AI transcript
0:00:02 People use the term AI and they’re like,
0:00:03 “Oh, everything’s going to change something.”
0:00:05 What does that mean?
0:00:07 Nobody wants to do data entry.
0:00:10 Nobody wants to sit in the back and read 100 facts
0:00:13 and try to input that into a system.
0:00:15 This is one of the most exciting times
0:00:18 to actually be going after some of these legacy markets.
0:00:20 There’s just so much untapped opportunity
0:00:22 that technology just wasn’t able to penetrate before.
0:00:24 Then now, with intelligent AI agents,
0:00:28 with voice agents, et cetera, you can now tackle.
0:00:31 The internet has given us so much,
0:00:34 whether it’s instant access to the world’s information,
0:00:36 platforms for new art and expression,
0:00:39 nearly every feature film or song in your pocket,
0:00:42 or the ability to work from just about anywhere
0:00:44 with a Wi-Fi connection.
0:00:46 The list, of course, goes on,
0:00:48 but there are still some promises
0:00:51 that haven’t quite hit their mark yet.
0:00:54 One example, robotic process automation,
0:00:55 otherwise known as RPA,
0:00:58 has been a buzzword for years.
0:01:00 A buzzword that promised a revolution
0:01:03 in automating repetitive tasks.
0:01:06 Companies like UiPath found it in 2005,
0:01:08 even promised to enable the, quote,
0:01:10 “fully automated enterprise.”
0:01:12 And while RPA did start strong,
0:01:16 its limitations became evident due to its rigid systems,
0:01:18 sometimes even hard-coding specific processes
0:01:19 down to the button click,
0:01:23 meeting with the unpredictability of most workflows.
0:01:27 Of course, UiPath, with a market cap of $7 billion
0:01:28 as of this recording,
0:01:31 and the RPA industry at large are no failure,
0:01:34 but they are poised for a facelift.
0:01:36 Enter intelligent automation,
0:01:38 a new paradigm powered by AI
0:01:41 that can handle messy, unstructured workflows
0:01:44 that exist in just about every organization,
0:01:45 precisely the types of tasks
0:01:48 that RPA previously could not handle.
0:01:50 So in today’s episode,
0:01:52 we discuss RPA’s second leg,
0:01:55 together with A16Z partner, Kimberly Tan.
0:01:58 Based on her viral article, “RIP to RPA,”
0:02:00 the rise of intelligent automation,
0:02:03 which, of course, will link in the show notes.
0:02:06 So what specifically about AI makes this possible today?
0:02:08 How should startups be thinking about this opportunity?
0:02:13 And why is this opportunity so much larger than the last era?
0:02:14 But before we begin,
0:02:17 I want you to think about the most annoying
0:02:19 and repetitive thing that you need to do.
0:02:22 Now, imagine that task automated.
0:02:25 That future may not actually be so far away.
0:02:28 So listen in to see how you can get involved.
0:02:31 As a reminder,
0:02:33 the content here is for informational purposes only.
0:02:37 Should not be taken as legal, business, tax, or investment advice,
0:02:39 or be used to evaluate any investment or security,
0:02:42 and is not directed at any investors or potential investors
0:02:43 in any A16Z fund.
0:02:46 Please note that A16Z and its affiliates
0:02:47 may also maintain investments
0:02:49 in the companies discussed in this podcast.
0:02:52 For more details, including a link to our investments,
0:02:55 please see a16z.com/disclosures.
0:03:03 Can really you wrote an article with a pretty fun title,
0:03:05 RIP to RPA?
0:03:06 So let’s jump into that.
0:03:09 But first, what is RPA?
0:03:12 RPA stands for Robotic Process Automation.
0:03:16 And it’s a way of basically automating very manual tasks
0:03:17 within an organization.
0:03:21 So things like data entry or invoice processing
0:03:24 that basically every business has to do,
0:03:25 but it’s nobody’s core competency.
0:03:29 Just one of the dirty, messy, internal things within an organization.
0:03:31 So historically, it’s been done very manually.
0:03:33 Like you would just hire a data analyst
0:03:36 or you would hire a back office operations person.
0:03:39 And there was this, I would say, innovation in the last 20 years
0:03:40 where people were like,
0:03:42 “Is it possible to automate these tasks?”
0:03:44 And so the historical way people have done it
0:03:46 is through Robotic Process Automation
0:03:50 where you basically build like a little software bot
0:03:53 that mimics the actual clicks that somebody would be doing.
0:03:54 It’s very deterministic,
0:03:57 meaning they’re literally clicking the different boxes
0:03:59 that I would be clicking as a human.
0:04:01 But organizations are messy
0:04:04 and the work we actually have to do is not perfectly delineated
0:04:06 by a very specific process.
0:04:08 So oftentimes, if something veers a little bit off course,
0:04:10 like maybe someone misspelled a name
0:04:12 or maybe a website changed
0:04:15 where the sign-in box physically is on a page,
0:04:18 then historically that would break the RPA process.
0:04:19 And as you can imagine,
0:04:22 there’s an infinite number of small little things
0:04:23 that could happen like that.
0:04:27 So RPA is often very good for doing 80% of the task,
0:04:29 but then like 20% of the time that it fails,
0:04:31 it’s still a manual person who has to come in.
0:04:35 So it’s just not reliable enough to actually do the full task.
0:04:38 And so you’re still left with having the back office people
0:04:41 that were the first generation doing these sorts of tasks there.
0:04:43 So I just think like with AI and LLMs now
0:04:46 because they’re able to process such unstructured data
0:04:49 and they’re able to intelligently collect context
0:04:51 and then figure out what the best course of action is.
0:04:53 The next generation of actually automating these back office tasks
0:04:56 should be like intelligent AI agents instead.
0:04:59 What can intelligent automation or what you refer to
0:05:01 as these LLMs in action?
0:05:03 What can they do that RPA couldn’t?
0:05:06 Let’s use the example of a company
0:05:08 that we were actually invested in called Tenor.
0:05:12 Tenor does referral management for healthcare practices.
0:05:14 So if I’m a primary physician
0:05:18 and I need to refer a patient to a specialist,
0:05:20 historically, the way that that would be done
0:05:23 is I would literally write something out on a piece of paper.
0:05:26 I would fax it to the specialist.
0:05:29 The specialist front desk person would take the fax,
0:05:31 look at it, look at all the information on it
0:05:33 and then input it into my own database,
0:05:35 check, you know, like the insurance policies,
0:05:36 check prior history, et cetera,
0:05:38 and then decide whether to accept the patient or not.
0:05:40 And that was a very manual task
0:05:43 that there’s just a little bit too much complexity
0:05:46 in the way that it’s done for RPA to be able to handle.
0:05:48 So it had to be some sort of administrative person,
0:05:50 like human who was gonna do it.
0:05:53 And with now like intelligent automation,
0:05:55 Tenor has come up with a very sleek solution
0:05:57 that is basically able to automate that whole process.
0:05:59 And it’s much more self-serve.
0:06:02 ‘Cause the way that RPA would historically work is
0:06:04 you would have to hire an implementation consultant
0:06:08 and they would sit next to whoever was doing the task
0:06:09 and they would basically just watch
0:06:11 what are the clicks that you are doing.
0:06:14 – Right. – And then program those clicks.
0:06:16 But someone like a Tenor, for example,
0:06:17 you’re not gonna have somebody sitting there
0:06:20 watching what the front office admin is doing.
0:06:23 Rather, they’ve created a really sleek UI
0:06:26 where it looks very much like a drag and drop
0:06:28 different process flows.
0:06:30 And they’re able to create their own automation process,
0:06:32 which to them feels very intuitive
0:06:34 so they can set it up themselves,
0:06:35 but actually has a ton of complexity
0:06:37 under the hood that’s being handled.
0:06:38 – I mean, one natural question that comes up,
0:06:39 I think for many people,
0:06:42 especially as they think about things like hallucinations,
0:06:44 is where is the technology in this arc?
0:06:47 Are we able to really achieve this idea
0:06:48 of intelligent automation today?
0:06:49 Are there barriers?
0:06:51 Like where do we sit in that trajectory?
0:06:52 – The way that we’ve seen it work best
0:06:55 is when there’s one very specific automation flow,
0:06:58 at least to start, that a company can just nail,
0:07:01 meaning it’s often industry specific.
0:07:04 So you can integrate into all the core systems there.
0:07:07 You can understand the context for that industry.
0:07:11 And it’s one very repeated, but very manual flow.
0:07:14 So for example, like data entry, I get on a phone call,
0:07:17 I hear the update on where an order is,
0:07:18 all the information from that order
0:07:21 can be parsed through that call
0:07:22 and input it into my main system.
0:07:25 That probably happens like thousands of times a day
0:07:29 for the large organizations, all manually done.
0:07:31 And that is one very specific flow.
0:07:31 And that’s just to start.
0:07:32 And then once you get there,
0:07:34 you can build deeper into other flows,
0:07:37 but I think that is a much more successful path
0:07:39 where you can actually understand the constraints
0:07:39 and build around them,
0:07:42 make sure that the agent performs correctly
0:07:44 versus tackling, let’s say, like everything within healthcare,
0:07:47 everything within legal and logistics to start.
0:07:48 – Normally I ask the question, why now?
0:07:51 But I feel like listeners know that AI is coming,
0:07:55 it’s here, LLMs are maybe the term that a lot of people use.
0:07:57 But is there a deeper why now
0:08:00 or specific technological advances
0:08:02 within the sphere of LLMs that you can point to
0:08:05 that actually make this possible?
0:08:06 – Yeah, I think one thing that we’re really excited about
0:08:08 is people use the term AI and they’re like,
0:08:10 oh, everything’s gonna change now because of AI.
0:08:11 But like, what does that mean?
0:08:14 There’s a lot of very distinct technological breakthroughs
0:08:16 that make different applications possible.
0:08:18 And specific to intelligent automation,
0:08:21 I think one of the things that makes it much more possible
0:08:23 than before is a lot of the fundamental research
0:08:25 coming out of the large labs.
0:08:26 So for example, recently,
0:08:29 Anthropic announced computer use,
0:08:31 which is basically a browser agent
0:08:33 that is able to intelligently understand
0:08:37 what is happening on the browser level of any sort of desktop
0:08:39 and be able to take actions accordingly.
0:08:41 So we talked about how historically,
0:08:45 RPA basically understood at a pixel level,
0:08:47 hey, I should click this thing and then I should click that.
0:08:49 But with something like computer use,
0:08:51 or I think OpenAI has something called operator
0:08:53 that they’re gonna release soon,
0:08:55 agents are gonna be able to browse the internet
0:08:58 and browse the web in a much more sophisticated way,
0:09:00 which is gonna open up a lot of possibilities
0:09:02 for what intelligent agents can do before.
0:09:05 So we think a lot of these intelligent automation startups,
0:09:07 they’re not gonna be doing fundamental research on their own.
0:09:09 There’s still tech that needs to be done
0:09:12 to make a browser agent fully work at scale.
0:09:13 But what’s really exciting for us
0:09:15 is that the large labs are clearly working on this
0:09:17 and clearly understand the opportunity.
0:09:19 And so as that tech gets better,
0:09:21 we think there’s gonna be a whole world of startups
0:09:23 who are able to leverage it
0:09:25 for all the different industries out there
0:09:28 that the large labs themselves are not gonna tackle.
0:09:30 And as you think about the opportunity,
0:09:31 you framed it in your article
0:09:34 as two different paths that people might take.
0:09:36 So one of them was the horizontal AI enabler
0:09:39 and the other was a vertical automation solution.
0:09:41 So tell us about that, the two different paths
0:09:43 that you see if people want to build in this space.
0:09:45 So the first is the horizontal AI enabler
0:09:47 and that’s something that we think
0:09:50 any company who’s doing any sort of automation,
0:09:52 intelligent automation, is going to have to do.
0:09:53 One very common example,
0:09:55 which I’ve touched upon a little bit already,
0:09:56 is data extraction.
0:10:00 Almost every intelligent automation path
0:10:03 starts with some messy unstructured data
0:10:06 that you need to pull key outputs from.
0:10:09 And today, a lot of people are just building that manually.
0:10:11 But we’ve started to see companies emerge
0:10:13 that are purely doing that path,
0:10:15 which is taking unstructured data
0:10:16 and pulling out the key pieces
0:10:18 to turn it into structured data.
0:10:21 And we think that could be one really interesting opportunity.
0:10:23 So anyone who is either building
0:10:26 their own automation in-house can leverage that
0:10:27 as a key component.
0:10:29 Or if you’re building like a full end-to-end solution,
0:10:32 maybe you input that as one of your components as well.
0:10:33 One thing that I’m personally really excited about
0:10:35 is the vertical automation path.
0:10:39 I think to make an intelligent AI agent very successful,
0:10:41 it often is helpful in the beginning
0:10:44 to have it be in a very constrained domain.
0:10:46 For example, in logistics or in healthcare and legal,
0:10:49 it is a domain that they can understand all the context for.
0:10:52 They have all the necessary inputs, integrations, etc.
0:10:54 And they’re able to automate one specific flow.
0:10:56 So what we’re really excited about there is,
0:10:59 let’s take an industry that does have a lot of manual work
0:11:03 that needs to be done, like a very large back office.
0:11:04 If you think about what are the things there
0:11:07 that actually have to be automated
0:11:10 that maybe RPA wasn’t able to tackle before
0:11:14 because it just wasn’t like a large enough individual customer.
0:11:16 Like it wasn’t one of the Fortune 500 customers.
0:11:16 That’s one thing.
0:11:19 It’s like what sort of industries fit that criteria.
0:11:22 And then thinking about what is an actual automatable flow
0:11:23 to start with.
0:11:25 And ones that get us really excited
0:11:27 are flows that are actually revenue generating
0:11:30 where the customer that you would sell to
0:11:32 was previously constrained on the amount of business
0:11:34 that they could handle because of this flow.
0:11:38 So that could be taking customer orders by voice
0:11:41 that was maybe not possible before that now you could do
0:11:43 or it could be like a referral management,
0:11:44 like I said before,
0:11:47 where you just couldn’t process that amount of data quickly enough.
0:11:49 But now you can.
0:11:51 I mean, when you think about the market size as well,
0:11:53 you were just talking about how effectively you’re targeting
0:11:55 what was previously done by labor.
0:11:58 What does that say about the opportunity and the scale of it?
0:11:59 It’s just so much larger.
0:12:01 There’s so many markets.
0:12:04 You look at the market from just like Bureau of Labor Statistics data
0:12:06 and you’re like, this is an enormous market.
0:12:08 And then you look at who the software incumbents are
0:12:12 and you’re like, they just don’t match up to the size of the opportunity.
0:12:14 And that was historically because as I said before,
0:12:16 software could not handle it.
0:12:18 Like the long tail of edge cases
0:12:20 of what these companies were actually doing.
0:12:22 Or they just didn’t have large software budgets.
0:12:26 But all these companies have large labor budgets
0:12:27 and they do have a lot of opportunity
0:12:32 that obviously they do want to wrangle and technology can empower.
0:12:33 And we think with intelligent automation,
0:12:36 this is one of the most exciting times to actually be going after
0:12:37 some of these legacy markets,
0:12:40 seeing whether or not you can actually serve them through AI agents
0:12:43 in a way that maybe traditional work for software couldn’t.
0:12:46 So I think it’s actually like a false comparison
0:12:49 to look at the historical software incumbents and say,
0:12:52 oh, this is the cap on what a company could become.
0:12:54 I think there’s just so much untapped opportunity
0:12:57 that technology just wasn’t able to penetrate before.
0:13:01 Then now with intelligent AI agents, with voice agents, etc.,
0:13:02 you can now tackle.
0:13:04 Yeah, I think you’re absolutely right that we were,
0:13:05 there was all this untapped potential
0:13:07 because the technology only went so far.
0:13:09 But now that we’re here,
0:13:11 how do you see the next five, 10 years evolving?
0:13:14 Because there is a shift that people have to do intellectually as well
0:13:17 as they’re thinking about their software budget to labor budget.
0:13:20 And they almost have to regear their brain to say,
0:13:23 oh, we actually can do this automation, which we previously couldn’t.
0:13:25 So how do you see that trajectory?
0:13:26 I definitely think it’s going to be an evolution.
0:13:30 And I think it’ll depend on the technology spectrum,
0:13:35 like how technology savvy or at the forefront that industry is.
0:13:37 But for a lot of these older industries that we’re talking about,
0:13:39 like the larger ones that are a little bit more on-prem,
0:13:42 a little bit more based in the physical world,
0:13:43 I think it will take time,
0:13:48 which is why I think doing the vertical end-to-end automation solution
0:13:51 is so exciting because you can actually build something
0:13:54 that is very tailored for their specific workflow
0:13:56 where it’s almost a no-brainer to use it.
0:13:58 Like, nobody wants to do data entry.
0:14:02 Nobody wants to sit in the back and read a hundred faxes
0:14:04 and try to input that into a system.
0:14:06 And that’s no company’s core competency either.
0:14:11 So if you’re able to build an intelligent AI agent specifically for that industry
0:14:15 that is tailored to exactly how they do their business,
0:14:16 it’s almost a no-brainer to do it.
0:14:19 And then the folks who were doing that before
0:14:21 can now focus on much higher value,
0:14:24 either customer-facing tasks or much more complex tasks.
0:14:27 And then over time, let’s say in the next five to 10 years,
0:14:30 the technology wave will continue to get adopted by more and more companies,
0:14:33 people that will become more knowledgeable
0:14:35 about what these agents can and cannot do,
0:14:37 more comfortable with the technology.
0:14:40 And then because you’ve integrated yourself with that customer base
0:14:41 with their core systems,
0:14:45 you’ll have the opportunity to take on more and more human labor
0:14:49 or core tasks that their traditional systems record could do.
0:14:51 So it’s a really exciting time, I think, to wedge in now
0:14:56 because there’s a clear opportunity to build something that is ROI-generating
0:14:59 and just an obvious boost to the company’s top line.
0:15:01 But you’ll still get in early enough
0:15:03 that you will have the right to win in the future
0:15:05 as these companies get more and more mature on the adoption curve.
0:15:06 Totally.
0:15:10 And so, obviously, we’re early in this arc, as you mentioned,
0:15:13 but there’s a lot of interesting, exciting things to come.
0:15:15 What would you like to see builders focus on?
0:15:17 What kind of builders would you like to hear from as well?
0:15:21 I would be really excited about people who are thinking about
0:15:23 what was not possible before.
0:15:25 We’ve talked a lot about what RPA does today
0:15:28 and the types of customers it’s able to target today.
0:15:30 When you think about the world of work
0:15:32 that could be intelligently automated away
0:15:37 and the amount of time and savings both employees and companies can get,
0:15:41 it’s just like an order of magnitude larger than what is currently possible.
0:15:43 And so, I’d be really excited about people who are thinking
0:15:46 about the bucket of types of tasks that were automatable
0:15:49 that RPA historically could not handle
0:15:53 and types of industries that it currently was not able to tackle
0:15:57 and really thinking about what are those first flows
0:16:00 or first automations within those industries that are possible
0:16:04 and really thinking about what are the clean UI or UX paradigms
0:16:07 that you could bring to bear for their solutions.
0:16:09 I love that. I love hearing that.
0:16:12 You’re not just interested in hearing from builders in finance or healthcare
0:16:14 but some of these really niche markets.
0:16:15 I think that’s a paradigm shift.
0:16:18 Yeah, and let’s say 10 years from now,
0:16:20 no one has to do manual data entry again
0:16:23 or no one has to get yelled at on the other side of the line
0:16:25 for an angry person in customer service.
0:16:27 I think that’ll be a win for everybody.
0:16:30 And then all these folks can then focus on much more creative, productive tasks
0:16:32 that probably make them happier too.
0:16:33 It finally sends out the facts machine.
0:16:34 Yeah.
0:16:39 All right, that is all for today.
0:16:41 If you did make it this far, first of all, thank you.
0:16:43 We put a lot of thought into each of these episodes,
0:16:45 whether it’s guests, the calendar tetris,
0:16:47 the cycles with our amazing editor Tommy
0:16:49 until the music is just right.
0:16:51 So if you’d like what we put together,
0:16:56 consider dropping us a line at ratethespodcast.com/a16z
0:16:58 and let us know what your favorite episode is.
0:17:01 It’ll make my day and I’m sure Tommy’s too.
0:17:02 We’ll catch you on the flip side.
0:17:05 (upbeat music)
0:17:13 [BLANK_AUDIO]
0:00:03 “Oh, everything’s going to change something.”
0:00:05 What does that mean?
0:00:07 Nobody wants to do data entry.
0:00:10 Nobody wants to sit in the back and read 100 facts
0:00:13 and try to input that into a system.
0:00:15 This is one of the most exciting times
0:00:18 to actually be going after some of these legacy markets.
0:00:20 There’s just so much untapped opportunity
0:00:22 that technology just wasn’t able to penetrate before.
0:00:24 Then now, with intelligent AI agents,
0:00:28 with voice agents, et cetera, you can now tackle.
0:00:31 The internet has given us so much,
0:00:34 whether it’s instant access to the world’s information,
0:00:36 platforms for new art and expression,
0:00:39 nearly every feature film or song in your pocket,
0:00:42 or the ability to work from just about anywhere
0:00:44 with a Wi-Fi connection.
0:00:46 The list, of course, goes on,
0:00:48 but there are still some promises
0:00:51 that haven’t quite hit their mark yet.
0:00:54 One example, robotic process automation,
0:00:55 otherwise known as RPA,
0:00:58 has been a buzzword for years.
0:01:00 A buzzword that promised a revolution
0:01:03 in automating repetitive tasks.
0:01:06 Companies like UiPath found it in 2005,
0:01:08 even promised to enable the, quote,
0:01:10 “fully automated enterprise.”
0:01:12 And while RPA did start strong,
0:01:16 its limitations became evident due to its rigid systems,
0:01:18 sometimes even hard-coding specific processes
0:01:19 down to the button click,
0:01:23 meeting with the unpredictability of most workflows.
0:01:27 Of course, UiPath, with a market cap of $7 billion
0:01:28 as of this recording,
0:01:31 and the RPA industry at large are no failure,
0:01:34 but they are poised for a facelift.
0:01:36 Enter intelligent automation,
0:01:38 a new paradigm powered by AI
0:01:41 that can handle messy, unstructured workflows
0:01:44 that exist in just about every organization,
0:01:45 precisely the types of tasks
0:01:48 that RPA previously could not handle.
0:01:50 So in today’s episode,
0:01:52 we discuss RPA’s second leg,
0:01:55 together with A16Z partner, Kimberly Tan.
0:01:58 Based on her viral article, “RIP to RPA,”
0:02:00 the rise of intelligent automation,
0:02:03 which, of course, will link in the show notes.
0:02:06 So what specifically about AI makes this possible today?
0:02:08 How should startups be thinking about this opportunity?
0:02:13 And why is this opportunity so much larger than the last era?
0:02:14 But before we begin,
0:02:17 I want you to think about the most annoying
0:02:19 and repetitive thing that you need to do.
0:02:22 Now, imagine that task automated.
0:02:25 That future may not actually be so far away.
0:02:28 So listen in to see how you can get involved.
0:02:31 As a reminder,
0:02:33 the content here is for informational purposes only.
0:02:37 Should not be taken as legal, business, tax, or investment advice,
0:02:39 or be used to evaluate any investment or security,
0:02:42 and is not directed at any investors or potential investors
0:02:43 in any A16Z fund.
0:02:46 Please note that A16Z and its affiliates
0:02:47 may also maintain investments
0:02:49 in the companies discussed in this podcast.
0:02:52 For more details, including a link to our investments,
0:02:55 please see a16z.com/disclosures.
0:03:03 Can really you wrote an article with a pretty fun title,
0:03:05 RIP to RPA?
0:03:06 So let’s jump into that.
0:03:09 But first, what is RPA?
0:03:12 RPA stands for Robotic Process Automation.
0:03:16 And it’s a way of basically automating very manual tasks
0:03:17 within an organization.
0:03:21 So things like data entry or invoice processing
0:03:24 that basically every business has to do,
0:03:25 but it’s nobody’s core competency.
0:03:29 Just one of the dirty, messy, internal things within an organization.
0:03:31 So historically, it’s been done very manually.
0:03:33 Like you would just hire a data analyst
0:03:36 or you would hire a back office operations person.
0:03:39 And there was this, I would say, innovation in the last 20 years
0:03:40 where people were like,
0:03:42 “Is it possible to automate these tasks?”
0:03:44 And so the historical way people have done it
0:03:46 is through Robotic Process Automation
0:03:50 where you basically build like a little software bot
0:03:53 that mimics the actual clicks that somebody would be doing.
0:03:54 It’s very deterministic,
0:03:57 meaning they’re literally clicking the different boxes
0:03:59 that I would be clicking as a human.
0:04:01 But organizations are messy
0:04:04 and the work we actually have to do is not perfectly delineated
0:04:06 by a very specific process.
0:04:08 So oftentimes, if something veers a little bit off course,
0:04:10 like maybe someone misspelled a name
0:04:12 or maybe a website changed
0:04:15 where the sign-in box physically is on a page,
0:04:18 then historically that would break the RPA process.
0:04:19 And as you can imagine,
0:04:22 there’s an infinite number of small little things
0:04:23 that could happen like that.
0:04:27 So RPA is often very good for doing 80% of the task,
0:04:29 but then like 20% of the time that it fails,
0:04:31 it’s still a manual person who has to come in.
0:04:35 So it’s just not reliable enough to actually do the full task.
0:04:38 And so you’re still left with having the back office people
0:04:41 that were the first generation doing these sorts of tasks there.
0:04:43 So I just think like with AI and LLMs now
0:04:46 because they’re able to process such unstructured data
0:04:49 and they’re able to intelligently collect context
0:04:51 and then figure out what the best course of action is.
0:04:53 The next generation of actually automating these back office tasks
0:04:56 should be like intelligent AI agents instead.
0:04:59 What can intelligent automation or what you refer to
0:05:01 as these LLMs in action?
0:05:03 What can they do that RPA couldn’t?
0:05:06 Let’s use the example of a company
0:05:08 that we were actually invested in called Tenor.
0:05:12 Tenor does referral management for healthcare practices.
0:05:14 So if I’m a primary physician
0:05:18 and I need to refer a patient to a specialist,
0:05:20 historically, the way that that would be done
0:05:23 is I would literally write something out on a piece of paper.
0:05:26 I would fax it to the specialist.
0:05:29 The specialist front desk person would take the fax,
0:05:31 look at it, look at all the information on it
0:05:33 and then input it into my own database,
0:05:35 check, you know, like the insurance policies,
0:05:36 check prior history, et cetera,
0:05:38 and then decide whether to accept the patient or not.
0:05:40 And that was a very manual task
0:05:43 that there’s just a little bit too much complexity
0:05:46 in the way that it’s done for RPA to be able to handle.
0:05:48 So it had to be some sort of administrative person,
0:05:50 like human who was gonna do it.
0:05:53 And with now like intelligent automation,
0:05:55 Tenor has come up with a very sleek solution
0:05:57 that is basically able to automate that whole process.
0:05:59 And it’s much more self-serve.
0:06:02 ‘Cause the way that RPA would historically work is
0:06:04 you would have to hire an implementation consultant
0:06:08 and they would sit next to whoever was doing the task
0:06:09 and they would basically just watch
0:06:11 what are the clicks that you are doing.
0:06:14 – Right. – And then program those clicks.
0:06:16 But someone like a Tenor, for example,
0:06:17 you’re not gonna have somebody sitting there
0:06:20 watching what the front office admin is doing.
0:06:23 Rather, they’ve created a really sleek UI
0:06:26 where it looks very much like a drag and drop
0:06:28 different process flows.
0:06:30 And they’re able to create their own automation process,
0:06:32 which to them feels very intuitive
0:06:34 so they can set it up themselves,
0:06:35 but actually has a ton of complexity
0:06:37 under the hood that’s being handled.
0:06:38 – I mean, one natural question that comes up,
0:06:39 I think for many people,
0:06:42 especially as they think about things like hallucinations,
0:06:44 is where is the technology in this arc?
0:06:47 Are we able to really achieve this idea
0:06:48 of intelligent automation today?
0:06:49 Are there barriers?
0:06:51 Like where do we sit in that trajectory?
0:06:52 – The way that we’ve seen it work best
0:06:55 is when there’s one very specific automation flow,
0:06:58 at least to start, that a company can just nail,
0:07:01 meaning it’s often industry specific.
0:07:04 So you can integrate into all the core systems there.
0:07:07 You can understand the context for that industry.
0:07:11 And it’s one very repeated, but very manual flow.
0:07:14 So for example, like data entry, I get on a phone call,
0:07:17 I hear the update on where an order is,
0:07:18 all the information from that order
0:07:21 can be parsed through that call
0:07:22 and input it into my main system.
0:07:25 That probably happens like thousands of times a day
0:07:29 for the large organizations, all manually done.
0:07:31 And that is one very specific flow.
0:07:31 And that’s just to start.
0:07:32 And then once you get there,
0:07:34 you can build deeper into other flows,
0:07:37 but I think that is a much more successful path
0:07:39 where you can actually understand the constraints
0:07:39 and build around them,
0:07:42 make sure that the agent performs correctly
0:07:44 versus tackling, let’s say, like everything within healthcare,
0:07:47 everything within legal and logistics to start.
0:07:48 – Normally I ask the question, why now?
0:07:51 But I feel like listeners know that AI is coming,
0:07:55 it’s here, LLMs are maybe the term that a lot of people use.
0:07:57 But is there a deeper why now
0:08:00 or specific technological advances
0:08:02 within the sphere of LLMs that you can point to
0:08:05 that actually make this possible?
0:08:06 – Yeah, I think one thing that we’re really excited about
0:08:08 is people use the term AI and they’re like,
0:08:10 oh, everything’s gonna change now because of AI.
0:08:11 But like, what does that mean?
0:08:14 There’s a lot of very distinct technological breakthroughs
0:08:16 that make different applications possible.
0:08:18 And specific to intelligent automation,
0:08:21 I think one of the things that makes it much more possible
0:08:23 than before is a lot of the fundamental research
0:08:25 coming out of the large labs.
0:08:26 So for example, recently,
0:08:29 Anthropic announced computer use,
0:08:31 which is basically a browser agent
0:08:33 that is able to intelligently understand
0:08:37 what is happening on the browser level of any sort of desktop
0:08:39 and be able to take actions accordingly.
0:08:41 So we talked about how historically,
0:08:45 RPA basically understood at a pixel level,
0:08:47 hey, I should click this thing and then I should click that.
0:08:49 But with something like computer use,
0:08:51 or I think OpenAI has something called operator
0:08:53 that they’re gonna release soon,
0:08:55 agents are gonna be able to browse the internet
0:08:58 and browse the web in a much more sophisticated way,
0:09:00 which is gonna open up a lot of possibilities
0:09:02 for what intelligent agents can do before.
0:09:05 So we think a lot of these intelligent automation startups,
0:09:07 they’re not gonna be doing fundamental research on their own.
0:09:09 There’s still tech that needs to be done
0:09:12 to make a browser agent fully work at scale.
0:09:13 But what’s really exciting for us
0:09:15 is that the large labs are clearly working on this
0:09:17 and clearly understand the opportunity.
0:09:19 And so as that tech gets better,
0:09:21 we think there’s gonna be a whole world of startups
0:09:23 who are able to leverage it
0:09:25 for all the different industries out there
0:09:28 that the large labs themselves are not gonna tackle.
0:09:30 And as you think about the opportunity,
0:09:31 you framed it in your article
0:09:34 as two different paths that people might take.
0:09:36 So one of them was the horizontal AI enabler
0:09:39 and the other was a vertical automation solution.
0:09:41 So tell us about that, the two different paths
0:09:43 that you see if people want to build in this space.
0:09:45 So the first is the horizontal AI enabler
0:09:47 and that’s something that we think
0:09:50 any company who’s doing any sort of automation,
0:09:52 intelligent automation, is going to have to do.
0:09:53 One very common example,
0:09:55 which I’ve touched upon a little bit already,
0:09:56 is data extraction.
0:10:00 Almost every intelligent automation path
0:10:03 starts with some messy unstructured data
0:10:06 that you need to pull key outputs from.
0:10:09 And today, a lot of people are just building that manually.
0:10:11 But we’ve started to see companies emerge
0:10:13 that are purely doing that path,
0:10:15 which is taking unstructured data
0:10:16 and pulling out the key pieces
0:10:18 to turn it into structured data.
0:10:21 And we think that could be one really interesting opportunity.
0:10:23 So anyone who is either building
0:10:26 their own automation in-house can leverage that
0:10:27 as a key component.
0:10:29 Or if you’re building like a full end-to-end solution,
0:10:32 maybe you input that as one of your components as well.
0:10:33 One thing that I’m personally really excited about
0:10:35 is the vertical automation path.
0:10:39 I think to make an intelligent AI agent very successful,
0:10:41 it often is helpful in the beginning
0:10:44 to have it be in a very constrained domain.
0:10:46 For example, in logistics or in healthcare and legal,
0:10:49 it is a domain that they can understand all the context for.
0:10:52 They have all the necessary inputs, integrations, etc.
0:10:54 And they’re able to automate one specific flow.
0:10:56 So what we’re really excited about there is,
0:10:59 let’s take an industry that does have a lot of manual work
0:11:03 that needs to be done, like a very large back office.
0:11:04 If you think about what are the things there
0:11:07 that actually have to be automated
0:11:10 that maybe RPA wasn’t able to tackle before
0:11:14 because it just wasn’t like a large enough individual customer.
0:11:16 Like it wasn’t one of the Fortune 500 customers.
0:11:16 That’s one thing.
0:11:19 It’s like what sort of industries fit that criteria.
0:11:22 And then thinking about what is an actual automatable flow
0:11:23 to start with.
0:11:25 And ones that get us really excited
0:11:27 are flows that are actually revenue generating
0:11:30 where the customer that you would sell to
0:11:32 was previously constrained on the amount of business
0:11:34 that they could handle because of this flow.
0:11:38 So that could be taking customer orders by voice
0:11:41 that was maybe not possible before that now you could do
0:11:43 or it could be like a referral management,
0:11:44 like I said before,
0:11:47 where you just couldn’t process that amount of data quickly enough.
0:11:49 But now you can.
0:11:51 I mean, when you think about the market size as well,
0:11:53 you were just talking about how effectively you’re targeting
0:11:55 what was previously done by labor.
0:11:58 What does that say about the opportunity and the scale of it?
0:11:59 It’s just so much larger.
0:12:01 There’s so many markets.
0:12:04 You look at the market from just like Bureau of Labor Statistics data
0:12:06 and you’re like, this is an enormous market.
0:12:08 And then you look at who the software incumbents are
0:12:12 and you’re like, they just don’t match up to the size of the opportunity.
0:12:14 And that was historically because as I said before,
0:12:16 software could not handle it.
0:12:18 Like the long tail of edge cases
0:12:20 of what these companies were actually doing.
0:12:22 Or they just didn’t have large software budgets.
0:12:26 But all these companies have large labor budgets
0:12:27 and they do have a lot of opportunity
0:12:32 that obviously they do want to wrangle and technology can empower.
0:12:33 And we think with intelligent automation,
0:12:36 this is one of the most exciting times to actually be going after
0:12:37 some of these legacy markets,
0:12:40 seeing whether or not you can actually serve them through AI agents
0:12:43 in a way that maybe traditional work for software couldn’t.
0:12:46 So I think it’s actually like a false comparison
0:12:49 to look at the historical software incumbents and say,
0:12:52 oh, this is the cap on what a company could become.
0:12:54 I think there’s just so much untapped opportunity
0:12:57 that technology just wasn’t able to penetrate before.
0:13:01 Then now with intelligent AI agents, with voice agents, etc.,
0:13:02 you can now tackle.
0:13:04 Yeah, I think you’re absolutely right that we were,
0:13:05 there was all this untapped potential
0:13:07 because the technology only went so far.
0:13:09 But now that we’re here,
0:13:11 how do you see the next five, 10 years evolving?
0:13:14 Because there is a shift that people have to do intellectually as well
0:13:17 as they’re thinking about their software budget to labor budget.
0:13:20 And they almost have to regear their brain to say,
0:13:23 oh, we actually can do this automation, which we previously couldn’t.
0:13:25 So how do you see that trajectory?
0:13:26 I definitely think it’s going to be an evolution.
0:13:30 And I think it’ll depend on the technology spectrum,
0:13:35 like how technology savvy or at the forefront that industry is.
0:13:37 But for a lot of these older industries that we’re talking about,
0:13:39 like the larger ones that are a little bit more on-prem,
0:13:42 a little bit more based in the physical world,
0:13:43 I think it will take time,
0:13:48 which is why I think doing the vertical end-to-end automation solution
0:13:51 is so exciting because you can actually build something
0:13:54 that is very tailored for their specific workflow
0:13:56 where it’s almost a no-brainer to use it.
0:13:58 Like, nobody wants to do data entry.
0:14:02 Nobody wants to sit in the back and read a hundred faxes
0:14:04 and try to input that into a system.
0:14:06 And that’s no company’s core competency either.
0:14:11 So if you’re able to build an intelligent AI agent specifically for that industry
0:14:15 that is tailored to exactly how they do their business,
0:14:16 it’s almost a no-brainer to do it.
0:14:19 And then the folks who were doing that before
0:14:21 can now focus on much higher value,
0:14:24 either customer-facing tasks or much more complex tasks.
0:14:27 And then over time, let’s say in the next five to 10 years,
0:14:30 the technology wave will continue to get adopted by more and more companies,
0:14:33 people that will become more knowledgeable
0:14:35 about what these agents can and cannot do,
0:14:37 more comfortable with the technology.
0:14:40 And then because you’ve integrated yourself with that customer base
0:14:41 with their core systems,
0:14:45 you’ll have the opportunity to take on more and more human labor
0:14:49 or core tasks that their traditional systems record could do.
0:14:51 So it’s a really exciting time, I think, to wedge in now
0:14:56 because there’s a clear opportunity to build something that is ROI-generating
0:14:59 and just an obvious boost to the company’s top line.
0:15:01 But you’ll still get in early enough
0:15:03 that you will have the right to win in the future
0:15:05 as these companies get more and more mature on the adoption curve.
0:15:06 Totally.
0:15:10 And so, obviously, we’re early in this arc, as you mentioned,
0:15:13 but there’s a lot of interesting, exciting things to come.
0:15:15 What would you like to see builders focus on?
0:15:17 What kind of builders would you like to hear from as well?
0:15:21 I would be really excited about people who are thinking about
0:15:23 what was not possible before.
0:15:25 We’ve talked a lot about what RPA does today
0:15:28 and the types of customers it’s able to target today.
0:15:30 When you think about the world of work
0:15:32 that could be intelligently automated away
0:15:37 and the amount of time and savings both employees and companies can get,
0:15:41 it’s just like an order of magnitude larger than what is currently possible.
0:15:43 And so, I’d be really excited about people who are thinking
0:15:46 about the bucket of types of tasks that were automatable
0:15:49 that RPA historically could not handle
0:15:53 and types of industries that it currently was not able to tackle
0:15:57 and really thinking about what are those first flows
0:16:00 or first automations within those industries that are possible
0:16:04 and really thinking about what are the clean UI or UX paradigms
0:16:07 that you could bring to bear for their solutions.
0:16:09 I love that. I love hearing that.
0:16:12 You’re not just interested in hearing from builders in finance or healthcare
0:16:14 but some of these really niche markets.
0:16:15 I think that’s a paradigm shift.
0:16:18 Yeah, and let’s say 10 years from now,
0:16:20 no one has to do manual data entry again
0:16:23 or no one has to get yelled at on the other side of the line
0:16:25 for an angry person in customer service.
0:16:27 I think that’ll be a win for everybody.
0:16:30 And then all these folks can then focus on much more creative, productive tasks
0:16:32 that probably make them happier too.
0:16:33 It finally sends out the facts machine.
0:16:34 Yeah.
0:16:39 All right, that is all for today.
0:16:41 If you did make it this far, first of all, thank you.
0:16:43 We put a lot of thought into each of these episodes,
0:16:45 whether it’s guests, the calendar tetris,
0:16:47 the cycles with our amazing editor Tommy
0:16:49 until the music is just right.
0:16:51 So if you’d like what we put together,
0:16:56 consider dropping us a line at ratethespodcast.com/a16z
0:16:58 and let us know what your favorite episode is.
0:17:01 It’ll make my day and I’m sure Tommy’s too.
0:17:02 We’ll catch you on the flip side.
0:17:05 (upbeat music)
0:17:13 [BLANK_AUDIO]
Intelligent automation is transforming industries by tackling messy, unstructured workflows that traditional Robotic Process Automation (RPA) couldn’t handle. In this episode, a16z partner Kimberly Tan discusses the shift from rigid RPA systems to AI-powered agents and why this evolution unlocks massive opportunities in legacy markets.
Drawing on her article “RIP to RPA: The Rise of Intelligent Automation,” Kimberly shares real-world examples of companies revolutionizing referral management, and highlights how startups can build impactful solutions in this space.
Resources:
Find Kimberly on X: https://x.com/kimberlywtan
Read Kimberly’s article: https://a16z.com/rip-to-rpa-the-rise-of-intelligent-automation/
Stay Updated:
Let us know what you think: https://ratethispodcast.com/a16z
Find a16z on Twitter: https://twitter.com/a16z
Find a16z on LinkedIn: https://www.linkedin.com/company/a16z
Subscribe on your favorite podcast app: https://a16z.simplecast.com/
Follow our host: https://twitter.com/stephsmithio
Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see a16z.com/disclosures.
Leave a Reply