Latest / State of the AI Union / Humans in the Loop
Transcript
- 0:11Hello, everybody, and welcome to this week's episode of the State
- 0:14of the AI Union. I am your host, Laura.
- 0:17And join with me today is Anirudh Shenoy.
- 0:20Anirudh, do you want to introduce yourself?
- 0:22Sure. Thanks, Laura.
- 0:23Thanks for having me. So I'm Anirudh, I'm a product
- 0:25manager here at Debra. I work on our AI platform,
- 0:28particularly on agents and search.
- 0:30I've been working in this field for about 6 years as a product
- 0:34manager earlier in the sort of pre ChatGPT era where I was
- 0:38working on more conversational AI, more NLP, more traditional
- 0:41ML systems, and then after that I've transitioned more into like
- 0:45generative AI product building and so on and so forth.
- 0:47So I've seen both the pre chart Jeopardy era of building AI
- 0:50products and then the posterior activity era of building AI
- 0:53products. Fantastic.
- 0:54Andy Ruud, you know, today we want to talk about like humans
- 0:58in the loop and this concept of humans in the loop earlier,
- 1:02earlier in the year, Garna, who I think is like the AI golden
- 1:07child of like using, using AI for efficiency in their, in
- 1:13their, you know, operation. They like need off hundreds of
- 1:17workers. And they basically said, you
- 1:18know, we're going to be able to use Gen.
- 1:20AI to to basically do a lot of our work.
- 1:23And, you know, there's a lot of news around them.
- 1:25And recently they've since rehired a bunch of these
- 1:29workers. A lot of them actually were
- 1:31customer service workers. And I think one of the, you
- 1:36know, in one of the studies when we looked at the future of AI,
- 1:40right, we see and the two sort of spectrums where, OK, one is
- 1:44fully human and one is fully AI. Are we going to move to the part
- 1:48where, you know, it's all going to be AI?
- 1:50And I think the answer is no, We're kind of going to land in
- 1:53the middle where humans are going to be a lot more efficient
- 1:56with the AI, right? So I wanted to ask you maybe
- 1:59first for your thoughts, Like what do you think about when you
- 2:03hear the term human in the loop? It's, it's, as the name
- 2:07suggests, right, There's a sort of loop running, which is sort
- 2:10of like an agent loop. And the idea is to bring some
- 2:14sort of human oversight, some sort of decision making where
- 2:18you probably want to have a human to oversee that sort of
- 2:20agentic loop, right? And you, you mentioned Clarina,
- 2:23right, which is a fantastic example of this pendulum sort of
- 2:27swinging in the extreme opposite direction, which is going from a
- 2:31fully human centered sort of customer support system to 0
- 2:36humans, right? And that's the pendulum swinging
- 2:38towards the extreme. And I think Klarna sort of
- 2:42realized that it's a lot harder than it seems to kind of take
- 2:45some of these agent take systems to production, which is true,
- 2:48right? Just like any other system out
- 2:52there, nothing is 100% foolproof.
- 2:54But when you sort of have that expectation that, hey, this
- 2:57system is going to run autonomously without needing any
- 3:00sort of human input at any sort of time, that expectation
- 3:03slightly becomes a little bit too much.
- 3:06And that's sort of, I think that's when Clara realized that,
- 3:09hey, we'll have to sort of dial it back a little bit, right?
- 3:12And bring in some sort of human and look, like you mentioned.
- 3:17So just to kind of summarize, when I hear the word human in
- 3:21the loop, it's basically how can we bring some sort of human
- 3:24judgment, some sort of human oversight into an agentic system
- 3:27so that it doesn't just go crazy and starts doing things on its
- 3:31own? Yeah.
- 3:32And do you think that it's actually both on the the people
- 3:38using the AI and the people experiencing the AI where on
- 3:42both sides they want to feel like a humanist listening to
- 3:48them? And I don't know specifically,
- 3:49for example, in this kind of story, whether their customers
- 3:52were complaining or whatever, right?
- 3:54But but to hear that, to hear that they've rehired, I think
- 3:58some of their the first people that they rehired were actually
- 4:00their customer support people make me feel like their
- 4:03customers wanted to feel like it could be more hard.
- 4:08I think it depends a lot on the sort of use cases that are also
- 4:11being automated, right? See, for example, if, if I'm,
- 4:15you know, calling my food delivery service provider, I'm
- 4:19calling the customer support there.
- 4:20And all I just want to know is, hey, where's my order?
- 4:22Like something very simple like that.
- 4:23I would just prefer it if I was just connected to some sort of
- 4:27AI agent or, you know, even if it's just through the app,
- 4:30right? Because it's a very simple
- 4:31thing. I don't want to even waste my
- 4:32time trying to figure out what's going on, right?
- 4:35Just tell me that, hey, it's on the way, It's delayed by a
- 4:36couple of minutes and you're done.
- 4:38But let's say I've waited for an hour, right?
- 4:40And I was promised that I would get my food in 15 minutes, then
- 4:43I'm going to be really, really angry, right?
- 4:45Like I want to, you know, imagine an AI agent coming in
- 4:49the way there even, and you know, it's just delaying things
- 4:52is just giving me this sort of robotic answer saying, hey, it's
- 4:55going to be there. It's going to be there, right.
- 4:56Then I want to talk to a human being.
- 4:58So I think it depends on the sort of use case, the sort of
- 5:01context that's there, the sort of background that's there to
- 5:04really figure out what needs a human in the loop and what
- 5:08doesn't need a human in the loop, right?
- 5:09I think there is a fine balance between the two.
- 5:12Yeah. And, you know, we've talked a
- 5:13lot about on this podcast, we've talked a lot about enterprise
- 5:16buyer and how there's a lot of conversation and lots of people
- 5:20are willing to have conversations.
- 5:21But there's still some hesitancy when it comes to sort of rolling
- 5:25it out to production or, you know, even adopting, adopting
- 5:29some kind of AI. And I think a lot of it has to
- 5:32do with the level of trust, right?
- 5:35Yeah. Tell me more about what you've
- 5:37seen when when it comes to sort of rolling out AI.
- 5:41Yeah. So this is pretty interesting,
- 5:44right? Because I've, I've seen I've
- 5:47worked in the AI space for like I said about 6-7 years now
- 5:50almost right. And any time we've sort of
- 5:53rolled out AI systems as part of a POC and as part of a larger
- 5:57pilot project, the POC phase has generally taken quite a bit of
- 6:01time in the sort of in the old school way of doing things,
- 6:05right? Because you had to sit up,
- 6:06manually train some of these models, you had to go and say,
- 6:09yeah, it should do this, it should not do this, it should do
- 6:12this, it should not do this, right?
- 6:13And it generally took a lot of time.
- 6:15But I think with AI, what you're seeing is hits.
- 6:17It's very easy to build quick and early prototypes, right?
- 6:21I could just like if you had an agent in mind, you could do it
- 6:24in like NA10 or you could do it in any of these systems in
- 6:27probably a couple of hours. I guess you can get something
- 6:29out there to quickly try and see what it takes, you know,
- 6:33experience some of these AI agents, but then taking it to
- 6:36production, that's a whole different ball game, right?
- 6:38You being the sole user of the agent versus you deploying this
- 6:42AI agent to 1,000,000 customers, 2 million users, 10 million
- 6:47users. That's where there's going to be
- 6:49some, you know, instances where this agent starts acting rogue,
- 6:53starts doing all sorts of crazy stuff and that's where the
- 6:56roadblocks come in. Testing it in a small group,
- 6:58that's fine. But then taking it to
- 6:59production, that's a whole different ball game all
- 7:01together, right? You need to have guardrails in
- 7:03place. You need to have proper
- 7:05observability in place. You need to start doing more
- 7:07stress testing. You need to start doing more
- 7:09regression testing. You need to start doing good
- 7:12evals. You need to have human in the
- 7:14loop, right? And I think that is where some
- 7:16of these PO CS, especially in the enterprise segment, are sort
- 7:19of falling flat, right? Because you know, you do these
- 7:22quick prototypes very early, but then you start rolling it out to
- 7:26a larger set of users, things start falling short.
- 7:29And I think there's also a little bit of an expectation
- 7:32mismatch some places because we are all used to a very, when we
- 7:36think of software, we all think of deterministic software,
- 7:39right? In the sense the software of old
- 7:42has always been, hey, you click this button and then this thing
- 7:45happens 100% of the time. So that's how we expect software
- 7:49to work. But when you talk about AI
- 7:51agents, they're all probabilistic systems, right?
- 7:54You ask a question, 59% of the time you'll probably get an
- 7:57answer. But it's all probabilistic
- 7:59systems, right? 1% of the time, it might do
- 8:01something else. You've used ChatGPT, right,
- 8:03Laura? Like you ask a question there,
- 8:05you get an answer. But the next time you ask the
- 8:06same question, you might get something slightly different,
- 8:08but it's still the essence of the answer is still the same.
- 8:11And these sort of things is what makes it a lot harder for people
- 8:15to digest what it takes to kind of take these agents to
- 8:19production. And this is I think one of the
- 8:20places where the human in the loop actually really, really
- 8:23matters because then you, your agent doesn't have to be 100%
- 8:26accurate, right? Whenever it's not sure about
- 8:28something, it has a sort of way to mitigate it by just asking a
- 8:33human, right? Hey, I think this is what it
- 8:35should be. What do you think?
- 8:36Can you approve this? I'm going to offer a refund to
- 8:39this customer. Can you approve this amount?
- 8:42I'm going to, I don't know, approve this insurance claim.
- 8:45Can you verify it once? You don't have to be 100.
- 8:48The agent, the AI agent doesn't have to be 100% accurate in that
- 8:52sense. When there's a human in the loop
- 8:54that sort of, you know, slightly reduces the expectation and the
- 8:58burden on the AI system, right? It doesn't have to be perfect.
- 9:01It can be probabilistic in that sense.
- 9:03Right. And you know what you said
- 9:04about, about us, the expectations of software to be
- 9:08deterministic. I think that's very true, right?
- 9:10Like to, to have an AI agent means it is going to be
- 9:15probabilistic and it's going to act like, like a human being,
- 9:17which we will never write 100% of the time.
- 9:20So, you know, designing a system where it, it can go and ask the
- 9:25human and it does it proactively.
- 9:27Is, is a, is a, is a really good solution here.
- 9:31Exactly. But before I think we we talk
- 9:33about maybe some design principles, I wanted to actually
- 9:37ask what you think are some differences in humans being in
- 9:42the loop versus just like team managers, you know, like what
- 9:45are some of the things we're going to have to change when we
- 9:48are managing AI agents versus managing human beings?
- 9:52Right. Yeah.
- 9:52So that's a fantastic question. And I think this is we're, we're
- 9:56learning a lot about this, you know, as as the technology
- 10:00matures, as you know, we start standardizing some of these
- 10:03things. In fact, I, I, I remember
- 10:07reading about this or listening about this in, in, in a talk
- 10:11recently given by to Sequoia. And one of the things that was
- 10:14very interesting there is, you know, they were mentioning that
- 10:19the ICS of today, like the individual contributors of
- 10:23today, the expectation from them is that, you know, they
- 10:26basically do work that's kind of assigned to them or work that
- 10:30sort of, you know, the coding work or whatever else it is
- 10:34there, right? Like they are the people who are
- 10:35like the leaf node, right? Like they don't delegate
- 10:38anymore. They actually do the work.
- 10:40But now with AIA, lot of these ICS are now individual
- 10:45contributors. They're now kind of becoming
- 10:47manager of agents. They are expected to figure out
- 10:51how to delegate work and not delegate work to another IC,
- 10:55delegate work to agents. So this is a big sort of shift
- 10:59in the way that work itself is going to change or in the way
- 11:04that ICS need to start thinking about the work now.
- 11:08For example, I'm an IC, right? I'm an individual contributor
- 11:10here. I'm a product manager and my job
- 11:13is really to sit and define some of these features and I'm I'm
- 11:15like the person who's like generating work.
- 11:17But of late, right, like of the last one year, as we have
- 11:20started incorporating more agents into the work, I just now
- 11:24start thinking more about, Hey, how do I need?
- 11:26What do I need to ask perplexity or what do I need to ask DP
- 11:30search or what do I need to ask, like our develop agent, right?
- 11:33How do I figure out what is the right question to ask rather
- 11:37than figuring out what the answer should be right.
- 11:40That being said, it's not like I'm not, I'm just going to take
- 11:42the answer that I get and just pass it along, right?
- 11:44I need to wait. I need to be the human in the
- 11:46loop in this case, right? But there's, there's this
- 11:49fundamental shift in the way that work is transitioning,
- 11:52which is it's, it's less about figuring out what the right
- 11:55answer is and more about figuring out what the right
- 11:58questions are to ask, right? And the best part about this is
- 12:01if some work used to take 10 hours for me to do, now I have a
- 12:05lot more leverage on that, right?
- 12:07I can do probably X or 2X of the work that I could do.
- 12:10Because now it's more about delegating it to the agents and
- 12:13acting as a human in the loop or a manager in the loop for some
- 12:16of these agents rather than just doing the work myself, honestly.
- 12:21Like I don't write up a lot of documents nowadays.
- 12:23I at least get the first draft written by some of these agents
- 12:27that even the internal agents that we use at the right.
- 12:29And that draft then acts like my base.
- 12:32And then I go and I say, you know, this needs to change.
- 12:34That needs to change. This needs to change.
- 12:35But it's it's basically like having, for lack of better
- 12:38words, large army of interns, right, Who you can delegate it
- 12:42to work to. Yeah.
- 12:44Maybe in your personal experience, it's a great example
- 12:46of how you've changed the way that you've worked, you know,
- 12:49with with AI. What have you seen yourself
- 12:53doing more of as you've been asking the questions?
- 12:56Have you have you sort of like changed the way that you're
- 12:59thinking about, you know, output from the agent or from your
- 13:04teammates? How has that helped shape you
- 13:07yourself in your, in your own thinking or your own like doing
- 13:10work? I think I.
- 13:11End up asking a lot of why questions now, right?
- 13:15Earlier it was more about like, you know, what do we need to do?
- 13:20Maybe how do we need to do it? But now since it's a lot easier
- 13:23to get these, why is this happening or why should we do?
- 13:28It's a lot harder to kind of get answers to those kind of sort of
- 13:32questions in the earlier era, right?
- 13:34Like if you go to Google and you just ask why is something this
- 13:37way, it'll probably take you to a link and then you have to
- 13:39scour the Internet to sort of get answers to that, right?
- 13:43But for example, if I just want to know, you know, why are our
- 13:46customers asking for this particular feature today?
- 13:49I can go to like the dev Rev agent and get these answers very
- 13:51easily now, right? Earlier I had to spend probably
- 13:54days are kind of going through tickets and kind of getting a
- 13:57sense of these things. But now it's a lot easier to get
- 14:00the insights like this a lot faster.
- 14:02So that's one part of it. I think the second part is
- 14:05simply that. Yeah, I mean it's it's it's the
- 14:09the way that we start thinking about work itself is like I
- 14:12said, changing a lot, right? It's more about how do I break
- 14:16down this large task into smaller tasks that I can sort of
- 14:19delegate rather than, you know, just figuring out, OK, I need to
- 14:23do this, this, this, this, that's right.
- 14:24It's, it's more the mindset has shifted more towards delegation
- 14:28and figuring out what can be automated rather than figuring
- 14:32out, oh, when should I do this work?
- 14:34Like earlier, I would just open up my calendar and kind of slot
- 14:37in and say, I should do this work at this time.
- 14:39Now I first break it down and say, hey, is there something
- 14:42that I can just quickly get it done by using deep research or
- 14:45by using an agent rather than just sitting and doing it
- 14:48myself? So if I'm hearing.
- 14:49You correctly, I think the the first thing was what you said
- 14:52was like answering the question why and probably that is giving
- 14:56you more clarity in your work too, because we're focused less
- 15:00on the how it gets done and more of the OK, why are we doing
- 15:02this? Are we actually answering the
- 15:04right question? So hopefully that will also give
- 15:07us like in general more clarity around the work that we are
- 15:10doing, why it's important, etcetera.
- 15:12And then the second part is like efficiency, just like getting
- 15:16the small bits out, right? It's what?
- 15:17What about? So let's go back to human in the
- 15:20loop design. What does good design look like?
- 15:22Have you seen it in other applications when we're
- 15:26designing for systems and agents?
- 15:29How do you envision some of these best practices?
- 15:32Yeah, I think a lot. Of these best practices are, I
- 15:34mean, this is the best part about the age of AI that we are
- 15:38in, right? Nobody knows the right answer.
- 15:40And that's good news as well as bad news, right?
- 15:42Good news is that it's it's like a level playing field for
- 15:45everyone, right? A Google, a startup, a Microsoft
- 15:49or maybe a startup in Palo Alto, a startup in India, all of them
- 15:53are on the same level playing field, right?
- 15:55Nobody knows the right answer. Everyone's trying to figure it
- 15:57out. So that's the best part of it.
- 15:59But I think there are a lot of interesting sort of design
- 16:02paradigms that are kind of going around, right?
- 16:05One of them that comes to mind is cursor.
- 16:07Cursor is a fantastic example of sort of human in the loop,
- 16:10right? You ask it to go and write code
- 16:13for something. It goes and writes that code.
- 16:15But any time it needs to take a dangerous action, it needs to
- 16:18run like a command line. You know, it needs to run a
- 16:22command on a terminal, it needs to delete a file, it needs to
- 16:25create a file, it needs to do a network call, whatever else it
- 16:27might be. You always see there is a box at
- 16:30the bottom that says, hey, I want to run this.
- 16:33Can you approve it? Right.
- 16:35Only when you go and click approve it, it runs that
- 16:37command. And the reason for that again,
- 16:39right? Like sure, Cursor does have the
- 16:42way where you can say, you know, auto approve.
- 16:43I don't care, right? Just do whatever you want,
- 16:45right? Don't ask me again.
- 16:46But you know that they they have made the default as ask because
- 16:51they know that some of these things can be dangerous, right?
- 16:53There could be this naive person who doesn't understand a lot of
- 16:56these things and the agent starts generating a lot of
- 17:00malicious, maybe not malicious, but maybe dangerous code.
- 17:03It goes out and deletes maybe your document.
- 17:05It goes ahead and starts creating all sorts of weird
- 17:08code, right. So that's that's one interesting
- 17:10thing. I think that's the other one
- 17:11that comes to mind. This sort of similar thing is
- 17:13what Apple Intelligence has done with Siri and ChatGPT.
- 17:16Like if you've used Apple Intelligence, there is an option
- 17:18to sort of when you talk to Siri, you can ask Siri to kind
- 17:21of ask ChatGPT and Siri comes back and says, hey, should I ask
- 17:25ChatGPT this question? And then if you say yes, then
- 17:29they go and do it. Now you can see that this is
- 17:31sort of a double edged sword, right?
- 17:32In some cases it's very cumbersome.
- 17:33It doesn't make sense for you to say every time, yeah, yeah,
- 17:35sure, go and do it. But again, they also have an
- 17:38always allow kind of option. So this is all very interesting
- 17:42stuff, paradigms like this where people are trying to figure out
- 17:45how to bring in human in the loop and sort of what's the best
- 17:49way, because at times it kind of gets annoying if the agent just
- 17:53reaches out to a human for all sorts of very dumb things,
- 17:56right? You know, if you go so the exact
- 17:58opposite direction, if an agent doesn't approach a human being
- 18:01for a very important thing, that's a bad thing as well,
- 18:03right? So again, there's that fine
- 18:05balance of when the agent needs to reach out to a human being to
- 18:09get an approval or to ask for the right thing to do.
- 18:12So I think a lot of these design choices are still being figured
- 18:15out. Even on the intelligence front.
- 18:16I think there's a lot that needs to be figured out.
- 18:19Agents need to understand when exactly to reach out.
- 18:22But I think that will happen as as time, yeah.
- 18:24So it's really about tuning, tuning the agent and customizing
- 18:29it to say when, when you should do something and when, when you
- 18:33should not do it. And when you're thinking about,
- 18:35when you're thinking about buyers today or either buyers or
- 18:38deployers, maybe of AI agents, what are some things you may
- 18:42encourage them to do? Maybe like on the buyer side,
- 18:46you know, as they're, as you talked about how Pocs are
- 18:48getting longer because we're having to, you know, introduce
- 18:51the human, like what are some things you may, you may
- 18:54encourage them to do so that they can actually get started
- 18:57without, you know, this. Yeah, I think.
- 19:00One is start thinking more from that probabilistic mindset,
- 19:04right? These are not traditional
- 19:05software systems. I've seen a lot of, you know,
- 19:09buyers come up with this, you know, very sort of fixed
- 19:14requirements for lack of better words, right?
- 19:16Which is if a user says this, this always needs to happen,
- 19:21right? If a user says this to my agent,
- 19:24it needs to use these exact same words, right?
- 19:27I think if that's sort of a mindset is that whenever you're
- 19:31trying to think of an AI system, especially a probabilistic AI
- 19:34system, it's like doom to failure, right?
- 19:37Again, like I've said earlier, like if you go and use ChatGPT,
- 19:40even today, you ask one question today and the same question
- 19:43tomorrow, you're probably going to get a different answer.
- 19:45The meanings remains the same. The output, the concept of
- 19:49whatever the answer is remains the same, but the wordings are
- 19:52going to change, right? It's going to be a little bit
- 19:54different. The structure of the answer
- 19:56might be a little bit different, but the answer still, the
- 19:58essence of the answer still remains the same.
- 20:00So I think, you know, approaching any of these buying
- 20:03decisions by having a very, you know, old school software
- 20:07mindset. It's kind of doomed to failure.
- 20:09You need to have this probabilistic mindset.
- 20:11You need to understand that these are more human like
- 20:15systems than in the past, right? Just like humans make mistakes,
- 20:19Just like how humans might make small changes here and there
- 20:23when they're replying to a person.
- 20:25These systems are very similar to that.
- 20:28Yeah, and it's interesting. It's interesting.
- 20:30Sorry to cut you off your butt. Like it's, it's interesting what
- 20:32you're saying because you're saying that, you know, the
- 20:35mindset that you have to go into buying AI is to actually think
- 20:39about AI like humans, right? And you know, that's, that may
- 20:44be like conflicting with what, with what buyers are entering
- 20:48the buy cycle into because they're thinking I want to
- 20:50replace my humans. But yet to make it successful,
- 20:53you have to, you have to think about them like humans.
- 20:56I think that's super interesting.
- 20:58Yeah, Yeah. What else?
- 20:59And I think we're also seeing that pricing is also changing a
- 21:02little bit, right? There are a lot of these AI
- 21:05start-ups out there, which is just sort of pricing it in the
- 21:07way where we're treating this as like, OK, if a human being was
- 21:11doing this, it would cost this much.
- 21:12But now you're an AI system, it's going to cost you this
- 21:15much, right? So it's it's more value based
- 21:17pricing as opposed to like, you know, seed based pricing that
- 21:21was there earlier. It's a lot of consumption based
- 21:23pricing sort of model now. And that's sort of takes me to
- 21:26the other point, right, Which is sure now that you need to think
- 21:29of these systems more from human tendency, like what how a human
- 21:32being would do it. You should also have a sort of
- 21:35sort of an escape mechanism, right?
- 21:37In case the AI system is not able to figure out what to do,
- 21:40what happens then? Does it delegate to a human
- 21:44being? Does it seek approval from a
- 21:45human being? Does it hand off to a human
- 21:47being? Right?
- 21:47There's a lot of different things that can happen there.
- 21:50The AI system could simply say, I can't do this.
- 21:53I'm moving it over to a human being.
- 21:54Or the AI system could say, hang on a second, let me figure out
- 21:58what to do. Let me contact someone, right?
- 22:00Let me contact my human superior, whoever is its
- 22:02manager. Let me see what I can do here,
- 22:04right? So does the system even have
- 22:07this sort of escape mechanism? A human in the loop perhaps to
- 22:11handle these sort of systems where situations where it
- 22:14doesn't know what to do, Like these are all questions I think
- 22:17that a buyer should start thinking about, right?
- 22:19One is, of course, having the probabilistic mindset.
- 22:21Second is figuring out what's the escape mechanism.
- 22:24And I think the third one is also, you know, on the lines of,
- 22:27you know, what sort of design philosophy is, is necessary when
- 22:31we are thinking about the interaction itself, right?
- 22:35A lot of these sort of agents are today chat driven in the
- 22:39sense you just go and ask a question and you get a response.
- 22:42But depending on your use case, that might not be the right
- 22:45interface, right? And that's one of the big
- 22:47innovations that actually, you know, folks like cursor brought
- 22:50in, which is, and you might even have seen comment from publicity
- 22:54recently, right? They've decided to just go out
- 22:56and build their own browser because they understand the
- 22:58value of a good UI. So these are just some some high
- 23:02level things to think, do you think?
- 23:03That when the end customer, right.
- 23:07So let's let's say somebody that goes on the website or whatever
- 23:10starts a chat. Do you think that for the
- 23:12customer it is valuable for them to know when they're talking to
- 23:18an AI versus when they're talking to a human?
- 23:21Yeah. It's a, that's $1,000,000
- 23:23question. I think we've, I've seen things
- 23:25kind of, again, it's, it's a big spectrum.
- 23:27I've seen the pendulum swing both ways, right?
- 23:30I feel like now more than ever, and this answer might change two
- 23:35years down the line. It, the AI should identify
- 23:38itself as an AI just so that the other side knows that there are
- 23:43limitations in its capabilities, right?
- 23:45That there are some things that it can do and then there's some
- 23:47things it can't do, right? And I think that instills some
- 23:51trust with your users instead of sort of misleading them and
- 23:55making them believe that it's not sorry.
- 23:57It it is an yeah, it's not an AI sorry.
- 24:00I've seen that, you know, some of these AISDR companies, you
- 24:04know, which do cold outreaches to prospects.
- 24:08It's like misfired a lot some places right where the AI
- 24:12doesn't really. You mentioned that it's an AI
- 24:15and the prospect just feels like this company doesn't even bother
- 24:18having a human being reach out to me and it's more like an they
- 24:21don't care. About me.
- 24:22They feel like they don't care exactly.
- 24:24Exactly. So yeah.
- 24:26Yeah, that's very interesting because you know, even now when
- 24:29we call contact centers, right and they like sometimes you are
- 24:34talking to to a robot, it's not even an AI agent, it's a robot.
- 24:38And then they go, oh, let me check, let me check on this.
- 24:41And you can hear them typing. And it's so weird, right?
- 24:44Because you know that it's the robot.
- 24:47And I think, I think what you're seeing is, is true.
- 24:49I mean, as consumers, we are also evolving in, in the age of
- 24:53AI where I think interacting with an AI agent is no longer
- 24:57going to be OK, cut my, my, my vendor doesn't care about me,
- 25:02right? But it, but as customers, I
- 25:04think we also have to be able to, to, to be flexible and say
- 25:08like, OK, let's guide the AI on the things that we we need and,
- 25:11and correct, I mean. It takes, it takes us back to
- 25:14what we discussed earlier on, right?
- 25:15It depends on what the use case is.
- 25:17If it's simple things, and I don't mind talking to any AI
- 25:20agent, right? Because it's saving me time,
- 25:21it's saving the company time for whatever.
- 25:23But for more complex things, especially when the AI is
- 25:27limited in its capabilities, right, when it really cannot
- 25:30handle the tasks that I have, then it doesn't make sense to
- 25:33just keep conversing with an AI or, you know, sort of faking it
- 25:36and saying that this AI can do these things.
- 25:38So that's when you really need like the humor.
- 25:40Awesome. Anirudh, thank you so much for
- 25:42joining me. Is there anything else that you
- 25:44wanted to tell our audience? No, this is.
- 25:47I think it's a very exciting age that we are in.
- 25:49You know, everyone's trying to figure out the right way forward
- 25:53and that's the most exciting part about it, right?
- 25:55It's it's a very level playing field for anyone.
- 25:57So yeah, I mean, I'm very excited that we're building some
- 26:00of really, really good agentic products, AI products and it's a
- 26:04very exciting time even as a summer learning opportunity
- 26:07point of view. So yeah, thanks for having me,
- 26:08Laura. It's been great.
- 26:09So. All right, thank you, everybody.
- 26:11Thanks for joining us in today's episode.
- 26:14And I think the TLDR of this episode was it's humans plus AI
- 26:18make work easier, not one or the other.
- 26:22And if you enjoyed this content, please let us know, leave a
- 26:25comment, give us a like and we will see you back next time.