

way? And it's because you're

23:03

a year smarter. 100 percent

23:05

agree. I have like 13

23:07

years of still sports presentations

23:09

in my tropics folder and

23:11

when I look at it,

23:13

it's fascinating to see what

23:16

is still true. And it's

23:18

interesting to see when some

23:20

recommendations have completely changed over

23:22

the course of the last

23:24

15 years. Because we are

23:26

learning, and I think we

23:28

need to be honest about,

23:30

look, yes, this was the

23:32

recommendation based on what we

23:34

knew. Here's what we learned

23:36

since, and here's why we

23:38

are recommending X today, that

23:40

is different. I think on

23:42

the other side, as professionals,

23:44

we need to remember. None

23:46

of us have all the

23:48

answers and what we knew

23:50

yesterday might have changed today.

23:52

So look, I love your

23:55

podcast. I love some of

23:57

the things that come out

23:59

of the various... blogs because

24:01

people share what they've learned

24:03

at a level of frankness,

24:05

including what we stopped

24:08

doing. And that is a fine off

24:10

of being learning humans. And

24:12

it's the best way to

24:14

be. Absolutely. How do we

24:16

learn to be better every single

24:18

day? Well, I feel like that's a

24:21

really good place to end

24:23

this episode on because... I always

24:25

want to learn more and I

24:27

appreciate you coming on the podcast

24:29

and helping everybody else learn more. It's

24:31

my pleasure. I know that, you know, there

24:34

are so many thoughts we can always get

24:36

into. I hope these sessions enable

24:38

more personal connections and if

24:40

you're listening to it and we

24:42

run into each other at an

24:44

event, let's grab coffee, let's talk

24:46

about data or life because we're going

24:48

to learn from each other, we will

24:50

make each other better and thank you

24:52

Mike for the opportunity. Absolutely.

24:55

So Mamet took us for

24:57

a ride from a

24:59

90-year-old general, his father, all

25:02

the way to data that

25:04

doesn't quite behave. And the

25:06

takeaway? Well, AI is like

25:08

a great intern. It's only

25:10

as good as the notes

25:12

you give it. So let's

25:14

feed it well and ask

25:16

better questions. But anyway, huge

25:18

thanks to Mamet for the

25:20

wisdom and the stories. If

25:22

you learn something today or

25:24

you just enjoyed the ride,

25:26

can you do me a

25:28

favor and just share the

25:30

podcast and spread the data love? Now

25:32

until next time, we'll see you in

25:34

the cloud.