Will AI Kill Software? — With Salesforce Co-Founder Parker Harris

Channel: Alex Kantrowitz

Published at: 2025-10-14

YouTube video id: qZD0Ti2a0xM

Source: https://www.youtube.com/watch?v=qZD0Ti2a0xM

What does the future interface of
software look like? Let's talk about it
with Parker Harris, Salesforce's
co-founder and the CTO of Slack who
joins us here in a YouTube exclusive
today. Parker, great to see you. Welcome
to the show. Awesome. Thanks for having
me, Alex.
Thanks for being here. So, let's just
start with this. You don't do many
interviews. I think the last interview
you did, if I can recall, was with
Nikkei in Japan.
Um
what makes you decide that this is the
time to do it? What's going on in the
tech world that's pressing you into
action? Uh well, first of all, I thought
it would be a fun interview.
Um
the style of this interview I thought
would be pretty awesome. And um
you know, we have our Dreamforce
conference next week. It's our third
23rd Dreamforce. We're going to have
50,000 people coming to the city and
we're going to talk a lot about AI. Um
and I'm just excited to be in the
industry right now. Um I think anyone
who is in a consumer-oriented company or
an enterprise company like Salesforce
and is we're all trying to figure out
how to leverage AI
for our customers. Um it's a pretty
exciting time. So, I'd love to talk
about AI and and every aspect. Um cuz
it's just uh
it's it's I think it's the biggest thing
I've ever experienced and I feel like
I've experienced a lot, you know, with
the rise of the internet and enterprise
software and phone and social and
mobile. But uh but AI I think uh tops it
all. Now, is there something about I
mean, obviously there's AI is an
interesting technology, but there's been
many technology shifts. So, is there
something about AI in particular that's
made you say this is the moment where I
need to go speak to the public about
what's happening? Yeah, for sure. You
know, I mean, we
we leveraged this shift to the internet,
you know, when like when we started the
company it was like the motto was you
know, we want um
software or Salesforce automation be as
easy as buying a book on Amazon. Um so,
the internet uh you know, software as a
service was a big shift in the market.
Um iPhone came out, uh transformed like
we're doing work from our phone. Uh
social came out and uh we added uh
social into enterprise software. Um we
had predictive AI back in 2014, 2016. At
least at Salesforce, that's when we
really were taking advantage of it.
Um but I really think generative AI is
bigger than all of that because
we've had to completely retool uh our
product strategy and our company uh and
we're still in the middle of it. And you
know, you'll see a lot of that uh next
week, but uh we're going as fast as we
can because you think about Salesforce
leveraged the internet to go after um
previous companies like Siebel Systems
uh who were doing who were basically
shipping software, handing you a CD
saying, "Alex, you need to buy some
computers, hire some great people." And
we said, "Oh, you don't have to do
that." Um and now we're in a world where
um if we don't go fast enough in this
world of agentic AI, some competitor
will leverage us leverage it to um to
skip past us and and we're not going to
allow that to happen.
So, actually I'm actually curious. When
you think about being disrupted by a
company
uh that is leading further into agentic
stuff than you are, can you just talk me
through the sort of brainstorm of what
would happen if you were to stay still
and how a company could disrupt
Salesforce if you were to do that?
I mean, so some of it is fiction right
now, but I'll I'll play out the fiction
for you. Um
What if
AI got to the point where it could just
listen to this conversation. It could
read your emails. It could read your
Slack messages. Um
It it could just be with you all day
long. And actually our our head of
revenue has some device. I don't even
remember what its name is, but he It
would record
>> or Oh, no, that Is that the um
Oh, I already forgot what it was called.
It The uh Apple people
>> everything he says and everything people
say to him. I mean, I'm sure there's
privacy implications there. Um but the
idea that um somehow it will take all
this information and process it and um
tell you what's important. It'll It'll
run your business. It will find out who
uh who out there might, you know, buy
some goods or services from you, uh you
know, prioritize those um without any
structure, without any business uh
logic, with business process. It'll just
figure it out.
And you know, maybe one day we'll get
there and maybe one day we'll get to
AGI. Um but you know, I think it's going
to be a ways out and uh you know, so so
the fear is someone's going to come
along and leverage AI in ways maybe we
haven't seen to um to basically obviate
the need
for
SaaS or enterprise software. You know,
Satya Nadella like this is the meme
going on out there. It's like, "Oh, SaaS
is dead, you know." And if you are a
model provider or
one of these frontier that. You
definitely want to be out there saying,
"Oh, yeah, like we're taking over
everything."
But you know, when I talk to our
customers, what they're saying is, "We
just want to get this stuff to work, you
know?" And there's a lot of great
demonstrations out there. Um you know,
we're we're going to try to show real
software and real customer stories next
week so you can you know, because
customers are looking for customer
success right now. They're looking for
trust. They're looking for customer
success. These are our corporate values.
Um and and I think we're getting fooled
a little bit by the state of the market
with with so many amazing demonstrations
of what's possible. Um but as you know,
Alex, like this technology can work, you
know, maybe five times out of 10, seven
times, nine times out of 10. Uh but in
business, you know, you need that
predictability and that assurance that
it's not going to mess with your brand.
It's not going to ruin customer
relationships.
Um it's not going to, you know, I think
there was that one example where uh was
it a travel company? They stood up an
agent and it was giving away crazy
discounts and then they said, "Oh, that
wasn't us. It was the agent." You know.
>> dealership that did that as well, I
think that was giving people
>> And so,
there's kind of this two things. There's
this FOMO of like, "I got to go as fast
as possible." All our customers, they
you know, there's a lot of
do-it-yourself AI happening out there.
Um and yet this fear of like, "Okay, I'm
going fast, but what happens, you know,
if I break things?" And some things I
just can't break. I don't want to break
my company. And so, um that's very
disruptive. And what we're trying to do
is come in with the right software, the
right uh values of trust and customer
success, and our people on the ground
with our customers. And that's what
we've been doing um really ever since,
you know, we launched agent force last
uh Dreamforce. So, it's only been a
year.
And uh now we have 12,000 customers on
it. Um we're in there with them. Uh you
know, this stuff needs some hand-holding
at first. You need to make sure like you
are tuning Do you have the right data?
Like if it doesn't have the right data,
it's not going to make sense. You know,
do you have the right security model?
Have you tuned uh the agents for brand,
for voice,
um for the right skills?
Uh and are you listening, you know,
based on that to the user's feedback?
Like is it working every day? And it
gets better and better.
Um but it takes iteration. It takes a
lot of focus. You know, and I think this
is going to get easier and easier, but
uh we're still in the early days.
So, having your CTO hat on, um
I really want to talk to you about this
just to go back to it once more. Uh
these claims that uh business software
or SaaS will go away. Um I remember
early days early on when ChatGPT came
out. And for viewers, this will be, you
know, a familiar story that I'm going to
tell. But Benedict Evans, the business
analyst, said the future of AI is not
going to be like chatting with your
Excel spreadsheet. It's going to be
maybe you put all your data into a
chatbot and then you query the chatbot
as opposed to like have Excel running as
a program and stapling an LLM onto it.
And from that moment I said to myself,
"Well, that makes If you can do that
with Excel, you can do it with all all
business software." And I think that's
what you're talking about. That's the
fear.
Uh but now, like you said, we're we're a
couple years into this. We're just about
to hit the 3-year anniversary of
ChatGPT. And I'm more skeptical of that
uh
throw everything into the large language
model um than I ever have been. If you
think about like I think you've already
hit on a couple of these things. Um
where do you store your data? What type
of security do you have? User
permissions.
Um maybe if you built like a very
elaborate database inside an AI
assistant, you would you would maybe be
able to replace SaaS. But to me, this
idea that business software is going to
go away and and fall, you know, we were
going to replace the graphical user
interface and it's going to fall away in
favor of the chat interface, I'm not I'm
not quite there yet. I think we're
headed there. But um but you know, you
can't just like I want to take all the
data of an enterprise and dump it in one
place, you know? And you're not going to
dump it into an LLM because the LLM has
no security model, you know? And it it'd
be great if you're the CEO. You'd be
like, "Yeah, just give me all the
corporate data. Put it in an LLM and let
me ask it questions." And and that could
be interesting if you're the CEO.
But if you're a public company and you
know, you're about to do an acquisition
or you you about to announce earnings,
You do not want all of that data exposed
to every single employee in your
corporation. So, you need data
governance and data quality, data
security.
Um and
like the promise like think about years
ago people were saying, "Oh, let's get
all the data in one place and I'm going
to put it in a data lake." And how many
failures we've had where it's like
you know, by the time you get it in one
place, the data's stale, it's not
accurate, um and you don't have that
security model.
Um what we're doing at Salesforce is
we've leveraged our our data cloud,
which is really think about it's a data
activation platform to connect to all
the source data sources. So, there's
data that's pouring into the data our
data cloud from all the services we run
on behalf of our customers. And then
they're connecting it to all their other
data sources so that you don't have to
move the data.
And uh and it's honoring the the
security model Salesforce. If it's
reaching out over to SAP, it's
leveraging the security model of SAP,
the security model of G Suite or Office
365
and bringing it together.
Um
and so, you're not going to get the data
one place and that's a hard problem to
solve. You know, you can't just like put
it, you know, if you're a corporation.
That's why you need vendors like
Salesforce to come in and help you.
Um and then, you know, you said, "Well,
um
you know, I might want to talk to a lot
of the data all at once." It's like,
"Give me Give me some insights." Or I
may be doing a specific task. I may want
to talk to that that spreadsheet or that
document. You know, maybe I'm building a
presentation
uh to articulate uh you know, a strategy
of my corporation or I'm doing a
customer pitch. I may want to talk to
some AI leveraging some data through a
secure channel to help me, you know,
like it's like
I might work with a number of humans to
do that, but what if there are a few AIs
in there with me helping and wouldn't
that be a good thing? And so, I don't
think it's one of the other. And And
then, the final point you made is
is the graphical user interface going
away?
Um
I think it's changing. You know, I think
the days of going into like I did so
much work building what we call
Salesforce Lightning. So, when we
started the company, the single page app
was
was not a thing. This was 1999.
We were just trying to get something
like when we started the company, one of
our um
you know,
goals was it we needed to be fast at
56K.
For those listeners, that's a modem.
You may not remember that, but 56K baud
modem is not fast. So, we're just trying
to get something work. It was like a
mainframe application. Let me go to the
server, give me a bunch of HTML. And
then, we came out at this as the
graphical user interface improved with
with web applications and you had more
of the single page app with things like
React and all the cool things today, we
built Lightning.
Uh and we worked like crazy to make it
an incredibly easy interface.
But do you really want to go in like if
you're if I'm a salesperson and I'm out
in the out on the road uh or I'm a field
service technician or you know, or I'm
just an employee and I'm working from
anywhere,
do you really want to you know, log into
a big interface if I could have it
coming to me conversationally? So, you
mentioned OpenAI's interest in being
this sort of catch-all place uh where
you would interact with your apps. You'd
interact We talked about this on the
show recently. You'd interact with
Spotify or
uh I don't know. I mean, it's the same
examples every single time. Call an
Uber, book a ticket. Uh I've heard the
same demo from OpenAI, from Apple, from
Amazon, from Google, and from Facebook.
Uh and and no one has built it
effectively yet. And then, you look at
the enterprise side. Um and I'm just
going to Let me Let me challenge the
idea here for a moment. We had um we had
Thomas Kurian. He did an interview with
me, the head of Google Cloud for my
Substack recently, and talked also about
using Gemini as a command line interface
to um
connect with a lot of your business
software. And we had a comment on there
uh from someone who's been through this,
you know, similar on the enterprise
side. Who's lived through these pitches
of centralization and using natural
language as the command line for
business software. And they said,
um "Enterprises are big, messy,
complicated places. I've been hearing
sales pitches for magic technology magic
technology to solve all their
information problems for 25 years. None
have lived up to the promises. You know,
third verse same as the first."
What do you think about that?
I agree and disagree. So, you know, I
think that um
it's very possible that the browser as
we know it, which is a it is a
centralized tool that we use,
um you know, whether you like it or not,
it's a centralized tool that we all use
to do a lot of our work or to do a lot
of our consumer activity. And it's
great. And uh is that changing? And And
you know, everyone's trying to move you
towards like the fear is, well,
does Google lose uh share because people
do not go into the Chrome browser and
search on Google even though I get my
Gemini results
and instead I go to ChatGPT. So, that in
the consumer world. So, they're all
fighting for that uh mind share.
Um
you know, I think in the enterprise, you
know, I agree.
I mean, the problem with AI is this this
magic demo. It's a great magic trick.
It's like, "Check it out, Alex. I can do
all my work just talking to AI. It knows
everything. It tells me what's
important. It's perfect." And it's a
demonstration.
And it doesn't always work that way.
So, I agree that, you know, we need to
move to the future where sometimes it
will work really well and sometimes I do
want to talk to AI and have it, you
know, tell me about my business from an
analytical perspective as I said, or I
want it to automatically you know,
automatically, you know, look for a bug
in my code and automatically post an
update to my code to get and you know,
and do a review of it. You know, yeah,
all of that is good. But it's humans and
AI working together. And so, what I
think again, I'll Sorry to keep talking
about Slack, but Slack is is the
multiplayer. So,
I've gotten all the cool kids. So, all
these young people work for me now and
so, they taught me the language that
it's a multiplayer environment where
many humans are working together. But
now, there's also AI working together
with you. And are you only working with
AI? Are you only you know, doing the
work talking to AI in a single, you
know, single player interface or even
the AI with you in a Slack channel, for
example? Not always.
You know, I I believe that uh I'm an
optimist. I believe that humans are
going to be are going to rise with, you
know, the power of AI. You know, they're
not going to be replaced. And I want I
still want to work with a lot of great
people. And so, when I need to get my
work done, uh sometimes I'm working with
AI, sometimes I'm working with humans.
When we talk about um
you know, Slack first is an internal
term we use like, how do we think about
working in that conversational
interface? I may be working with AI or I
may be just bringing in the assets from
these other just like you know, the
engineering groups have done, you know,
where I'm not logging in to get
directly. I'm not logging into Workday.
I'm not logging into Salesforce. I can,
you know, I can introspect my data. I
can work with pieces of uh you know, the
workflow, the UI, and all of that in
Salesforce without logging into
Salesforce.
And so, I think it's going to be a
mixture and we're going to listen to our
customers. Um and they're going to tell
us, yeah,
this works, this doesn't work.
Um and you know, I
I I like simplicity and I like, you
know, when things work well.
And the AI doesn't always work well. And
so,
you have to have a fallback. You can't
say like the AI is going to be the only
way of doing things.
Um and so, maybe that's a a different
answer to your earlier question, you
know, is SaaS going away or is the
graphical user interface going away?
It's not going away.
There's still a lot of good use cases
for it, but you know,
our job as leaders in this industry are
is to try to push forward to the future.
We're going to show the future. It will
provoke those reactions of that's not
the only thing. It's kind of like when
we started the company,
we had a tagline no software. It's like
no software. And people would say, "I
know you have software."
You know, you're building a lot of
software.
And that was just a motto just to say,
"We're not going to ship it. We're not
going to make you deal with it." Um
and so, I think we're just in that in
that, you know, mode again, but on
steroids with AI.
Okay. Uh I'm actually curious to hear
your perspective about the underlying
technology.
You mentioned earlier in the
conversation, you know, maybe we'll get
to AGI. Uh
there's also a conversation now that a
lot of the progress of the models has
leveled out and it's mostly about
orchestration now.
Where do you fall on that front?
You know, I I think um
the pre-training
I'm I'm I'm not an expert in this, but
it does seem like post-training is where
the best innovations are happening now
and the pre-training and the amount of
data like they've we've used up a lot of
the data. They're trying to create
synthetic data to try to improve uh
model performance. The models keep
getting better though. Like when we keep
get new new drops of models from
Anthropic or from Open AI or you know
any of the other providers.
They are getting better. Um and maybe
better is, you know, not always
significantly smarter, but just more
deterministic, you know.
Um Less hallucinations. Less
hallucinations.
Um
But for us, like I don't need
a radically smarter model. I just need
to make these models work well for like
back to the simplicity, for the inner
use cases in the enterprise.
Um
We're coming out with something called
Agent Script uh here at Dreamforce. And
Agent Script, it's kind of funny, like
a lot of the companies like we look at a
lot of companies, partners, companies we
want to buy,
competition. Everyone's kind of doing
the same thing, which is
we started with like put everything in
the LM, make the LM the brain, and tell
it what to do, and you know give it the
instructions, and it's going to be
amazing. And what they found is
like let's say you Adidas is one of our
customers. Let's say that they, you
know, want to you know, they add, you
know, return.
Uh I want to return my sneakers. And
there's a series of steps Adidas
expects.
Um or William Sonoma, you'll hear a
story about William Sonoma at uh at
Dreamforce.
If you tell the LM, do this and do this
and then do this and do this,
it doesn't always do every step. You
know, it's non-deterministic.
And so what we've done is we've pulled
out um in in the reasoning loop in the
brain,
we've added a state machine
for lack of a better term. So, uh
workflow state machine, what have you.
Um but we made it really simple. It's
called Agent Script. So to basically
say, I want you to do these things, and
as you do those things, I want to use AI
on each step.
And it adds a blend of determinism with,
you know, non-determinism or you could
say determinism with the brilliance of
large language models.
And I think that is what our customers
really have been looking for, one of the
things that they've really been looking
for
um to make AI work. And so, we're not
waiting for AGI. We're just trying to
get the the amazing AI I mean that the
AI we have today can transform every
corporation. We just have to make it
easy to use, fast to deploy,
um a little bit more deterministic, and,
you know, and a platform around it to
know that that's all true.
So, when you think about the ways that
the labs could improve models that would
basically increase the amount of tokens
that you might use, is there anything
they can do now just making it
error error make it yeah, less
error-prone or capable of handling I I
mean it seems like this is really
getting to the end of
what's needed?
I you know, I I think maybe we're going
to hit I don't know, but maybe we're
going to get to a point where um
we we don't we're not a model provider.
We have a research group that's doing
some work in models and
um and we're playing around the idea of,
you know,
would would fine-tuning some of these
models into the use cases of what we do?
I mean the domain of what we provide is
so much smaller than the domain of what
an Open AI is trying to solve. It's
trying to be the brain of everything and
ask any question and it'll answer any
question.
We're trying to provide services to our
customers to improve their sales,
improve their customer service,
make their marketing more targeted and
more personalized, improve e-commerce,
give you better analytics, better
collaboration with Slack. All of these
are finite areas actually and doing it
in industry industry-specific areas. So,
we're also looking at, you know, would
it help to have a financial services
fine-tuned model for CRM? For specific
use cases. ADI and you know, you say,
well, you know, are these frontier
models you say, okay, you don't need to
speak Shakespeare. It's okay.
I don't need Shakespeare, but, you know,
I do need you to understand Salesforce
automation and Salesforce automation in
healthcare and in in in payer provider,
you know, maybe it's the payer market.
So you know, it keeps getting more
narrow.
Maybe we could have um more effective AI
with fine-tuning that would give you
better determinism because it just
understands that space better,
and it understands it doesn't have to
have the that wide-open realm of
questions that it needs to answer.
Right. Okay, talk a little bit about the
pace of adoption. You know, us on the
outside, we see these studies the like
the MIT study, which I'm sure has some
problems with it, that a lot of that the
vast majority of companies aren't
getting
uh an ROI on their AI deployments now.
And then we saw some headlines recently.
Uh there's a couple of them like this
one from Reuters, Salesforce projects
weak growth sales fueling AI anxiety.
Um is the the adoption slower than
people are expecting or were the
expectations out of whack? What's
actually happening on the ground?
Well, what I see on the ground is um
everyone every CEO said, "Everybody
needs to do AI now."
And this is like 2 years ago. So,
they're like, "Okay, great." Everybody
went off, grabbed models, and started
like, you know,
creating prompts, giving it data,
and creating all these use cases. So it
was it was DIY all over the place. And
there's a lot of failure in that.
Um you know, it was it at first it was
like, "Oh, so cool. Look at look at what
it can do." But again, that
repeatability, that trust, that
security, the data access, all the
things we've talked about. And so I
think we have this, you know, the past
is littered with a lot of failure
of do-it-yourself. And, you know, what
we're trying to do is really just focus
on use cases we do to make our customers
successful. So a year ago, we started
with a simple use case. At Dreamforce,
we said, "Hey,
you know, if you go to salesforce.com,
you go to help.salesforce.com, we
replace the bot with an agent from Agent
Force to answer our customers'
questions." And and a year ago at
Dreamforce, we said, "You can do that,
too."
And we we did a quick uh way of like
just give us your website, and our AI
would then look at your website and
understand if you crawl a website, you
can understand what the company is. And
we could quickly create uh an agent that
you could put on your website. And so
that was a simple use case, answer
simple questions.
And there was a lot of adoption of that.
And then the customer said, "Okay,
that's great, but now I want to do
things. I don't want it to just answer
questions. I want it to, you know,
process a return, you know, check on my
order, check on the status of delivery,
whatever it is, um
uh take action for me, which means I
need more data, needs more skills." And
so we've increased that, you know, with
a lot of what, you know, uh a lot of
people in the market are calling forward
deployed engineers. So we took, you
know, Sreeni Telapragada, our president
of technology, took over service and
support well over a year ago.
And we really really rethought our human
capital and said,
"You know what? We actually, because
this help.salesforce.com the Agent Force
on the website
is really helping us with this, you
know, super simple support scenarios.
We don't need as many support reps to do
that job.
What we do need is to retrain a lot of
those people and get them in our
customers and get them to help our
customers more
uh you know, for those higher level
skills of, you know,
taking action. So, um
you know, I do think it's still early
days. I think in the enterprise, there
is a lot of adoption,
um but there's also a lot of false
stories out there.
Um
my advice is you just got to listen to
the customers and not, you know, the
vendors. That's why at Dreamforce next
week, we're telling customer stories.
You know, we could spend all week, you
know, doing we'll see a lot of demos and
we can, you know, talk about technology
and talk about AI and then, you know,
there could be some amazing stuff, but
listen to our customers.
That's what we've been doing for 26
years is, you know, make a customer
successful, ask their permission to use
them for marketing where they tell their
stories.
Um our customers, we don't wall them off
from each other. They talk to each other
Dreamforce. The good and the bad, too.
Like, hey, if they've got a problem with
something, if maybe an agent's not
working quite correctly, they can tell
each other, but we want to hear it, too.
And we'll be in there hands-on making
sure that, you know, we turn that around
and make it work really well for them.
Um so
>> Parker, I know we're running Yeah, go
ahead.
I know we're running running out of
time. I just want to ask you one last
question. Um do you have a personal use
of AI that uh that you do either for fun
or some
something in your life outside of
business uh that you can share with us
and then we'll close.
>> I use it for learning, you know, I I use
it as my, you know, I'm on a board of my
uh my college um and uh you know, we're
trying to figure out how students should
uh you know,
enter the workplace and figure out how
to be productive and, you know, train
them, you know, what's the pedagogy for
that? Um
And so I use it, you know, I've I've
used it just to like I entered listen to
Andrej Karpathy, you know, try to
explain how all this works, you know,
like
That's great. This is complex stuff.
Like the math behind it, like and they
don't no one fully understands exactly
the magic uh of of where we are right
now.
But, you know, I stopped at linear
algebra in college. And so, I'm using AI
to try to reteach me some things to help
me understand this technology.
It's like and I can ask it questions and
not be embarrassed where I if I asked
you, I might be embarrassed. You'd be
like, "Parker, you don't know that?" And
be like, "Well, you know, I'm still
There's definitely always things to
learn." So. Yeah.
No, it is amazing. I I often think,
"Well, if ChatGPT was a person, they
would think that I was both a real big
dummy and very annoying."
And that's kind of the nice thing about
having it just be technology on the
other side. So,
uh
Fascinating stuff, Parker. Thank you so
much for joining. Appreciate you coming
on this show in particular as you
make a rare public appearance and hope
to speak to you again. Thanks, Alex.
It's a pleasure.
All right, everybody. Thank you so much
for watching. We'll be back on the feed
with another episode soon.