Executives from OpenTable, Land O’Lakes, and Salesforce share how they’re deploying AI agents to transform industries, from restaurant reservations to agriculture and beyond. Discover how these innovators are using AI to enhance customer service, boost productivity, and reshape the future of work.
Category
🤖
TechTranscript
00:00So I'm going to save the lightning round, I think,
00:03for a little bit later.
00:05But can we first just level set here?
00:08And whoever wants to take this, please go ahead.
00:12Can you explain where in any of your organizations
00:17you are actually testing agents today
00:20and in what type of use cases?
00:22Because there's a ton of talk, and I'm
00:25looking for a little more tangible examples
00:29to share with the group.
00:31Maybe that's you, Sagar.
00:32I'm looking right at you.
00:34You're looking right at me, so I'll take it.
00:36So on the OpenTable side, we look
00:39at agents both on the restaurant and the diner
00:42side of the marketplaces, as well as
00:44internal employee productivity.
00:46On the restaurant side, we've seen a lot of early success.
00:50So one use case that comes to mind,
00:55last year we launched a tool that
00:58when a diner leaves a review for a restaurant,
01:02oftentimes the restaurant responds to that review
01:04and says, oh, I'm sorry you had a bad experience.
01:06Oh, I'm glad you had a great time and enjoyed the steak.
01:09That used to be entirely human driven.
01:12So some restaurants would have teams of people going through
01:17and just responding to reviews over and over and over again,
01:21trying to make them personalized, contextual,
01:23relevant.
01:24We took that, passed it into a large language model,
01:28did some prompt engineering, and said, hey,
01:32can you create an appropriate review response?
01:34Here's like 100 examples.
01:37Turns out that more than half the review responses
01:40of restaurants that use this feature now
01:44have AI assistance within them.
01:46So just one example of taking something
01:49that was a lot of work and trimming it down
01:53into something that's much, much simpler.
01:54And are there humans involved at all in this anymore?
01:57And if so, how?
01:58Because I would think as a diner,
02:01maybe I like that little touch from that.
02:04So right now it's an AI assistant.
02:07And so it is the co-pilot that will help
02:10craft the appropriate response.
02:12And then the human editor will review it and decide
02:15whether to publish or not.
02:16And I think there is a future where
02:20we trust these more and more.
02:21We're able to tailor them to the tone and the thing
02:25the restaurant wants to do to let it go on autopilot.
02:30We're not there yet with that feature.
02:32Other things on the restaurant side,
02:34we have some restaurants that have adopted voice AI
02:36to answer the phone.
02:38And traditionally, you're like IVR system.
02:42Everyone hates talking to it.
02:43It's awful.
02:44Nobody wants to press 1 to talk to this thing
02:48or 2 to make a reservation.
02:50These are way different.
02:51They sound like humans.
02:52They answer the question.
02:54They get you to what you need to do.
02:57And there's always a human backstop.
02:59So if you want to talk to a human operator, you can.
03:02You probably won't even notice you're talking to an AI
03:05with some of these tools we've integrated into the platform.
03:08But it is authentic.
03:11It's real.
03:12And the customers actually really like it.
03:14I think one of our restaurant partners
03:16who is using a voice AI partner of ours
03:19said actually 95% satisfaction rate
03:22with that call, which is pretty different than what
03:26you'd expect from professionals.
03:27Definitely different.
03:28And I have so many questions about the workforce impact.
03:31And we're going to get there, because that's a key theme
03:34here.
03:34But Teddy, I want to just turn to you for a second.
03:36So you're chief technology officer at Land O'Lakes.
03:40So a bunch of different divisions at the company.
03:44If someone doesn't work in ag tech or agricultural industry
03:50or grocery, they might not realize that Land O'Lakes
03:54has a chief technology officer.
03:55Can you talk about the ways that Land O'Lakes has
03:58used predictive AI in the past?
04:00But are you experimenting with agents at all now?
04:03And if so, how?
04:04Yeah, absolutely.
04:05So Land O'Lakes has four major groups.
04:09And the one you referenced is the ag side.
04:11So we are a crop inputs distributor.
04:14So we buy from the basic manufacturers,
04:16sell that to ag retailers who sell it to farmers.
04:19The key ingredient in all that is not necessarily
04:21the bag of seed or the jug of crop protection
04:23that we might sell, but it's the insight
04:25that comes along with it.
04:26And this insight in the past resided
04:28with a lot of agronomists.
04:30These were master's degree in agronomy.
04:32They knew for the soil type, the type of climate you're in,
04:35what seed fit best that profile, as long as what
04:38practice go along with it.
04:40So really what predictive AI has been given us to do
04:42is to take that knowledge and make it available,
04:46not to just a master agronomist, but to a lot
04:48of the younger agronomists that are available today.
04:50Matter of fact, we have this thing called a crop protection
04:53guide.
04:53It's like the Bible of crop protection is 1,000 pages.
04:56And that's when we turned into a large language model.
04:59So now the agronomist can have a conversation
05:01with this large language model, and then
05:03turn and have a really fruitful conversation with the farmer.
05:06So that's an area where we've been using predictive AI
05:09for quite some time.
05:10We've had good agronomic knowledge over the years,
05:12good data, and so we've been able to turn that
05:14into a really good use case.
05:17You talk about AI agents.
05:19One area where we are using AI agents,
05:21and that's been really interesting,
05:23is on the animal nutrition side.
05:25So the Purina brand, not the dog and cat, that's Nestle.
05:28But every other animal you can think of,
05:30which is horses, chickens, cattle,
05:33Land O'Lakes is the owner of that.
05:35The cooperative is the owner of that company.
05:37Well, we have a lot of big dealers,
05:39like Tractor Supply, that order from us.
05:41But we also have a lot of mom and pop dealers.
05:43And Mark talked about faxes.
05:46I'm going to hope 2025 is the year we eradicate faxes.
05:50Right?
05:52It's still there.
05:53Some of these groups still fax us orders in,
05:56or pick up a phone and call.
05:58So one of the things we're trying to do is, at least,
06:00can you send us an email of what you want?
06:02From there, the agent will pick it up, figure out who you are,
06:05what your account looks like, what your credit limit is,
06:07what your ordering pattern looks like.
06:09And then, based on the email you sent,
06:11what should your order be for this week or this month?
06:13And we're seeing good adoption with that,
06:15and more and more automation coming there.
06:17The customer service rep now, instead
06:19of having to just do that repetitive work,
06:21they're taking the time now to look at the order,
06:23make sure it's right, maybe upsell or cross-sell
06:26if that comes to fruition.
06:28And then, maybe even call the dealer back if they have to,
06:31and have the conversation there.
06:33Super interesting.
06:33Lori, when we spoke briefly, when we were introduced maybe
06:38last week, you talked about, inside Salesforce,
06:43part of what you're focused on or looking at or talking
06:46through as it relates to agentic AI
06:50is how to disrupt careers inside your company.
06:55Now, I don't know, maybe because everything
06:58scares me as a journalist, that could be scary,
07:02or it could be really exciting, depending
07:04on the level of the employee, what their role is,
07:07how disruptive AI could be to it.
07:09Can you talk through what you mean by that,
07:12and what the discussions are going on inside Salesforce
07:16today on that topic?
07:17Yeah, and the conversation sounds scary,
07:20but as a group of employees, so I
07:22own the technical team, the architects and the solutions
07:25engineers, who generally like to geek out on this stuff anyway.
07:28But our CEO, Mark Benioff, has asked each of us
07:32in the divisions to say, what would you
07:34do if you could let agents loose to work side
07:37by side with your team?
07:38What work would they do?
07:40Let's experiment with it.
07:41We've run competitions.
07:43I know my team, in particular, they've
07:45taken things like, we build narratives
07:47when we produce a demo for our customer to say,
07:51here's what this technology could do for you.
07:53We have to build a narrative that goes with that.
07:55Sometimes that can be a 45-minute long demo,
07:58and that narrative production can be
08:00a big lift for any individual.
08:03We'll take workloads like that and put it to an agent
08:07and let that work with us as we build the actual technology
08:12side by side.
08:13Wow, and so at least for your organization,
08:16there's more, you don't have to, there's
08:19not a lot of convincing that agents are working with us
08:24and not replacing us.
08:25Because that discussion is, I think
08:28it's a fair discussion to have.
08:30It is, it is.
08:31But again, we're bringing this technology to our customers,
08:34so we need to be able to use it ourselves.
08:37And it's making our lives easier.
08:39I know we talked early on about ways
08:42that we're using the technology.
08:44So I run a go-to-market organization.
08:46One of the things that we need to do
08:47is first unleash all the skills in the organization.
08:50So imagine having a skills LLM.
08:53I've got thousands of people that
08:54have done incredible work over their last 20 or 30
08:57years in the tech industry.
08:59How do I encapsulate everything they've done
09:02and then make that available to all of Salesforce,
09:05never mind to our customers?
09:07So I can create a skills LLM around that.
09:09Just being able to use these agents
09:12against different workloads or use cases like this
09:15and experiment with it is really fun for everyone.
09:18So you're taking, when you say skills LLM,
09:20these are skills that individual employees have that
09:24are now being fed into, can you explain that a little more?
09:26So if you played with Notebook LLM at all,
09:29it's probably the best example.
09:31You can take any topic.
09:32So imagine dissertations that you've written,
09:36books that you've written, blogs, articles,
09:39and work that you've done, and putting that into Notebook LLM.
09:42And now all of a sudden, we can ask questions.
09:44But imagine 20 of your colleagues
09:46who have done different pieces of work.
09:48Now you can ask questions of that notebook
09:50and try to understand, well, who would I
09:52talk to if I want to know more about Walmart versus Amazon?
09:57She's plugging my book right now, by the way.
10:00Winner sells all, Amazon, Walmart,
10:02and the battle for our wallets.
10:05Look at you right there.
10:06Wow, I did not pay him in anything other than Bitcoin.
10:11No, but so that's sort of what you're
10:14working through right now.
10:16Sagar, at OpenTable, you talked about the restaurant side.
10:19Can you talk about any use cases of agents internally now
10:24and sort of how autonomous these agents are
10:28and how sort of you think about human working with agents
10:32versus, I won't say being replaced by,
10:35but having tasks taken over fully by them?
10:39Sure, so we are using agents in our customer service chat.
10:46Now, this is something that we've
10:48done in collaboration with Salesforce.
10:51Thank you, Lori.
10:52It's a good tool.
10:53So that replaced a traditional rules-based chat system
10:58where you have to kind of go through the system
11:01and then eventually get to humans.
11:02Now you can actually upgrade the chat to an AI agent
11:04and try to answer the question.
11:06And I think that for me, it's like, well,
11:08the customer service team is now able to focus on harder tasks,
11:15on doing things that are more proactive for our customer
11:18base, right?
11:19They might call a restaurant and say, hey,
11:21I noticed there's some problem in how
11:22you've set up your reservations in your books,
11:25and proactively reach out and solve problems,
11:29rather than answering questions that, again, the AI can do.
11:37I run a technology team, so obviously I'm
11:39always thinking about how do we use it for coding
11:42and making our engineers more productive.
11:44And I tell my engineers, this is not
11:47going to necessarily replace you,
11:50but if you don't get really familiar with it,
11:53somebody who's better at using Copilot will, not necessarily,
11:58might replace you, right?
11:59So it's not going to replace you,
12:01but the person who uses AI might, right?
12:03And so it behooves you to get really comfortable and familiar
12:06with these tools, to learn, to experiment,
12:09to share knowledge with each other on how
12:10to really get good at this.
12:13And today, a lot of the coding tools
12:16are more centered around code completion
12:19and being a very good assistant.
12:24I think over time, more and more tools
12:27will come out that allow you to, say, assign it, even a ticket,
12:32and say, complete this end-to-end, run it,
12:35run some tests against it, and show me the result.
12:39And there, the trained software engineer
12:41will still need to be there to shepherd it
12:43through to production.
12:45But there's going to be more and more done autonomously
12:50over time, which aren't there yet.
12:53I will take questions throughout if we
12:55have some along the way.
12:57So don't be shy.
12:59Raise a hand, and I'll try to, I have a pretty good view here.
13:04That's not a hand here, right?
13:05OK.
13:06Teddy, in an organization like yours,
13:08which technology-focused in some ways,
13:11but maybe not how Silicon Valley thinks about technology
13:15focused, how do you think about prioritizing
13:18during this wave of inundation?
13:20I know as a journalist myself, it's
13:22hard at times to wade through what's real and what's not.
13:28And I think in a large part, we know the revolution is here,
13:33but there's still a lot of piecing through what's
13:36real and what's not.
13:37So for an organization like yours,
13:41that seems like that'd be a lot of responsibility
13:43on your shoulders to prioritize.
13:45Yeah, there's a lot of transformation
13:47both on the digital side and the AI side, right?
13:49And AI in the broader sense, including practical AI.
13:53So one of the things that's really worked for us
13:55is almost like distributing technology teams.
13:59So instead of having the centralized technology
14:02team where all the requests come through,
14:04and then we try to figure out where's the waterline
14:06and how do we prioritize number one versus 100,
14:09we've actually broken up our teams into local teams,
14:12and then they're assigned to a sales and marketing
14:15organization for a given division, or an operations
14:17team, or the HR department.
14:20And that team, there's several AI teams, actually.
14:25And those AI teams will be consisting
14:26of data engineers, data scientists, and visualization
14:29folks.
14:30And they sit every day with the folks
14:32that they are supporting and working with.
14:35And as they get closer and more embedded
14:37with those functional users, the use cases come to them,
14:42and they together will prioritize what
14:44comes to the top of the list.
14:45So somebody in sales and marketing
14:47might make the AI agent the number one
14:50thing the team works on, while they also maintain what's
14:53in production at the moment.
14:54So they also know where all the things that don't work so well
14:57and why they have issues every day and what to improve.
15:00So that's been helping figure out how to prioritize that.
15:03And at the end of the day, my job
15:05ends up becoming, how do we not make sure
15:07that these distributed teams get lost,
15:09that there's still a feel of a technology organization that
15:11comes together?
15:13Now, by doing this, one of the great things
15:15answers some of the questions you're asking earlier about,
15:17is the job going to be replaced?
15:19So a job is a collection of activities.
15:21And the AI agent will do some of those activities.
15:25So now, our teams are educating their counterparts
15:28on, let's figure out what these activities are,
15:31and then engage in new activities
15:34that will help us even reinforce your job as it exists today.
15:38So that's been a great way to get people
15:40to learn AI a bit more, as well as just
15:42get the technologist is learning the business better,
15:46and then the business counterpart
15:48is learning technology better as well.
15:50So you raise a couple of really interesting points there.
15:54One, just to go to this one more time.
15:57So I'm imagining CFO conversation at some point,
16:02and productivity is a favorite word, right?
16:04And I'm just like, part of that productivity conversation
16:08has to be about, yes, changing tasks,
16:11but also, where can we trim?
16:14And are you all telling me, we're
16:17going to see some of that, right?
16:19Or is that not the focus today, and so you're not
16:25thinking about that right now?
16:27I mean, to Teddy's point, I'm finding
16:29that there's more interest in experimentation
16:32about, what else could we do?
16:35What new professions are going to come from this?
16:38What new workloads could we tackle
16:40that we couldn't tackle before?
16:41So at this point, similar to the conversation earlier,
16:45I think there's so much possibility.
16:47People's heads and activities and budgets
16:50are more at the new use cases.
16:53Do you see any new senior roles popping up already,
16:58whether informally or formally inside your organizations
17:02or around the industry?
17:06I sat at a dinner a couple months ago,
17:09and surrounding me on all sides were chief transformation
17:13officers, and their jobs were sitting
17:16at the center of data and AI and the CTO's office, which
17:21I had never, I thought that was relatively new.
17:24I don't know if you guys are seeing this
17:26in your industries as well.
17:29I feel like the jobs are changing.
17:31So maybe we, I don't know if the titles have caught up
17:34to the job changing itself.
17:36But we have traders, for example,
17:38that trade commodities, particularly
17:40in that feed business, buying corn, soybeans, selling them,
17:42hedging, and all those types of things.
17:44And they have a lot of knowledge from years and years
17:47of experience looking at the market.
17:49Well, you can actually develop a model
17:51to do some of that predictive AI,
17:53and think about where the market's going to go,
17:56and then make those adjustments on the positions.
17:59Well, as the modeler is building the actual model
18:05to be used, they actually need input from that trader.
18:08And the trader is learning how to model at the same time.
18:11Because they're the ones that look at and go,
18:12that doesn't make any sense.
18:13Because this and this will happen.
18:15I saw that happen 15 years ago, or whatever.
18:17And so now, what do you call that new trader that actually
18:20knows how to model?
18:21I don't know if we have a name for that, a data
18:24scientist that does trading.
18:25I don't know.
18:26But that is what's emerging.
18:28So I think that's what's really cool.
18:29We just have to figure out where we go with that.
18:31Yeah, just to add to that, one thing I tell my team
18:34is, you don't need a PhD to use most of this, right?
18:39While data science, machine learning,
18:41the core techniques are hard, the APIs that have been
18:45developed are so easy that any engineer can do it.
18:49And obviously, you don't even need an engineering degree
18:51to be able to do some of this stuff, right?
18:53And so we've really pushed and emphasized
18:57that building features on top of AI
18:59isn't a data scientist, or a machine learning,
19:02or research science job anymore.
19:03It is everybody's job to be able to do that in the organization.
19:08Do we have any questions from the audience for this group?
19:12Otherwise, ooh, actually, I'm going
19:15to come back to you in a second.
19:17Right here in the middle, can you just tell us
19:19who you are and who you're with?
19:21Yes, Marian Johnson.
19:22I'm the chief product and technology officer
19:24for Cox Automotive.
19:26One thing that I've been thinking about
19:28that I want to get some of your opinions on
19:30is, as the barriers to entry to develop code,
19:34as you were just saying, speak code into existence,
19:37how is that going to change the value prop of SaaS companies
19:41when the barriers to entry to create code is now just
19:44changed?
19:47Yeah, I mean, one thing that is true
19:52is that there are certain things that still exist.
19:54Network effects, as an example, still exist.
19:57And so for a company like OpenTable,
20:00when we have such a large network of restaurants,
20:03it is still true that diners will
20:05come to us to search for those restaurants
20:07to make reservations to find them.
20:10The other aspect is that data is the thing that
20:14feeds all of these AI systems.
20:16And so if you've been around for a long time
20:18and you have the most data, then you
20:20can build the best AI systems just from a first principles
20:25perspective.
20:26And it is true that it's easier to develop things
20:33in Greenfield.
20:33But if you don't have any users, you don't have any data,
20:36it's hard to necessarily build the smarts and intelligence
20:38on top of it.
20:40Any other thoughts on that one?
20:43I would just add probably cybersecurity dimensions of it,
20:48and then even having legal paperwork in place.
20:51There are things that make it easier
20:53to do business with certain types of companies,
20:56never mind the network effect, that I think will hold true.
21:03Do we have a question here?
21:06Sure.
21:08Yes.
21:09There's a mic right here.
21:12You could tell us who you are, Allie.
21:14Sure.
21:15Hi.
21:15We've met.
21:16Hi.
21:16Allie Garfinkel, senior finance reporter at Fortune Brainstorm
21:19co-chair.
21:21One of the things I find very difficult
21:23about the agentic AI conversations,
21:25it feels like agents can go everywhere.
21:27So I guess my question is, where can agents absolutely not go?
21:34We had this conversation, right?
21:36Jason said, yeah, I have some bad jokes,
21:38but I'm going to leave them at home.
21:40But you said, I don't want it to read a book to my kid.
21:44I don't want the AI, although we do
21:46have Alexa at home that does that sometimes.
21:48So it does happen.
21:49But I do think there are tasks that are, as an example,
21:54I'll do it today.
21:55So we had one of my coworkers, one of my direct reports,
21:58his dog took him to the vet.
22:00It's not looking good.
22:01And I almost used the AI to write the nice note back,
22:05like, oh, I'm so sorry.
22:06He could probably put in better words than I would.
22:08But at the end of the day, I'm like, that's not really,
22:10I should be sending that.
22:11And so at the end of the day, I ended up writing the note.
22:14Probably not as good as what the AI would have done,
22:16but at least it was mine, right?
22:17And so I think those are the areas,
22:19there's going to be this dichotomy of like,
22:21it's going to be better with the AI,
22:24but it still should be us, right?
22:27It's better if it's us.
22:28And I think those are kind of the things
22:30we'll have to struggle with over the next decade or so
22:33to figure out where we stand there.
22:35Yeah, I would say one thing I,
22:37I mean, as a father of two kids,
22:39one thing I worry about is just,
22:41okay, they need to learn how to do these things
22:46that are now possible to do much faster with AI.
22:51Otherwise, will they actually not be able to form
22:54their own like cohesive story
22:56or their own cohesive essay, right?
22:57So that's one thing that I think just personally,
23:01like, keep it away from that education,
23:06the aspect of education where you're starting
23:07to develop original thoughts, right?
23:09Because otherwise you've created somebody
23:13who's somewhat dependent on the AI
23:16to take them to the next spot.
23:17And I think there could be some risks there.
23:20Can you just tell us who you are, please?
23:22I'm Hari, I'm the CTO at GE Healthcare.
23:25Prior to that, I was running AI for Microsoft and Oracle.
23:29Quick question, like as I'm learning
23:33and graduating on the AI space,
23:35so more into the ethical AI stuff and federated learning,
23:41there's a lot of ways to go, right?
23:45And especially in the healthcare,
23:47like I am stuck on innovating because of complaints
23:51and PHI and PII and I can't move fast.
23:55How are you tackling this?
23:57Yeah, so let me see if I can take a crack at that.
23:59So the Crop Advisor LLM that I talked about earlier, right?
24:03It's the book that essentially we've made into an LLM.
24:07And you can ask questions about what to do
24:09about a treatment of a disease
24:11or an insect infestation on a field.
24:14So what could happen
24:15if it gives you the wrong recommendation, right?
24:17Particularly because the wind speed was different that day
24:20and the LLM said one thing and you did another.
24:24So this is where I actually do think
24:26the human side doesn't disappear
24:28because training the model on the book that we had
24:32didn't take very long at all, right?
24:33And it was kind of giving you the,
24:35it's a kind of a retrieval model anyway, right?
24:36So you ask a question, it gives you the answer back.
24:40The harder it becomes in context
24:42and where you are, what's happening,
24:44what are the situations, et cetera.
24:46We spent more time having agronomists,
24:48master agronomists train this model
24:51to make sure and ask it a million different questions
24:53to see if it would like to make it bulletproof
24:55at the end of the day.
24:56And so we've had to have multiple committees
24:58where one was when it first started
25:00and then at the very end before,
25:02we're about to release it now to say,
25:04do we get to the right spot?
25:05So I think that's where the human side comes in.
25:10It's how do you make sure these things
25:11are actually doing what they're supposed to do
25:12and they don't go into places
25:13where they're not supposed to
25:14because they will make connections that don't really exist
25:17or maybe are the wrong connections as well.
25:19I think unfortunately we are all out of time.
25:22If you have more questions for this great group,
25:25you can find them at table one.
25:27Enjoy your dessert.
25:28Thanks so much, everyone.