AI in B2B

0:00:05 The content here is for informational purposes only, should not be taken as legal business
0:00:10 tax or investment advice or be used to evaluate any investment or security and is not directed
0:00:14 at any investors or potential investors in any A16Z fund.
0:00:18 For more details, please see a16z.com/disclosures.
0:00:19 Hi, everyone.
0:00:21 Welcome to the A16Z podcast.
0:00:22 I’m Das Rush.
0:00:27 In this episode, we talk with Ola Griginski, the founder of People AI, a platform for sales
0:00:28 and marketers.
0:00:31 Well, the broader question we’re tackling is beyond just sales and marketing, to really
0:00:36 how AI is taking data from across the enterprise to change how we work, regardless of our
0:00:37 function or industry.
0:00:42 Also joining us in this conversation is Peter Lauten from the A16Z Enterprise investing
0:00:43 team.
0:00:46 We’re going to cover everything from how to design AI to be a co-pilot for knowledge
0:00:51 workers, to the founder’s playbook for spotting the next AI opportunity, and then how to take
0:00:53 a product to market once you do.
0:00:58 First, we begin with why AI and B2B, that is business to business, is so different from
0:01:00 business to consumer.
0:01:05 In the last five years, more and more areas of human existence or activity are adopting
0:01:06 the AI.
0:01:12 And in particular, AI is finally proliferating B2B Enterprise world, which is very different
0:01:15 to work with than your classic B2C scenario.
0:01:17 B2C is much more reactive.
0:01:22 You build something, you throw it at a million users in Australia, and then see what happens.
0:01:24 And if it doesn’t work, you can pull it back.
0:01:28 With B2B, you cannot do it because you have five customers and each one of them is paying
0:01:29 you.
0:01:34 So you better go ahead and do the research in advance and make something enterprise is
0:01:35 one.
0:01:41 And so with that, it’s a very different level of risk, but also a very different level
0:01:45 of data instead of having a large number of users out there that all contribute a little
0:01:47 bit of data to you.
0:01:53 You have very few users that contribute a lot of very valuable, very private, secure
0:01:55 and needle-moving data.
0:02:00 It’s much harder because you have to make those companies very comfortable with the
0:02:06 fact that their data will be anonymized, aggregated, will not be shared with competitors, will
0:02:08 not be leaked if you get hacked.
0:02:13 And one of the big calculations that particularly big customers would be doing is how do I compare
0:02:18 the gains of machine learning being applied this way against the cost of actually doing
0:02:19 it.
0:02:24 And so if you think about the cost of collecting, cleaning and maintaining the data, like that’s
0:02:28 certainly gotten cheaper over time, particularly in the last decade with cloud compute and
0:02:31 AWS, GCP, Azure, etc.
0:02:37 But there are scenarios where the value out of the data you get maybe isn’t so obviously
0:02:39 worth the cost of actually collecting it.
0:02:46 And so when we think about how this generation of technology will get deployed over even
0:02:52 the scale of decades, a lot of it is, where is the greatest yielding application of these
0:02:54 technologies by industry?
0:02:56 And then let’s aim our efforts there.
0:02:58 And then what’s the next best industry?
0:03:03 So it’s also a function of just how valuable is the promise you’re making to customers
0:03:06 against their cost of being able to achieve it.
0:03:07 Interesting.
0:03:12 That’s how companies are evaluating AI, but what about on the startup side?
0:03:15 How should an entrepreneur look at the market for AI?
0:03:20 For aspiring entrepreneurs, there is a pattern that if there is an area of human activity
0:03:24 or existence that generates activity data of how humans do something and the data is
0:03:28 not being captured today, there is a ripe opportunity.
0:03:34 There is a time series activity stream, someone doing something, moving shipping containers,
0:03:40 ordering readings from a wind farm, recording emails from a salesperson, recording location
0:03:42 of an Uber driver.
0:03:48 And as long as you can collect reliable, comprehensive, non-manually entered high-volume activity
0:03:54 data from many wind farms or from many containers and ships or from many salespeople in one
0:03:59 place, aggregate and start seeing the big picture and then use that big picture to analyze
0:04:04 the micro trends and predict what’s the best next action for the operator, the salesperson,
0:04:10 the Uber driver is, that’s where it brings in network effects and significant acceleration
0:04:11 of growth.
0:04:17 So my theory is that every industry is going to switch into this business model of collecting
0:04:21 activity data, understanding its scale and turning into best next actions in the next
0:04:23 five, ten years.
0:04:26 You mentioned network effects there as an accelerator of growth.
0:04:29 What exactly do you mean by those network effects?
0:04:35 The more sensors, edge computing devices, salespeople, Uber drivers you have in the
0:04:40 network, the more data you collect, the more the patterns of behavior you see.
0:04:45 When you put them all into one centralized graph, the smarter this graph becomes for
0:04:52 everybody, the better the predictions it can produce about the best next actions.
0:04:55 Now this is where the second loop of the network effect starts.
0:05:00 The better the predictions are, the more money Uber drivers make, the more money salespeople
0:05:02 make, the less wind farms break.
0:05:08 So more and more players will be lured in with this better efficiency to become part
0:05:11 of the network, contribute even more data into this shared graph.
0:05:13 So it becomes kind of this virtuous cycle.
0:05:14 Exactly.
0:05:18 Imagine you’re a wind farm operator and you know that your wind farms are going to break
0:05:23 after they broke and there’s a wind farm next to you, same wind, same everything, but they
0:05:27 come in and practically fix them and have zero downtime because they start collecting
0:05:30 the data about their wind farms breaking five years before you.
0:05:34 If your competitor automated the business process and you did not, you are at disadvantage
0:05:38 but you can go to the same vendor and buy the same assembly line and catch up.
0:05:42 If you miss the AI boat, the results are very different.
0:05:47 If you did not collect the data early enough, there’s nothing you can do to make your AI
0:05:49 better than your competitors.
0:05:54 Even if the data is shared, the AI has been trained on your peculiar behavior and the
0:05:56 market responds to it for three years longer.
0:05:59 It’s seen more examples, more samples, it’s just smarter.
0:06:07 And so unlike automation, like industrial revolution from 100 years ago, AI arms race is a zero
0:06:08 sum game.
0:06:13 So my prediction is that 10 years from now, Fortune 500 will look very different than now
0:06:19 because some companies did not get into collecting the data and training their machinery models
0:06:20 early enough.
0:06:25 Is there a tipping point where this balance of power around the data shifts from the customers
0:06:31 you’re serving it up to, to the vendors because your AI, your model becomes such a competitive
0:06:34 advantage that they’re going to have to play by your rules.
0:06:38 They’re going to have to allow you access to that data and to reuse that data for training.
0:06:41 The competitive dimension part of that is really interesting.
0:06:46 In theory, if one customer is contributing their data to that corpus, certainly their
0:06:51 data is anonymized and no one else in the customer base would really have access to their data.
0:06:56 But in theory, anything you observe at one customer is informing generalized models that
0:06:58 all the customers benefit from.
0:07:03 And so I think part of it is just being super honest in the sales process that yes, of course,
0:07:04 that’s how it works.
0:07:08 But I think the calculation that most of the customers are doing is, okay, that’s fine.
0:07:13 I realize I’m helping everyone else out a little bit, but you’re helping me so much
0:07:17 that I’m willing to contribute a little bit of the signal from my data to help everyone
0:07:18 else.
0:07:22 It’s kind of like you push the boulder uphill getting first customers, but then you roll
0:07:27 downhill with the network effect becomes so strong and the value of joint data set, joint
0:07:30 knowledge becomes so big that people just go ahead with it.
0:07:35 And so one reaction to how nervous customers can be around, “Oh, I don’t know if I want
0:07:41 other people to benefit from my data at all,” is that’s kind of analogous to people writing
0:07:46 contracts that say, “You can only exclusively sell your software to me,” because in abstract,
0:07:49 these are just tools and technologies that everyone’s going to get access to eventually.
0:07:55 And so any hesitation I think customers have about it, that’s going to be a relic years
0:07:56 from now.
0:08:00 One trend, and actually I think I heard it from Marc Andreessen here, is that when data
0:08:03 model changes, systems of record die.
0:08:10 So there were systems of record like CRMs built on, not relational, but on hierarchical
0:08:12 databases in the 70s.
0:08:13 Do you know of any today?
0:08:15 They don’t exist anymore.
0:08:19 Then the next data model that happened was on-prem SQL, C-Belt.
0:08:21 Then the cloud SQL happened.
0:08:23 And now we’re talking about Salesforce.
0:08:28 The next generation of data model is likely to be a graph, which is the data model that
0:08:35 allows you to train AI in the best way possible, federated shared graph of data.
0:08:40 Together with data model shifts, the way you consume the software has been changing.
0:08:47 Instead of you pulling data and looking for it in bunch of Excel reports or Salesforce
0:08:53 reports or websites, it’s actually being pushed to you in a pre-packaged, we call it personalized
0:08:58 actionable insights way, where everything you need to know to complete this action is
0:09:01 right here pushed to you through the channel through which you’re most likely to engage
0:09:02 with it.
0:09:08 Most of the systems of the future will have this feed or minimum choice, a maximum focus
0:09:10 on one thing in front of you.
0:09:13 And once you complete it, the next thing will come and then the next thing will come.
0:09:17 And that will make you much more productive and much more focused at what you’re doing.
0:09:20 So you mentioned that the way we consume software is changing.
0:09:21 Say a little bit more about that.
0:09:24 What does that change mean for product design?
0:09:26 It’s not the case that the product is simpler.
0:09:31 It’s that the UX is designed such that there’s a level of intent that you can observe from
0:09:32 the user.
0:09:35 You should generally know what they’re trying to do such that you don’t have to expose the
0:09:39 entire complexity of a product to the user and overwhelm them.
0:09:42 You just say, we’re quite certain this is what you actually want to do at this stage
0:09:43 of the product.
0:09:44 That’s all we’re going to show you.
0:09:47 And you’re going to be able to very efficiently execute that function.
0:09:52 It sounds like the prediction here is this next wave of AI companies, this next generation
0:09:58 of enterprise software, it starts to make worker lives easier because it gets simpler.
0:10:00 And give you the last tabs in Chrome.
0:10:01 Totally.
0:10:07 And part of that is just the interaction between how do you map UX and design against this
0:10:11 like increasing orchestration of the knowledge worker labor force?
0:10:14 So what does that mean then for kind of the enterprise worker of the future?
0:10:18 It seems to me like in that scenario, you’re getting workers to be increasingly focused.
0:10:21 Does that mean they’re also getting increasingly specialized?
0:10:22 I think it’s two things.
0:10:29 They’re able to focus on what really needs human judgment, really complex rare situations
0:10:33 that maybe there’s not a sufficiently big data corpus to make an interesting inference
0:10:35 on with a machine learning application.
0:10:39 And so what you get is humans can not only do the hardest IQ work that compute can’t
0:10:43 really replicate yet, but also they do become more empathetic.
0:10:47 Your accountant spends more time thinking about your specific needs and how they communicate
0:10:50 with you versus them doing simple math.
0:10:53 There are two modes in which AI operates with people.
0:10:56 We call one autopilot and another one is copilot.
0:11:01 So autopilot is when you are doing as a human, you’re working on something mundane, some
0:11:04 repetitive low value at task.
0:11:07 And so AI can easily automate that for you.
0:11:12 So think of our Uber analogy of them receiving a phone call, typing in the computer, picking
0:11:15 up the ham radio, delivering it to you.
0:11:20 That’s just not the best use of all those billions of neurons in your brain.
0:11:24 So AI comes in and automates those functions, freeing up your time to do something much
0:11:26 more productive and much more effective.
0:11:31 The second bucket in which we see AI play in the role is what we call copilot.
0:11:38 In copilot cases, AI is augmenting your ability to make decisions.
0:11:42 Think of you are in a race car and there’s someone next to you that knows what’s around
0:11:47 the curve and is whispering in your ear what to be ready for, what’s the best next thing
0:11:48 you could do.
0:11:54 While autopilot makes us super productive, copilot allows us to focus on more human,
0:12:00 more face to face, more EQ driven things that machines will not be able to do for a long,
0:12:01 long time.
0:12:07 Oleg, I’m curious, a lot of these companies that have some sort of copilot goal, there’s
0:12:12 some sensitivity around the employees and the users not wanting to be tracked.
0:12:15 For some people who didn’t grow up on the internet, they say, oh, this is a violation
0:12:16 of my privacy.
0:12:20 I don’t want Oleg to know the magic of how I deliver my sales.
0:12:24 To be honest, if I have some really good weeks and I have some bad weeks, I don’t want them
0:12:26 to know about my bad weeks.
0:12:28 How do you get the end user comfortable?
0:12:30 Record the 10x rule.
0:12:38 You have to be visually in a very simple, explainable way, promising and delivering 10x the value
0:12:42 of being on the system than being off the system.
0:12:47 Now, to simplify that, we go to our customers, we say, great, let’s not put you on the platform,
0:12:51 let’s A/B test and see what happens in another quarter when the guy next to you or the girl
0:12:58 next to you did 10x your results and the only variable we have is you were afraid of being
0:13:00 a user of a modern system.
0:13:02 So break that down for me a little bit more specifically.
0:13:05 You’re making this technology for sales and marketers.
0:13:08 What are sales and marketers doing today that they’re not going to have to do five years
0:13:12 from now and how is AI going to change the way that they work?
0:13:17 Is this just about their efficiency or is it going to fundamentally shift how those industries
0:13:18 work?
0:13:20 It will definitely shift how the industries work.
0:13:22 So let’s think about the day of a salesperson.
0:13:28 In an average week, salespeople spend about a third of their time on manual data entry.
0:13:35 Another third of their time is supposed to be spent on prospecting, so finding who else
0:13:41 looks like my current customers, which from computer science perspective is a very classic
0:13:45 lookalike problem we’ve solved 30, 40 years ago and then only a third of that time is
0:13:51 spending face-to-face meetings actually selling, actually building relationships.
0:13:55 Now the first two parts, the manual data entry, it should be gone.
0:13:58 It should be fully automated through autopilot capability.
0:14:04 The second part is who should you talk to next is also very easy to solve with the help
0:14:11 of lookalike modeling, machine learning, pointing you to very high value ad prospects
0:14:15 of customers and actually helping you automate the outreach to them.
0:14:21 And then a third part, which is being face-to-face with customers, truly building relationships
0:14:23 is where machines cannot replace humans.
0:14:29 So five, 10 years from now, salespeople or any white collar kind of knowledge workers
0:14:37 will be actually focused on EQ-driven relationship-building activities still with copilots help because
0:14:40 by then copilot will recommend how to better build a relationship.
0:14:46 So you can think of, in particular, the sales org of the future is really high EQ people
0:14:51 showing up and then compute is aiming them at where their labor is best suited.
0:14:54 I mean, this all sounds really fascinating, but then I think about the older generation
0:14:58 of Salesforce out there being told they have to use this new tool.
0:15:03 We’ve been going about their methods of selling and they’ve got their kind of day-to-day.
0:15:07 How do you go about with that habit change for those sorts of workers?
0:15:12 Or is this a generational shift in terms of how AI technologies get adopted in the enterprise?
0:15:18 If you think about it, the salespeople themselves, they are ingrained in their habits because
0:15:21 something they’ve done over and over again worked.
0:15:26 What autopilot does, it frees up a bunch of their time to experiment more.
0:15:32 And so if we approach any problem that involves AI with autopilot freeing up your time and
0:15:38 copilot teaching you the new ways, eventually you will be able to retrain the people who
0:15:42 are ingrained in their old ways into new ways by showing them that trying new stuff actually
0:15:44 yields results.
0:15:49 This training at the end point exactly where the knowledge worker is doing their job, we’ll
0:15:51 say these are the next top priorities for you.
0:15:55 This human can learn from that and they start to understand, okay, this is the best way to
0:15:56 sell.
0:16:00 The way that this is playing out is it’s kind of unbundling these horizontal, like learning
0:16:03 management systems of two decades ago.
0:16:09 This kind of dance between the knowledge worker and the software, the knowledge worker and
0:16:12 the machine learning algorithms trains them live on the job.
0:16:16 You learn in the software while you do your job.
0:16:20 We hear a ton about we have to retrain the workforce, but this is, I think, the first
0:16:24 time I’ve heard it articulated that some of that retraining gets built into the products
0:16:30 especially at the point where it sounds like the products are augmenting what humans do.
0:16:32 How are products going to look different?
0:16:34 What does the product of tomorrow look like for me?
0:16:40 In the next 15 years, every area of our existence will probably go through a transition and
0:16:45 will be, instead of being inundated by a bunch of Chrome tabs, it’ll be inundated by a bunch
0:16:47 of feeds that tell us what to do.
0:16:53 I think there are cycles of bundling and unbundling playing out here already in the latest wave
0:16:55 of machine learning applications.
0:16:59 The use cases are getting increasingly precise and tailored to the end user.
0:17:03 One of the consequences of that is as these systems get richer and deliver more promises
0:17:08 to the user and then help them do their job in more and more ways, I think you just get
0:17:11 much fatter workflow applications.
0:17:13 We see that already happening in B2C.
0:17:17 The bundling of services as they figure out the right workflow.
0:17:18 Think of Uber.
0:17:19 It used to be just rides.
0:17:20 Now it’s food delivery.
0:17:22 Now it’s freight and stuff like that.
0:17:23 It’s coming from the same app.
0:17:27 You can ship something, you can get somewhere, or you can get food delivered or whatever
0:17:29 you want to have delivered.
0:17:36 Once the company finds an optimal user interface that allows for suggesting best next action,
0:17:41 then it just makes sense to bundle in more and more functionality and take over more and
0:17:43 more of the attention span of the user.
0:17:46 I think this actually goes to a question that I’m really curious about, too, which is the
0:17:49 broader trends of how these products are coming into the market.
0:17:54 Are you finding that you have to really drive that adoption at the enterprise through your
0:17:58 user first, or is it still more of a traditional top down?
0:18:02 You’re going in, you’re selling to somebody because they want to automate what their sales
0:18:03 force is doing.
0:18:07 So that has been an interesting point for us because we want to build a bottoms up approach.
0:18:08 Right now we’re top down.
0:18:11 We come in through kind of standard procurement channels.
0:18:17 The part that is tricky there is that you have to make the organization comfortable with
0:18:20 your security posture, your privacy posture.
0:18:26 And so having that is what slows down the bottoms up approach, you cannot give someone
0:18:30 an AI that’s going to learn from individual user without checking in with the company
0:18:31 first.
0:18:35 It’ll be a very tricky balance played out over the next few years where the users will
0:18:40 be demanding more and more powerful and data hungry products, while the enterprise will
0:18:43 be saying, well, I still want to be in control of it.
0:18:47 I think there’s a distinction between the types of enterprise apps.
0:18:55 The apps that build AI based on user behavior and that do not require merging of that behavior
0:19:02 data with proprietary enterprise data such as Zoom or Slack, those will have much easier
0:19:06 time with bottoms up adoption because they just need the user to log in and do stuff.
0:19:10 It’s a workflow or a utility tool that a single player can use in a single player mode.
0:19:15 The flip side is when the value you create is based on or is significantly amplified
0:19:20 by the historical proprietary knowledge of the company.
0:19:25 So in our case, at people AI, we need to know what’s in your CRM in order to not give you
0:19:27 random suggestions.
0:19:33 And so having access to this proprietary information that the company has under the lock by the
0:19:39 security IT is where you have to go and be very transparent and open with the kind of
0:19:46 top down IT security infosec teams and work with them to get access, but also work with
0:19:49 them to explain to them the value that the end users are going to get.
0:19:53 You’ve both been working around these AI technologies for a few years now.
0:19:57 I know, Oleg, you’ve gone through a few different companies and iterations, getting to people
0:19:58 AI.
0:20:01 What have been the biggest surprises for you in this space and what have been some of
0:20:03 the biggest lessons learned?
0:20:07 So the first company I joined, I started my career as an inside sales person back in
0:20:08 2006.
0:20:12 I was pounding the phones before LinkedIn existed and before Twitter existed, Smile
0:20:13 and Dial Style.
0:20:21 The company that I was at went out on Toronto Stock Exchange right before the downturn.
0:20:23 So the first lesson I learned was timing is everything.
0:20:27 And if you get an opportunity to move forward aggressively.
0:20:32 Second lesson I learned was with Cementria, the company I started in 2011.
0:20:36 That’s where I realized why the size of the market really matters.
0:20:43 We started a sentiment analysis API, beautiful technology, scales nicely.
0:20:47 The only problem is there was only 20, 30 companies that really needed it in the world.
0:20:53 So Cementria ended up being a very small market and we ended up owning probably 80% of the
0:20:59 market within three years while still doing single millions of revenue, which was kind
0:21:00 of crazy.
0:21:03 It’s really striking that you mentioned you’ve been an inside sales, you’ve done the Smile
0:21:05 and Dial, you’ve been a marketer.
0:21:09 How much did your personal experiences inform your product development?
0:21:10 Oh, 100%.
0:21:16 I’ll never forget the moment in 2007 when CEO of the company, where I was actually at
0:21:20 that point leading inside sales team, grounded me and my whole team in a room in a pretty
0:21:27 small sweaty conference room for a week, having us go and clean Salesforce record by record.
0:21:32 And then a week later, we came out and we used the new clean, amazing Salesforce for
0:21:33 a week.
0:21:37 And by the end of the month, it was just as bad as it was before.
0:21:41 And so that was one of the first formative experience where I knew something is really
0:21:45 wrong with how we do sales and marketing today.
0:21:48 And then I had to run my own sales team.
0:21:51 And when I was running my own sales team, I could not get the data.
0:21:52 I like data.
0:21:53 I understand it.
0:21:57 I could not understand what’s going on in the sales team and my people were working really
0:22:03 hard, but I could not pinpoint why it’s not working out, why we are losing deals, why
0:22:09 we are not ramping quickly, why we need to hire more people while seemingly our productivity
0:22:12 seems to be fine, but we are not doing enough.
0:22:16 And so all these why questions that you’re supposed to have the data in the CRM, but
0:22:23 you never actually have it led me to really start thinking about starting people AI.
0:22:27 I love that the way that like the product development has been informed by your personal
0:22:28 experiences.
0:22:32 I think that’s a beautiful tie-in to just how you can have all the data in the world,
0:22:37 but there’s still something about human experience, human empathy, that you really can’t replace
0:22:39 with a computer or a data set.
0:22:44 I get to spend almost every single day of my life meeting founders on the frontier of
0:22:47 building products for knowledge workers.
0:22:52 And there’s a huge disconnect between their optimism for these people against what you
0:22:57 hear in the policy realm and all this doomsday thinking.
0:23:02 This whole wave of automation and augmentation that is making people pretty nervous at the
0:23:07 macro picture, I see all this micro level evidence that it’s just really transforming
0:23:12 the workforce and people really get more meaning out of their jobs and they really start to
0:23:15 love that they can focus on exactly what they’re good at.
0:23:21 You’re really good at something, like you want to be the best musician, you want to
0:23:24 be the best artist, the best computer scientist.
0:23:28 That’s what deep inside we’re all striving towards.
0:23:35 And that self-mastery is what AI, the co-pilot specifically, helps you amplify.
0:23:40 The autopilot takes away the time that you spend on stuff that you don’t care about on
0:23:42 your way to mastery.
0:23:47 And then the co-pilot actually gives you guidance on how to become better, more effective,
0:23:51 learn faster on becoming a master at what you do.
0:23:56 So I think the side effect or the primary effect of AI eventually is going to be that
0:24:01 a lot more people will be insanely good at that specific thing.
0:24:02 Great.
0:24:04 I think that’s a fantastic note to end on.
0:24:07 So I want to thank you Oleg for coming in and thank you Peter.
0:24:08 Thanks for having me here.
0:24:09 It’s been a lot of fun.
0:24:10 Yeah, thanks to us.