Latest / The Learning Corner by Precursor / Episode #87: Patience Pays In Venture, Why Humans Beat AI, Managing Your Agent Workforce

Transcript

Episode #87: Patience Pays In Venture, Why Humans Beat AI, Managing Your Agent Workforce · The Learning Corner by Precursor

• 0:00
  If AI is so ready for prime time, why do we need so many forward deployed engineers?
• 0:04
  Wildly promising technology requires massive rethink of the way the companies work.
• 0:08
  This just to me reinforces the idea that this technology is gonna take longer to generate
• 0:13
  business value than people think because to get the real value, you have to, like, rip up and
• 0:18
  redo your processes.
• 0:20
  It's the week of July 20.
• 0:22
  Welcome to The Learning Corner, a Precursor Ventures podcast.
• 0:27
  Welcome to the Learning Corner by Precursor.
• 0:29
  I'm Mia Farnham, and I am joined by our GP and managing partner, Charles Hudson, where we will
• 0:34
  be talking through the articles and Goodreads that caught our team's attention over the past
• 0:38
  week.
• 0:39
  Our first Goodread this week is a blog post from Nick Charles from Asylum called Patience
• 0:45
  and on Owners versus Tenants.
• 0:47
  Charles argues patience isn't just a virtue investors should aspire to.
• 0:52
  It's actually a strategy to make more money.
• 0:54
  He points to Venture's growing obsession with speed chasing revenue at all costs, one to two
• 0:59
  year fund deployment cycles making a comeback, and metrics like time to 100,000,000 in revenue
• 1:05
  that ignore whether that revenue is any good.
• 1:08
  He borrows Jeff Bezos' owners versus tenants framing.
• 1:12
  Tenants churn through what they hold quickly while owners make decisions built to compound
• 1:18
  over years.
• 1:19
  And Nick backs this up with examples like Clay taking seven years to nail its product before
• 1:23
  scaling past a 100,000,000 in ARR, and OpenAI needing a full decade before Chachu PT became a
• 1:30
  hit.
• 1:30
  And his closing point is that patience doesn't mean moving slowly.
• 1:34
  It means being willing to look worse in the short term in exchange for building something
• 1:40
  that actually lasts.
• 1:43
  And, Charles, I have a bunch of questions here and want your broader Damn.
• 1:46
  Reaction, of course.
• 1:47
  But I think the largest point of tension in this piece here is how to align on what
• 1:54
  patience means between fund managers and their LPs.
• 1:57
  I think the original version of early stage venture inherently was tied to patients.
• 2:03
  And now the conversation has become what is a reasonable level of patience that LPs and then
• 2:08
  therefore GPs should have as it pertains to liquidity.
• 2:12
  But I'll pause for your broader thoughts first.
• 2:14
  I emailed Nick right after I read this.
• 2:16
  I'm glad he wrote it.
• 2:18
  It's funny.
• 2:18
  I went on when I was on, The Peel, I talked to Turner about this.
• 2:23
  There's this whole meme about, like, the game on the field in Venture.
• 2:26
  We talked about this.
• 2:26
  Oh, there's the game on the field.
• 2:27
  Can you play the game on the field?
• 2:29
  Should you play the game on the field?
• 2:30
  And my general view is it doesn't matter whether you wanna play it, but you must be
• 2:34
  aware of it.
• 2:35
  And I think about our model, which is to get in super early, usually pre product market fit.
• 2:41
  And I would say there have been times in the last fifteen years I've been doing some form of
• 2:46
  precursor where being super early pre product market fit and taking time to build things has
• 2:51
  been very much in fashion.
• 2:52
  There have been other moments where everyone's just like, I want just that water, instant
• 2:56
  growth.
• 2:57
  And I've been on the same journey that Nick explains in this post.
• 3:01
  You know, I went to a lot of our LPs and said, guys, some of these things that are getting to
• 3:05
  a $100,000,000 in revenue will come right back down.
• 3:07
  This revenue may or may not be durable.
• 3:09
  We don't really understand.
• 3:10
  And I think in a time of great uncertainty around what, you know, what moats are, what's
• 3:15
  gonna work in AI, things just out of people said, I'm just gonna chase the things that are
• 3:19
  the fastest rabbits.
• 3:20
  The things that get out of the gates quickly because to some extent, revenue is a signal of
• 3:25
  something.
• 3:25
  Whenever I ask people like, hey, Bubble or some of this stuff, they go, but look at the
• 3:29
  revenues.
• 3:30
  I go, look at them.
• 3:31
  Actually, at them.
• 3:32
  Look at, like, what underlies them, whether they're going to persist, whether there's any
• 3:36
  margin potential there.
• 3:37
  And there's almost no appetite for that conversation.
• 3:40
  And I think we live in this feedback loop where a lot of LPs venture as a part of their
• 3:46
  portfolio.
• 3:47
  It's not the whole thing.
• 3:48
  And in many cases, I meet limited partners whose primary interest is, are you in the hot
• 3:54
  companies?
• 3:55
  And I think in some ways, it's because the hot companies are easy to understand if venture is
• 4:00
  one of 10 things on your plate.
• 4:02
  You've heard of OpenAI.
• 4:03
  You've heard of Anthropic.
• 4:04
  You've heard of SpaceX.
• 4:05
  You've heard of Andoril.
• 4:06
  Maybe there's twenty com you've heard of Databricks.
• 4:08
  But maybe there's a great slow growth company that on your statement is still flat.
• 4:14
  And the investor is like, tell me why I should be excited about this company.
• 4:17
  It's been flat for five years.
• 4:18
  Replit.
• 4:19
  Like, are a bunch of these companies.
• 4:20
  And I think there's very little appetite for that narrative right now, which means if you're
• 4:25
  gonna pursue that strategy, you need patience yourself, you need fortitude, and you need
• 4:31
  people who will help keep you in business for that good news to turn out to be true.
• 4:35
  And I think this also has trickled down impacts for founders because VCs are responding to the
• 4:41
  institutional incentives, which I think right now are to be in the hot, visible, high growth
• 4:46
  things.
• 4:47
  Some of them will work out, some of them won't, but they're all it's funny.
• 4:50
  We talked about what's legible to capital, that whole legibility thing.
• 4:53
  And I think we're now in the same era for VC managers, which is like, well, what's legible
• 4:56
  to LPs?
• 4:57
  And I think legibility is just a huge shortcut to stop you from asking hard questions and
• 5:01
  thinking.
• 5:02
  Oh, this person went to Stanford and worked at OpenAI, so, like, they're probably smart.
• 5:06
  Maybe.
• 5:07
  I mean, yes, they're smart.
• 5:08
  It doesn't mean the the startup's gonna work, but it's easy to underwrite.
• 5:11
  And I think in some ways, what Nick is talking about is the extension of this legibility
• 5:15
  concept to fund managers, which is like, hey.
• 5:17
  If you're in a bunch of names I've heard of, whether I have an opinion on those names or
• 5:20
  not, you're in things I recognize and know.
• 5:22
  And that's easier to talk about it at a committee than a company that a manager has
• 5:26
  great belief in, but it's taken five to seven years to show anything.
• 5:30
  Yeah.
• 5:31
  And I think that's, I guess, the even more deeper tension is the conundrum of maybe the
• 5:36
  arbitrage opportunity is to be patient right now while everyone else is racing ahead to
• 5:42
  deploy in into things that have a 100 plus million revenue that's maybe not defensible.
• 5:46
  But will you get the opportunity to be patient in the time required for you to kind of end up
• 5:50
  six, seven years later, and perhaps you're right?
• 5:52
  That's what I think about a lot.
• 5:55
  Like, who's the patience, yes.
• 5:56
  I think, like, if everyone could just take a beat and be like, wait, should we just take
• 6:00
  stock into what we're actually deploying into?
• 6:02
  I think most people would say, yes, that's a reasonable thing to do.
• 6:04
  But you have to deploy money.
• 6:06
  Fund managers need to recognize or LPs want would like to recognize and and showcase the
• 6:11
  great portfolio they have.
• 6:12
  So do you get that chance and leisure to be patient?
• 6:15
  I think sometimes is the question in my mind.
• 6:17
  Feels like patience is a luxury right now.
• 6:19
  Yeah.
• 6:20
  But also, I think a lot about I have a handful of my friends who are hedge fund managers or
• 6:26
  other public markets people.
• 6:27
  Go, look, the market can remain irrational longer than you can remain liquid sometimes.
• 6:31
  And I think that's the equivalent of patience and venture, is like if you're really patient
• 6:35
  and right, you might go out of business as a fund manager before you're proven right.
• 6:40
  You'll still make money.
• 6:41
  You just won't have a fund anymore.
• 6:42
  There are examples of this in my history as a fund manager where funds that were counted out.
• 6:47
  LPs didn't back them.
• 6:49
  The people sat there and managed out the tail, and the tail ended up being pay partners is is
• 6:53
  one example I could think of.
• 6:55
  The tail ended up being very, very interesting, and people had sort of written written it off
• 7:00
  because it wasn't performing on the time cycle that people wanted.
• 7:05
  Yeah.
• 7:05
  And so I don't know.
• 7:06
  I I think about that a lot in our business, is like there's this tension between the things
• 7:13
  you need to explain and show to limited partners to continue to earn their trust and
• 7:17
  dollars to stay in business and the things that will actually work and generate returns for you
• 7:22
  in the future.
• 7:23
  I do think the one thing that's very clear is, to your earliest point, being able to see the
• 7:27
  game, understand it, and know when to participate as a fund, knowing all these
• 7:31
  different dynamics at play.
• 7:32
  Yeah.
• 7:32
  I do wanna go back to how this impacts founders because I was thinking about this through the
• 7:36
  lens of investors, of course, and the relationship between LPs and GPs.
• 7:39
  But patience is obviously something that is inherently important for founders in this age
• 7:44
  too, and certainly media hasn't done a great job in terms of making people feel like being
• 7:48
  patient can be rewarded.
• 7:50
  I also think about it as people who work at startups right now as well as nonfounders,
• 7:55
  people who see things on LinkedIn or different media outlets and see like, well, my company is
• 8:00
  not growing at a 100 x year over year.
• 8:02
  Should I leave?
• 8:04
  And I think, generationally, this is something that a lot of my peers have been talking about
• 8:07
  is, like, the itch to feel like they have been working for a while and not seeing the instant
• 8:12
  reward.
• 8:12
  And I continuously come back to the patience piece of, like, I think most of the stories
• 8:17
  that you read are maybe there's, like, one in a million that are truly anomalies, and I think
• 8:22
  that's just what true venture are.
• 8:24
  The rest is a lot of posturing, and kind of just optics to make things seem a different
• 8:30
  way, but it does trickle down to every facet of our ecosystem.
• 8:33
  In the last ten years, gambling and speculation, I'd argue starting with crypto,
• 8:38
  but you could go crypto to sports betting to prediction markets.
• 8:42
  It's just become a bigger part of social life.
• 8:45
  Yeah.
• 8:46
  And I think investing in patience is about getting more later by giving up something now.
• 8:52
  And I think part of this is just a change in social values and social interest and the
• 8:58
  feeling that long termism is for suckers.
• 9:00
  That's the old playbook, and it doesn't work anymore.
• 9:02
  And if you're still following it, you're a sucker.
• 9:05
  And I think that's, like, a very damaging social value because there are lots of things
• 9:08
  that take time.
• 9:09
  But I also understand the impetus for the people who feel that way given the world that
• 9:14
  they grew up in.
• 9:14
  And I think also it's hard for startup founders.
• 9:16
  I talked to someone in our portfolios.
• 9:17
  Like, I tripled this year.
• 9:18
  I go, nobody cares.
• 9:19
  And he goes, what do mean?
• 9:20
  I go, hey, man.
• 9:21
  Five years ago, tripling would have been amazing.
• 9:22
  And I don't wanna diminish how hard it is to triple a business.
• 9:26
  And when I say nobody cares, I mean the people you wanna go raise money from are gonna see
• 9:30
  your triple, and it's not gonna make them jump out of seats.
• 9:34
  They're gonna say, well, you tripled.
• 9:35
  I have a company that's three x ing month over month, not year over year, month over month.
• 9:39
  And so that's what I'm looking for.
• 9:41
  There is a collective obsession with these, like, weird then if I see another one of these,
• 9:47
  like, nine charts that all have this, like, super vertical shape, and, of course, the start
• 9:52
  up that's being talked about is the leftmost line Of course.
• 9:54
  For the fastest, you know, to a $100,100 dollars in thirty minutes.
• 9:58
  We're the fastest start up in the history.
• 9:59
  Adam, whatever.
• 10:00
  It's gonna be something silly.
• 10:02
  That's where people's energy and attention is, and it's very hard to fight Yeah.
• 10:06
  Unless you have a really good emotional and financial support system.
• 10:10
  Well, our next good read is a piece from David Deming's Fort Lightning Substack titled
• 10:15
  Efficient Social Learning is the Human Advantage over AI.
• 10:19
  David makes the case that as AI gets more capable, humans will keep a real edge in one
• 10:24
  specific area, learning efficiently from very little data, especially in social and
• 10:29
  relational tasks, something that we talked about a month ago, I believe.
• 10:33
  He walks through research showing people can infer someone's preferences and motivations
• 10:37
  after watching just a handful of interactions while large language models still struggle to
• 10:42
  carry what they learn about one person over to a new context.
• 10:46
  The core economic idea is comparative advantage.
• 10:49
  Since AI will eventually get cheap and commoditized, the tasks where sample efficient
• 10:54
  human judgment matters most, reading a room, building trust, adapting to one specific
• 10:59
  person, become disproportionately valuable.
• 11:02
  He's really careful to separate this from the question of whether people will pay for a human
• 11:07
  in the loop, which he calls a very separate and unresolved debate.
• 11:11
  His point is really about where human labor still has a structural edge regardless of what
• 11:15
  customers end up wanting.
• 11:17
  And David ties it back to evolution, arguing humans got good at social reasoning because our
• 11:21
  survival depended on transmitting knowledge to each other, not because we practiced abstract
• 11:27
  reasoning, which is actually the newer and harder skill for us.
• 11:31
  I love this piece.
• 11:32
  Aliyah from our team shared it, and I thanked her for doing so.
• 11:36
  One line from David really stuck out with me.
• 11:40
  Quote, the context of a single relationship is constantly evolving and requires multimodal
• 11:46
  learning from accumulated common ground.
• 11:49
  Social learning doesn't scale easily.
• 11:51
  Building a deep relationship with one person doesn't let you skip steps with others.
• 11:56
  And he makes the point that LLMs are also decent at predicting how a rational person
• 12:02
  should behave, but fall apart the moment someone acts irrationally or out of character.
• 12:07
  So wanted your, obviously, your broader reactions to the piece, but wanted to know, do
• 12:11
  you think that's the real lasting ceiling for AI in today's world?
• 12:14
  Or is it more that these models just haven't been trained enough on the messy, super non
• 12:19
  relational parts of how people actually behave yet?
• 12:22
  I read it, and I was like, well, this is a grounded optimistic take of how this all plays.
• 12:29
  Hopefully.
• 12:30
  Yeah.
• 12:30
  I thought the idea of comparative advantage is a good principle to introduce, which is like,
• 12:35
  yes.
• 12:35
  Even if the LMs have an absolute advantage in everything, there'll be some areas where having
• 12:41
  humans involved still makes sense.
• 12:43
  And I do think, one part of this article that I really agree with is there are places in the
• 12:48
  world already where access to humans is a you know, if you if you're a high status airline
• 12:53
  flyer, you're you don't have to navigate an AI tree or voice prompts if you don't want it.
• 12:58
  You will get a human on the phone.
• 13:00
  And oftentimes, the human it's funny.
• 13:02
  I, you know, I read a lot more code now, and there are sometimes I'm just like, that took
• 13:06
  way too long.
• 13:08
  If I had explained that task to a human, they wouldn't have gone through the process of
• 13:11
  breaking down every single step, then and in the process, they would have said, I think I
• 13:15
  know what you mean.
• 13:16
  Or, hey.
• 13:16
  Let's go back and forth.
• 13:17
  And I I read it.
• 13:17
  I was like, you know, a part of what I think makes ironically the most successful companies
• 13:22
  that I work with successful is the ability of leadership to communicate complex ideas up and
• 13:27
  down the chain through to and through people, and those people's ability to absorb a message
• 13:33
  and figure how to translate it to the next person in the chain.
• 13:36
  And I think there's a lot of brute force tasks that LLMs are just good at because once they
• 13:42
  have learned how to do them, the variance surface is known.
• 13:46
  And I have some of those tasks running already at precursor.
• 13:49
  I'm just like, 99% of the time, the known outcomes for that task are fully mapped.
• 13:54
  And the 1% that aren't are the things that the team and I spend time on.
• 13:58
  There are other things I think about, like, most of the founder calls I have.
• 14:01
  I have a vague idea of what we're gonna discuss when we get on the phone.
• 14:06
  I don't always know where they're gonna go Yeah.
• 14:09
  What topics they're going to bring up.
• 14:11
  And I think a lot about I was reading something else that that I think that the NextView team
• 14:16
  had written.
• 14:17
  And I was like, yes.
• 14:18
  I do the same thing all the time.
• 14:20
  I talk to 30 founders in our portfolio every week on Zoom or in person.
• 14:26
  Those those those calls have the same structure.
• 14:29
  They have nothing in common.
• 14:30
  In two dimensions, one, every company has a different set of issues, and more importantly,
• 14:35
  each founder is their own computer.
• 14:38
  They're their own emotional computer with their own idiosyncrasies and their own beliefs and
• 14:43
  their own strengths and weaknesses.
• 14:45
  And my view is, like, the point of that meeting is for me to do the mapping between, like, who
• 14:49
  is this person, what is this issue, and what am I trying to communicate to them, and that is,
• 14:55
  like, not scripted.
• 14:57
  Yeah.
• 14:58
  It's like a little bit of intuition.
• 14:59
  There is some pattern matching.
• 15:00
  Like, hey.
• 15:01
  I've had 30 meetings with this person.
• 15:02
  I know they get triggered by x y and z thing, or I've had 30 meetings with this person.
• 15:07
  I know that they can take direct feedback without any kind of cushion on it.
• 15:13
  But there are other things that come up randomly that are really de novo, and you're
• 15:17
  like, well, what I'm really doing is synthesizing all the information I know about
• 15:21
  this problem and this person and contextualizing it on the fly.
• 15:25
  I'm not sure that LLMs are great at that task yet.
• 15:29
  Going back to your example of strong leadership, being able to kind of tell people
• 15:34
  what to focus on and that being such a high highly valuable skill, is that one in the same
• 15:39
  to managing people?
• 15:42
  Because I read this piece then I was like, oh, are good people managers now going to be, like,
• 15:46
  the most valuable people in your organization?
• 15:48
  Now that's obviously with the belief that, like, people will continue to work in your
• 15:51
  organizations and Yeah.
• 15:52
  Just be a bunch of agents.
• 15:54
  Increasingly, when I talk to our portfolio companies, I ask them, like, well, what what
• 15:58
  mistakes are you making with AI?
• 16:00
  Like, oh, we built the wrong thing.
• 16:02
  I'm like, do you mean?
• 16:02
  Like, well, we built something, and then we got three quarters of the way down the road, and we
• 16:06
  realized, like, it wasn't right.
• 16:08
  I'm like, oh, so you had bad product management.
• 16:10
  Or, oh, we built a bunch of things that customers ask for, and now we've lost all
• 16:14
  product coherence.
• 16:16
  And I've always felt like part of what good human leadership does is it creates a shared
• 16:21
  understanding of the job to be done among a group of people.
• 16:24
  Right now, people are devaluing that step because it's so fast and easy to build, but I'm
• 16:30
  like, you still have to build the right thing.
• 16:32
  Yeah.
• 16:33
  And I think part of what a good leader does is they spend time with people talking through
• 16:37
  this is what we're building and why, and they create space for, like, alternative viewpoints
• 16:42
  to come up and get debated and maybe integrated or dismissed.
• 16:45
  Or for other people to say, why aren't we talking about this element of it?
• 16:48
  And people go, oh, yeah.
• 16:49
  You know, you're you're you're right.
• 16:52
  As one of our founders has been helping me with a software development project, and he's like,
• 16:57
  you figured it out.
• 16:58
  By the time you build the first version, you know what you actually should've built.
• 17:02
  And I think the what to build, how to build it, and the consensus building, as long as there
• 17:07
  are people involved, I think that will become an even more important skill.
• 17:11
  Because pre AI, you had to do all of that work before you started because building was
• 17:15
  expensive.
• 17:16
  And now because building is cheap, a lot of people are like, well, let's just skip all that
• 17:19
  and build.
• 17:20
  And I'm like, well, you're gonna end up with the same problem just at the back end instead
• 17:23
  of the front end.
• 17:24
  And all the same workarounds Mhmm.
• 17:26
  And time required to fix.
• 17:27
  Yep.
• 17:28
  Well, our last Goodread this week is a post on X from George Sowellka called, You Just Hired a
• 17:34
  Million Bad Employees.
• 17:36
  George draws a direct parallel between agent workforces and human workforces, arguing the
• 17:42
  two break down in exactly the same ways so we can use decades of management lessons to fix
• 17:47
  how we run AI agents today.
• 17:49
  This is a really good tie to what we were just talking about, actually.
• 17:52
  He walks through a set of specific pairings.
• 17:54
  Token maxing is just throwing bodies at the problem.
• 17:57
  Loops are meetings about meetings.
• 17:59
  Wasted tokens are the new headcount bloat, and context hoarding is the modern version of
• 18:04
  employees guarding tribal knowledge as job security.
• 18:08
  He argues the fix in both cases is really the same.
• 18:12
  Define what good looks like, which is why he thinks evals are becoming the new OKRs.
• 18:17
  Coding to what you just said, Charles, is the one function where AI has actually delivered,
• 18:22
  and that's only because coding has built in evals.
• 18:25
  Coding either compiles and works or it doesn't.
• 18:28
  While most other business functions have no equivalent way to measure success, George also
• 18:33
  raises a political problem.
• 18:34
  Employees who hold the most valuable tacit knowledge have the least incentive to teach it
• 18:38
  to an AI system that might replace them.
• 18:41
  His bet is that the biggest winners won't be new AI native startups, but transformation
• 18:46
  companies that go deep enough into an existing business to actually build the evals and
• 18:51
  workflows needed to make agents reliable.
• 18:55
  My favorite parallel he talked about in his piece was about loops being the equivalent of
• 19:02
  meetings about meetings and Yep.
• 19:04
  Simply people just not knowing what to ask for or what success actually looks like.
• 19:09
  So you're really just stuck prompting in circles.
• 19:13
  I didn't know if you had a favorite comparison that George mentioned or one that stood out to
• 19:18
  you the most.
• 19:18
  I think people's, like, utopian idea like, I tell you, oh, we have this self managed, self
• 19:24
  running company.
• 19:25
  I'm like, but is there anybody checking to decide if it's, like, heading in the right
• 19:28
  direction or making good decisions?
• 19:30
  And everyone's like, full autonomy.
• 19:31
  I'm like, I run a bunch of agents in semi autonomous mode because I like to intervene.
• 19:37
  In many cases, I'm like, oh, that's close, but not right.
• 19:41
  That is not that does not mean my quality shipping bar.
• 19:43
  Let me shoot.
• 19:44
  In the same way I would help an intern be like, hey.
• 19:47
  We wouldn't publish that or we wouldn't ship that at precursor.
• 19:49
  It's close, but let me tell you what is required.
• 19:52
  That's, I think, the state of the art in tech.
• 19:54
  And I I keep wrestling me with this, like, fundamental contradiction that I feel like
• 19:59
  whenever I mention it to people, they give them this puzzled look.
• 20:01
  I'm like, if AI is so ready for prime time, why do we need so many forward deployed engineers?
• 20:06
  They're like, well, they're not really service peep they're not, like, traditional because and
• 20:10
  I'm like, fair.
• 20:11
  But, like, what are they doing?
• 20:13
  Yeah.
• 20:13
  They're like, well, they're helping these companies with business transformation.
• 20:16
  I'm like, yes.
• 20:17
  The companies can't adopt this stuff on its own, which is always the case with new
• 20:22
  technologies.
• 20:23
  I'm like, what if we're closer to where I think we are than other people think we are, which is
• 20:26
  wildly promising technology requires massive rethink of the way that companies work to get
• 20:31
  full value, and companies don't have the internal sherpas to do this themselves Yeah.
• 20:36
  At scale.
• 20:37
  So either the vendor has to become basically the transformation agent or you need a Deloitte
• 20:43
  or a McKinsey or a third or OpenAI or Yeah.
• 20:47
  You need an FDU shop.
• 20:49
  And I'm just like, this just to me reinforces the idea that this technology is gonna take
• 20:53
  longer to generate business value than people think because to get the real value, you have
• 20:58
  to, like, rip up and redo your processes.
• 21:00
  You can't just drop an agent into your there again, there's some discrete functions,
• 21:05
  customer success and a few other things, customer support, and that are discrete where
• 21:09
  you could drop an agent and say, we're just not gonna necessarily do a total transformation.
• 21:14
  And I just think about, like I mean, I think about what we talked about in a previous
• 21:18
  episode about Sierra.
• 21:19
  Yeah.
• 21:19
  They're in this business, and they're very smart people.
• 21:22
  And even they were like, this is hard, and we're in the business of selling AI powered
• 21:26
  customer support to other people.
• 21:27
  Mhmm.
• 21:28
  And so I'm like, well, what how much harder would it be for a company that doesn't have the
• 21:31
  same technical resources and isn't starting basically with a blank slate?
• 21:35
  For most companies, I think the issue is they don't know what they don't know.
• 21:38
  So the forward deployed model works perfectly of just like, I'm gonna have people sit with us
• 21:42
  day to day and almost, like, coach me on what good looks like or what it can actually become.
• 21:48
  But then probably, like, halfway through the project or the deployment or the transformation
• 21:52
  project, the folks who are around the table were actually at the actual org are like, oh,
• 21:56
  wait.
• 21:56
  Now I understand what you're building.
• 21:57
  I have all these edits because I had no idea and context of what was being created.
• 22:02
  I mean, again, consulting gets to live another bunch of years thinking about it more broadly
• 22:06
  of this is exactly what that industry was built upon, which was not staff aug, but let me carry
• 22:13
  the burden of figuring out this mess, and I'll bring you in at the right times to get your
• 22:17
  perspective.
• 22:18
  And I think we're seeing every organization, small or large, now feel like they require that
• 22:23
  outside skill set to be the sounding board as they navigate what they wanna do here.
• 22:27
  One thing that he mentioned in the beginning that I disagreed with a little bit was his
• 22:33
  quote that said, managing AI is harder than managing people because AI scales dysfunction
• 22:40
  instantly.
• 22:41
  And in my mind, I was thinking, sure.
• 22:43
  Like, if what you've built is distributed across an entire set of integrations on a pure
• 22:48
  task perspective, then it's rather difficult.
• 22:52
  But I find unwinding that dysfunction to be quite easier.
• 22:56
  I guess I'm only speaking from my experience with a smaller org, not talking about working
• 23:00
  for like a thousands of people org.
• 23:03
  I do find undoing years of dysfunctional teams is nearly impossible to pull off.
• 23:09
  You're almost better off just firing everyone and then starting from scratch.
• 23:14
  Didn't know if you had a perspective on that one as well.
• 23:16
  I think, what I found is that the more autonomous the AI, the more problems and
• 23:21
  trouble it can create.
• 23:23
  And, it's not always good at telling on itself.
• 23:26
  And so I think the difference is, like, with humans, we've built in all of these checkpoints
• 23:30
  on work that are generally designed to limit the amount of damage someone can do.
• 23:36
  And the more trust you have in somebody, the less often you check on their work.
• 23:40
  But that's usually an earned right.
• 23:42
  A lot of the AI is like, we're just gonna go off and do this thing.
• 23:44
  And, if you don't have eval if you don't have, like, if you don't have software driven
• 23:49
  checkpoints, you can look up and be like, oh, wow.
• 23:51
  This thing has, like, been off the rails for a very long time, and nobody noticed.
• 23:57
  And sometimes because it's confident and wrong, it's not gonna tell you, I think I screwed up.
• 24:03
  It'll just be like, just processing just crank it away over here, sir.
• 24:07
  And you have to look under the hood, and I think there's this belief that, like, oh, these
• 24:11
  things, once you set them up, they're they're not set it and forget it.
• 24:14
  The model's changed.
• 24:16
  Sometimes, like, you accidentally inject new stuff.
• 24:18
  Like, things break.
• 24:20
  And I find that, like, because AI agents basically will work at whatever frequency you
• 24:24
  tell them unlike humans, sometimes they'll just run and run and run, and they can they can run
• 24:28
  for weeks before you realize, oh, this thing's broken.
• 24:30
  It's been broken for weeks.
• 24:32
  That would almost never happen with a human.
• 24:34
  Yeah.
• 24:34
  Like, the thing that makes agents so wonderful on paper that they'll never burn out and they
• 24:39
  can run things at scale, like, is the limit.
• 24:42
  They they don't carry guilt or internal stress in a way that if you're a lower level employee
• 24:50
  and you have questions or you know something is running poorly or bad, you'll think about it
• 24:56
  overnight, every night, until you finally get it like, a manager will get it out of you.
• 25:00
  Yeah.
• 25:01
  What do you actually think is going on as opposed to just executing it like, here's the
• 25:05
  whole list of things I have concerns over that are not letting me sleep, and I don't think I
• 25:10
  can push this project forward if I don't surface them to you.
• 25:13
  Like an agent, like, why would they ever bring that up?
• 25:15
  They just wants you to continue to use it.
• 25:17
  Also, think I basically have management for my agents now.
• 25:20
  I'm like, okay, Claude code, help me write code.
• 25:23
  But Codex is the VP of engineering, and we don't ship anything without Codex.
• 25:26
  And Codex always finds something
• 25:28
  Yeah.
• 25:29
  That it doesn't like.
• 25:31
  That's usually an improvement.
• 25:32
  Then I'm like, cool.
• 25:33
  We're gonna run this for a little while, then we're gonna, like, check on it.
• 25:35
  So I'm like, I manage my agents the way I would manage, basically, a junior employee.
• 25:41
  The difference being that this junior employee can work twenty four seven, but it still makes
• 25:46
  mistakes.
• 25:46
  And, again, it to your point, it has no guilt or shame.
• 25:49
  So sometimes it makes mistakes, and it's like, oh, I I made I think I made a mistake, but I
• 25:52
  just, like, pushed through it.
• 25:53
  Yeah.
• 25:54
  And I'm like, that that isn't what I wanted you to do.
• 25:56
  Yep.
• 25:56
  Well, we are all out of time today.
• 25:58
  Thank you, Charles, for jumping in.
• 25:59
  Thank you to our listeners for joining as well.
• 26:02
  Please like and subscribe, and we will talk with you all next week.
• 26:05
  Thank you.
• 26:05
  Bye.
• 26:06
  Bye.