E295: Why AI Agents Will Quietly Replace 80% of Investment Teams

Why are humans — not models — still the biggest bottleneck to AI progress, and what happens when that bottleneck becomes a business? In this episode, I talk with Ali Ansari, Founder and CEO of micro1, about the hidden layer powering today’s AI breakthroughs: high-quality human intelligence. Ali explains how micro1 pivoted from an AI recruiting startup into a critical data infrastructure company for frontier AI labs, why expert-generated data is now the limiting factor in model performance, and what needs to change for AI agents to actually work in production. We also explore how focus, market timing, and ruthless prioritization enabled micro1 to scale more than 30× in a single year.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00David Weisburd?

  At a high level, how do you explain Micro One?

• 0:02Speaker 01

  Micro One is the AI platform for human intelligence. So what that means is we vet highly skilled people, mainly PhDs, professors, and industry experts, mainly in medical, finance, and legal, but also many other domains. And we help train frontier large language models. So you could think of, you know, the AI labs, the way they're kind of improving their model capabilities is by gathering net new human data for their post-training pipelines. And we help them gather that net new human data.

• 0:33David Weisburd?

  And who are your customers?

• 0:35Speaker 01

  Customers are the frontier labs that build foundational models. And we also have enterprise customers, you know, Mac Seven and kind of Fortune 500 broadly, that are building also foundational models, but also they're building enterprise agents that we help them evaluate and kind of get ready for production.

• 0:55David Weisburd?

  One of the reasons I wanted to chat today is because upstream of the LLMs improving, there's these improvements to the models. Maybe you could unpack on why are LM models improving? And how much of that is this recursive AI improving itself? And how much of it is the PhDs and these other professionals?

• 1:17Speaker 01

  It's almost entirely humans that teach the models in some way or another. Of course, that started with the pre-training phase where humans taught models by first creating the internet. Of course, that was the, you know, the largest set of human data that we that we had initially, which the models kind of took a unsupervised route of training. And, you know, that was kind of the initial initial state to the foundational models. And then afterwards, where the models really got useful is when humans kind of started to do a bunch of preference labeling, choosing which answer is better and so forth, based on the model responses. And then, once we passed that phase, now we're in this kind of expert data training,

• 2:02Speaker 01

  where humans are creating really complex data from scratch, whether it's doctors, lawyers, finance experts, and investment banking and other areas.

• 2:13David Weisburd?

  2025 was supposed to be this year of AI agents. Some people think it's gonna happen 2026. What needs to happen for AI agents to gain traction in the general market?

• 2:25Speaker 01

  Really, there's just one fundamental bottleneck that needs to be resolved. And that is enterprises need to dedicate large portions of their product budgets and really just implement in their product workflow. This notion of evaluations. And so what I mean by that is if you think about like, what does product development look like in any given enterprise or just any company in general? There's usually a phase of design. You design, you know, whatever software you're trying to build. There's some approval processes and then you get into development. The programmer, you know, develops it. There's full stack backend development, frontend development, et cetera.

• 3:04Speaker 01

  And then you put that into some QA engineering phase where there's a usually one QA engineer that goes in and kind of tests the products, says, okay, this works. And it's kind of a binary thing, like the software either works or it doesn't. And then, and then it goes into production. And that, that needs to change. And the, the, the part that needs to fundamentally change is the QA part. Where there's no more just one QA engineer going in and saying, okay, this software works and we can put it to production. But instead, there needs to be an evaluation framework for each of the actions that the probabilistic software needs to do. In other words, the agent needs to do. Because the agent, there's no notion of the agent works or it doesn't work.

• 3:46Speaker 01

  It, it's instead, what is the action space of this agent? What are all the things that I want it to do? And what are all the things that it should do? And, and basically what the experts do is they, you know, they create human data to measure exactly the capabilities of each of those functions. And then, once the, once the threshold is met, then the agent can move into production with confidence. What's happening right now is that there's a lot of good demos. Because if the agent works one out of five times or one out of 10 times, you, you'll just record that one out of five and it looks really impressive. But then it doesn't work four out of Five times and you cannot have that in production.

• 4:24David Weisburd?

  One of the hardest things of investing is seeing what's shifting before everyone else does. For decades, only the largest hedge funds could afford extensive channel research programs to spot inflection points before earnings and to stay ahead of consensus. Meanwhile, smaller funds have been forced to cobble together ad hoc channel intelligence or rely on stale reports from sell-side shops. But channel checks are no longer a luxury, they're becoming table stakes for the industry. The challenge has always been scale, speed and consistency. That's where AlphaSense comes in. AlphaSense is redefining channel research. Instead of static point in time reports, AlphaSense's channel checks delivers a continuously refreshed view of demand, pricing and competitive dynamics powered by interviews with real operators,

• 5:07David Weisburd?

  suppliers, distributors and channel partners across the value chain. Thousands of consistent channel conversations every month deliver clean comparable signals, helping investors spot inflection points weeks before they show up in earnings or consensus estimates. The best part, these proprietary channel checks integrate directly into AlphaSense's research platform, trusted by 75% of the world's top hedge funds with access to over 500 million premium sources. From company filings and brokerage research to news trade journals and more than 240,000 expert call transcripts. That context turns raw signal into conviction. The first to see wins, the rest follow. Check it out for yourself at alphascense.com/HowIInvest.

• 5:49David Weisburd?

  So another way for AI agents to scale, they need to behave like smart humans or ideally smarter than the smartest humans. In order to assess that, you need to have some framework in mind to assess the AI agent's performance versus a smart human.

• 6:06Speaker 01

  Exactly.

• 6:07David Weisburd?

  Every trend seems to have a killer app in the beginning with social, it was Facebook. With iPhones, some would argue it was Instagram. What's the killer app for AI agents?

• 6:18Speaker 01

  The obvious example is coding. I would argue actually the only use case that is very useful in production now. But I think that's a, that's actually an exciting thing that it's really the only one that's working super well. Because we've seen the immense amount of speed it's added to programming and really like how productive it's made. Software engineering in general. And so imagine kind of applying that same thing to essentially every other domain.

• 6:46David Weisburd?

  Sustained alpha is contingent on oftentimes having asymmetric information, having access to information or data. Other investors don't have. What are some early case studies for how investors are using AI? In order to get an information edge over the competition?

• 7:06Speaker 01

  So makers of private equity investors and you know, they're, they're creating LBO models or they're, they're manipulating them in some way. And models are getting quite good at that. So you know, what, what the, the data that we've been helping kind of a lot of foundational model companies create is, is around these, around these kind of manipulation of spreadsheets generally, which helps investors in their, in their day to day work, which allows them to again, work on the kind of higher level of thinking that any investment requires. What we do at Micro One is we try to kind of simulate this real world environments that investors usually work in.

• 7:48Speaker 01

  And so what we try to do is um, to get the models good at these capabilities, you have to try to replicate the, the same workflows that investors go through in terms of like the collaboration they go through and the kind of like multi expert, task creation that happens and the overall kind of peer reviews that happen in the process. Um, and so that's really the goal for us.

• 8:10David Weisburd?

  Now the model could take care of that. Now they could focus on which industries they want to go to, meeting the right people, meeting the right co investors, selling themselves to the investments themselves if needed, like in a venture capital and focus on higher level activities than just being in that model.

• 8:27Speaker 01

  That's exactly right.

• 8:28David Weisburd?

  You're part of this new generation of AI entrepreneurs, these AI native entrepreneurs. How do you look at building a business that maybe the previous generation? Built differently.

• 8:40Speaker 01

  We we pretty religiously follow this notion at Micro One, which is we have to try to get every function within the company to eventually have some AI agent that a human helps operate. And of course, there's a lot of functions where that's not remotely possible yet, but we have to still kind of strive towards it. And the company's overall velocity will be very much defined by this idea of, you know, how many agents exist within the company? And whether almost every function is not automated, automated is not the right word, but but kind of operated with humans running agents versus just humans doing the job on their own.

• 9:22David Weisburd?

  Let's say you're a private equity fund or venture capital fund in 2028 or 2030.

• 9:30David Weisburd?

  Give me an example of how a day to day might look like where humans are working next to AI agents completing tasks.

• 9:38Speaker 01

  Often, this is kind of explained as coworkers. And I would actually kind of disagree with this notion of coworkers. I don't I don't think AI agents are going to be coworkers. I think

• 9:53Speaker 01

  instead, what AI agents are going to be are kind of systems that actually change

• 10:02Speaker 01

  the domain of any given function. So what I mean by that is, you know, investment bankers are not going to have the same set of functions as the investment banking agent. Instead, the investment banking agent will take,

• 10:22Speaker 01

  you know, the investment banker kind of humans do currently. And what will happen is the investment banker, you know, human will only focus on that kind of.

• 10:35Speaker 01

  10% that really requires human creativity and focus. And the rest will be taken by that that agent, which the investment human kind of, you know, helps manage.

• 10:47David Weisburd?

  One thing that I've learned from talking to hundreds of investors, founders and business leaders is that planning for the future isn't just about growing assets, it's also about protecting the people who matter most. Life comes with a lot of important milestones. Maybe you're buying a home, starting a family, planning for college, or simply thinking more intentionally about your financial future. As those responsibilities grow, it's worth making sure you protect those you love. Not sure what kind of life insurance you may need? Concerned about how much it may cost? Think life insurance is too complicated? Your local State Farm agent can help answer those questions. They can help you explore your options for life insurance and choose coverage that you can adjust as your family's needs change over time. Financial protection for your loved ones when it matters most. Learn more about life insurance at statefarm.com. Like a good neighbor, State Farm is there.

• 11:32David Weisburd?

  In that future where AI is doing the work, what should humans be focused on? And how should they prepare for that future?

• 11:40Speaker 01

  It's gonna be a really, a really nice future. And the reason is, humans are naturally gonna find new things to do. What's gonna happen is the human is just gonna make their job more fun and come up with new things to do within their job. If you think about like, why does a human choose to do work every day? I mean, obviously, part of it is like to make sufficient cash and so forth. But, but, but in most cases, the other part of it is that it's like, actually like, pretty meaningful, like you're, you're doing something that you care about and you're impacting the world in like some, you know, cool way.

• 12:24Speaker 01

  And, and so like, I don't think humans are gonna wanna just stop that. They instead will do more of that because those functions will actually be, because what they do will actually be even more impactful because of agents. So I think what humans will do is they'll basically figure out ways to continue expanding on what they love doing, which will be their, their work in, in most cases. You know, what this means is that there's essentially gonna be net new functions created pretty rapidly by, by humans in every domain.

• 12:57David Weisburd?

  One of the concerns humans have is that Fear of losing meaning through their work. The second one is this Terminator case where the AI becomes sentient and becomes basically self-acting. What probability do you prescribe to that? Or do you think it's complete science fiction?

• 13:18Speaker 01

  It's a very unlikely case where models become completely, you know, ha- have the ability to completely learn on their own and also have the ability to kind of create versions of themselves and, you know, in some way reproduce. It's very, very unlikely that those two things become true to the extent that is true for humans. And without that sort of positive feedback loop existing, it's really hard for these systems to really get out of hand truly. So I think that's a very unlikely case, but but it doesn't mean that it's a case that we should

• 14:02Speaker 01

  kind of ignore. Safety evaluations is is a very important part of what model providers do, what what enterprises do and and and should continue doing. But I think it's really just that, like if you if you have sufficient budget and kind of care and efforts spent towards safety evaluations and and red teaming and so forth, then I think we will be just fine.

• 14:30Speaker 01

  And in fact, I think this is actually a really good area for the government to focus on. The Trump administration is doing a great job of like not slowing progress in AI in any way. I think they're accelerating it really nicely. But I would say like one area that the government should

• 14:45Speaker 01

  should probably focus on is actually this this this exact notion of coming up with a safety evaluation framework that requires a lot of like science and engineering to come up with good frameworks here that needs to be updated basically every day.

• 15:01David Weisburd?

  What's one piece of advice you wish you could go back four years ago and give a younger Ollie on how to better run Micro One, how to maybe avoid mistakes or or scale faster?

• 15:13Speaker 01

  One thing that I've actually realized quite recently is

• 15:19Speaker 01

  Market matters a lot. And I think, you know, being a very product oriented, you know, entrepreneur and really just caring about building a good product and and sort of assuming the rest will come, which is sort of true. And I and I like to believe that that's continues to be true, but

• 15:41Speaker 01

  I've come to a pretty important realization where the market you're in really matters. And, you know, the growth that we had was um by far last year when we decided to only focus on this application of human data and build this data infrastructure for labs. We were kind of split into these like bunch of different markets and long story short, we decided to focus on the application of the AI recruiter agent that we built, which was just human data and only focus on that, which of course meant we had to develop a lot of other things. It didn't, you know, it didn't stop at they are recruiter. We had to build the data platform and a bunch of other things that that came afterwards. But once we'd made that decision

• 16:27Speaker 01

  of of just focusing on this kind of one application where the market was really hot and like there was a lot of demand um in the market, the company, you know, more than thirty X than in one year, which was which was last year. Of course, like previous years to that, we we, you know, three X, five X, you know, whatever, like these these numbers were still good, but but twenty twenty five, we we literally more than thirty X. And and so so thi- this made me realize that we had focused on one specific application where the market really had demand and things blew up. So so the lesson is like really focus on, don't neglect focusing on the right markets.

• 17:05David Weisburd?

  What was upstream of that? You had to fire your customers and focus the team?

• 17:09Speaker 01

  So unfortunately, we had to stop serving um the customers in terms of startups that would hire engineers from us and and and other types of customers that we had. We, you know, had to stop serving them and slowly kind of phase them out in terms of being customers and only focus on on the AI labs and and, you know, the Mac seven that uh are building foundation models. And then we also started to focus.

• 17:36Speaker 01

  On building our product around exactly what the AI labs need.

• 17:42Speaker 01

  And so that kind of changed the product roadmap a good amount. And then I would say that the third thing is we made this decision to go all in on on data. It really it really changed the branding of our company as well. Like we were able to freely explain on our websites and overall kind of like sales materials that we are data infrastructure for for labs versus we're building a recruitment engine. And this, you know, this allowed us to actually close the labs like pretty quickly because of it.

• 18:12David Weisburd?

  So it just goes back to the innovator's dilemma. How in the world can a startup compete against a $10 billion company? And the thing that the startup always has is focus as the most finite resource. And if they could focus on one thing, then, you know, downstream of that, you could disrupt a 10 billion, 100 billion, trillion dollar industry.

• 18:33Speaker 01

  Exactly. And I think in these cases, focus is like the industry we're in is it's interesting because the reality is we actually have to balance how we focus. The focus is we are all in on data, you know, as I've said earlier, but but we also can't actually focus on any one data niche because of how fast these data niches change and how many different structures there are. Like, for example, if we focused on just finance data or just coding data, it actually wouldn't make so much sense because the same customers have so many different needs that they want to use a very small amount of vendors for. And if you focus on like one modality, you would actually not, you would not be a good vendor.

• 19:17Speaker 01

  So naturally, we have to build the product in a kind of paradoxically focused way where where the focus is actually to be able to vet all types of skillsets and and and build this like data platform that can actually handle all data modalities.

• 19:33David Weisburd?

  Running an AI company today is a practice in truly first principles thinking. How do you become a better CEO with such uncertain terrain in front of you?

• 19:44Speaker 01

  It's a good question. I uh, I am, you know, asking that every day. And one thing I do every single day is I, I try to cancel as many meetings as I can the next day. Like I look at my calendar and I question every meeting, like from the ground up. Doesn't matter when it was set, uh, maybe it was set a few weeks ago and it actually is not relevant anymore. And so I, I actually end up cancelling like roughly 30% of meetings every single day by just questioning them. And this saves me, you know, many hours a week. And so there's sort of like this notion of constantly questioning what I spend time on is probably the most important.

• 20:21David Weisburd?

  The best CEOs are always trying to get to ground truth. There are structural ways to do that. Elon basically removes all the middle layers. So there's an organizational structure, but also just getting to ground truth really means talking to the customers. And ultimately even more important than whether the product is good or not, is whether the customer is happy or not. It works best when those things are together. But getting to ground truth, which is the customer feedback, seems to be one thing that every single CEO that's scaling fast has in common.

• 20:55Speaker 01

  There's no alternative than the CEO and really the whole exec team talking to customers very frequently. I am practically an account executive at Micro One and, and it needs to stay this way for a while, especially because we have such a small amount of customers. It's the clients that we have, but it's also the experts that we have that, that actually help us kind of train these models and so forth. And we look at the experts also as customers. And so, you know, I try to be very close to our expert community. And these sort of things I think are, are important to really understand like what in this case, both of our customer types really want and what really makes them, uh, stay with the Micro One.

• 21:38David Weisburd?

  Ali, this has been absolute masterclass. Thanks so much for jumping on.

• 21:41Speaker 01

  Yeah, thank you, David. Thanks for having me.

• 21:43David Weisburd?

  That's it for today's episode of How I Invest. If this conversation gave you new insights or ideas, do me a quick favor. Share with one person in your network who'd find it valuable or leave a short review wherever you listen. This helps more investors discover the show and keeps us bringing you these conversations week after week. Thank you for your continued support.