AI Transformation: Will Consulting Be Replaced by Autonomous AI Systems?

Thanks, Utsav it's a great intro.

Well, I think ⁓ we're recording this ⁓ a after, two weeks after the Sequoia article that just came out that next legandry company is going to be a ⁓ tech company that ⁓ is cloaked as a services company. ⁓ I find it a little peculiar that people are now catching up to it because almost exactly two ago, NFX ⁓ said ⁓ there's an article NFX that says, ⁓ ⁓ Basically coined the term services as software not service ⁓ Service as a sorry software as a service but service as a software, right and that basically the new ⁓ AI age is going to be taking ⁓ Taking budget from the labor market which is about five trillion dollars in their estimation and putting it into these new services, right? So these new classes services. So even two years ago I was looking at it like, okay, this is where we need to go. Where I think the Sequoia article is on point right now is that we have gone through the cycle of building a platform, a marketplace, etc, ultimately to go back, would say, was talking to a friend today, to get to about 360 degrees back at the very beginning, we're like, okay, this actually needs to be sold as consulting.

This needs to be sold. the same way people are used to buying it because there's so much noise in the tooling space versus you need to deliver outcomes anyways and customers buy outcomes anyways. ⁓ the behaviors of I'm buying this outcome and there's a person that I can talk to and there's a lot more other people, the analysts that do all the work, etc. I don't really get to talk to them that much. I talk to this one person that sold me that thing and I trust and I'm going to yell them. yell at them when things don't go well, The accountability of it, that's still there. People still want it, right?

So it's almost immaterial how it gets done as long as you can trust that it's done very, very well. So for me, the journey over the past actually three years is kind of going full circle to a model that's become very similar to what we used to do, but much larger consultancy with much better margins.

Well, I started coding and designing when I was 14. I'm a product guy. And I always say I randomly became a CEO of an agency we or consultancy. ⁓ We became over a hundred people and that's not what I kind of set out to do. I set out to kind of build systems.

And when we got to a place where we understood this is how you do venture building and we had a kind of a meta process that worked across any industry. ⁓ I was like, basically I lost. interest in it, right? Like I solved that problem, but the problem was, ⁓ you need people, you need to always hire more people as like the same process, right? ⁓ ⁓ before Chat GPT, the first version came out, ⁓ I had drafts of how I want to transform, ⁓ you plus that was the consultancy ⁓ something like a, like a Fifth Row right now. ⁓ But the tech wasn't there. So I was kind of waiting for the stack to come in. So my frustration was like, we keep doing the same thing, the same process.

It has to be productized. How can we productize it? And then about three years ago, think February, February, 2023, I started seeing things that my friends brought me from Open AI hackathons in the Bay area. And I'm like, okay, this is my whole team.

Like this part of my team that, you know, is doing this for two weeks. And now it was done in like 30 seconds. Of course the quality was so-so, but you could kind of see the trajectory that was going. So, so that was the main impetus for. for us going trying to build ⁓ first version of Fifth Row which was venture building automation, just our own process. ⁓ then it scaled from there.

I think the shifts are continuing and continuous, right? So as new capabilities come out, everybody, I feel like everybody is, it's always big shifts. We work with Fortune 500s mostly, large organizations in the US or Europe. discussions are always, let's say three to six months, maybe nine months behind the bleeding edge. So what we have to do is we have to... ⁓ translate that and translate the FOMO of, but Claude, co-work exists now, right?

But you have processes from 19 whatever, right? So what are we gonna do about this? So we increasingly find that it is no longer a challenge for people to believe that these technologies can meaningfully replace the work of consultants or analysts. ⁓ But you have to redesign the processes, right? And you need to redesign the organization.

So that becomes a topic discussion a lot more. Plus, ⁓ it becoming a lot more proactive and always on, right? So instead of, Hey, I have this like venture building thing that I can run anytime. It's like, no, it's running by itself, right?

It's like a thing that I have now that I didn't have before that I previously would have hired consultants, external, external people, internal people, or just. bought some reports as a snapshot. Ultimately it comes down to how do we make better decisions faster, shorten the cycle time so that we are just faster in doing business than our competitors. So it's the process change and it's the always on-ness of these systems.

So in terms of the models, it's, ⁓ you know, early on, nobody knew anything, right? So everybody was just trying to build something. We had the advantage of having, you know, 40 experts that did the thing every day, ⁓ basically writing prompts and building it. summer 2023, we basically built very large prompts 40 times over and ⁓ made it a production. The one thing that we, the one thing that we've done from the very beginning was, We always build a custom RAG a custom database of sources that we scrape just broadly from the internet. we knew that that's the thing that we have to do because our customers are immediately going to ask us like, where's this from?

Like, how do know that's true? So, and that evolved even now about 50 % of our infrastructure costs overall go to fact checking, right? So we found one thing like if people are okay, atleast enterprises are okay to wait, if for it to be correct. Like it has to be correct.

The correctness, the provenance, where things come from sourcing explainability of what's going on is the number one thing. It's, it's, it's table stakes. If you don't have that and none of like Chat GPT cloud, etc still don't have it. Right. ⁓ for enterprise, this is something that we spent a lot of time on early on ⁓ we of course had customers where it worked really well.

Customers where ⁓ were just blindsided by behaviors of the models. Right. So we had to kind of build it up. ⁓ ⁓ the, that was like one of the core tenants. So like, how do we know that this is coming from a reputable source, that it's good, etc.

So we had to basically, we built like a version of deep research in 2023 Uh, and the first version of fifth row was like 50 deep, deep research is right. Kind of strung together. And, and until I guess a year and a half ago, like, uh, September, 2024, nothing went out at a platform until a human saw it. Right.

So we had a human Q & A layer, right? So we would check for all the hallucinations, make sure that everything makes sense, etc ⁓ was easy to do then, ⁓ because we had one process, right? Like it was one app. But increasingly ⁓ we ⁓ the venture building app, we, the venture building app basically starts with research, goes through ideation, then ultimately build a landing page and drive traffic to it.

Right. And if, ⁓ and if people sign up, then you have proven it with much less cost, but then you can go and talk to the people. Right. This is something that we would have charged half a million dollars, 12 months, sorry, 12 months, half a million dollars, 12 people, three to five months, depending on how long it would take.

And now it took two hours, right, to prepare it. And then the market test would run for a month, but you could run a lot of them in parallel. So that proven to be like superhuman very quickly. Basically we got to results by the end of 2024 that in 87 % of cases we outperformed the cost per signup. on a new venture built on first try, right?

By doing all of this research, just taking our consulting process, packaging it to, it ended up being about, I think, 60 agents, right? And running it. So we did a lot of work on that orchestration. So we started, so that was part of the reason why we said, okay, well, U Plus has been doing this.

There's like no reason for it to exist. Let's focus on Fifth Row. So that's where we kind of switched and Fifth Row became. the butterfly out of U Plus And then I think around, yeah, like mid 2024, customers started coming to us and said, well, this venture building thing is great. And we understand what you did.

It's super impressive, but we don't do venture building. We do startup scouting. We do due diligence and regulatory compliance and market intelligence and whatever, SCC filing analysis and things like that. We have that process like you had for venture building.

Can you take that process and can can you automate it for us as well? It's like, okay, this is interesting. It turns out we did not automate venture building. We automated how to do consulting at scale.

So let's, let's actually dive deeper into that. So in terms of the models, right? The models started getting better and we had this big discussion with my CTO. is two years ago now when he said, well, we need to have more analysts at all of these different verticals. And like, but the LLMs know all of it.

Like it was trained on everything, right? Like if you ask, if you know how to ask, it's going to tell you what the process for strategic foresight or regulatory monitoring or whatever is, what is the best in class, McKinsey BCG, whatever process, right? ⁓ So that was one thing, one key decision. The other key decision was it needs to be able to, when it makes mistakes and need to be able to see it and adjust, right. And continue working.

So we started working with autonomous agents, what now everybody knows, right? But two years ago, nobody, barely anybody knew. So we started working on that, but it was GPT-4 turbo, I remember, like expensive and not that turbo. And then in June, 2024, if I remember the dates correctly, 4.0 came out.

It was for the first time that it was cheap, fast, and intelligent enough so these agents could start working. And we could suddenly describe like a process, like a generic process. It's like, this is how you do strategic foresight, right? And it will be able to go and do it, right?

And then we'll have its thoughts and call some tools and things like that. And it was great, ⁓ but one agent wasn't able to do that much, right? It was just, it was very limited. And the other thing that we found is it would be super lazy, right?

So it would oftentimes say, okay, I'm done. Like I looked at stuff and it's done, right? And the work wouldn't be great. So we're like, okay, so this is something we need to manage.

If we strung a bunch of them together, they would like chain the hallucinations and at the end you would have something very unusable, right? So we're like, okay, well, this is another problem that we have to solve. So ⁓ we built a product or part of the platform that we called Answers, which is just basically a set of tools that ask questions. Use different answer engines. we initially, we just basically looked online and script information, etc. ⁓ we're benchmarking Open AI, Excel, perplexity, all of the different web search, know, ⁓ search capabilities that are there.

They give you direct answers, right? But they still hallucinate, right? ⁓ then a judge. So the concept of LLM as a judge where, ⁓ where a quality control looks at the answer of the previous ⁓ ⁓ basically evaluates if it's correct. And if the facts are truly there, so like we need to have this on every single question the agent asks.

That slows it down quite a bit and then makes it lot more expensive. But it's the only thing that you can do if you really want to know that things work out. So we have to build that. The other thing that we had to do is we had to build autonomous quality control for these agents that basically can restart these agents and look at, ⁓ well, you haven't done enough of the work that you're saying you've done, right?

Like you try different things. Right. So we built this agent as a judge that looks at the other agents. Right.

And we said, okay, these are the two key things that we have to just build. And the third thing was for us was composability, which is how can we take individual teams of agents and, and combine or rather hand off their output into another set of agents that don't have to know about each other at all. Right. So we have kind of Lego blocks and we can assemble different applications, right.

Because you do TAM assessment for. venture building, you do it for a startup analysis, you do it for patent, for invention analysis, if you're a tech transfer office, for example, right? So the same thing can be multiple different places, but it cannot be connected, right? It's just like a workflow. So you have to build these three things.

We launched it like a year ago at the beginning of 2025, and it took us until May to like dial it in for it to actually work. And then we were surprised again, because we found that you know, everybody's like, N8N exists and stuff like that. Like you say, Hey, models exist. ⁓ found that our customers who are mostly, I would say VP director level plus analysts underneath them in large organizations, they will not build their own agents. It's just, they will not prompt anything.

They will not do N8N anything. ⁓ It's just going to happen. ⁓ And, ⁓ they're able to chat, right? So we had to kind of figure out how can we like create a chat that builds these agents for them. ⁓ But even better. can we go and canvas the most used workflows for consulting in these different verticals and build a catalog. So by the summer, we launched a catalog of about 150 different ⁓ apps, which are these workflows, teams of agents that do consulting grade work with traceable sources. ⁓ It's typically tens of agents, asks thousands of questions, everything is fact checked, right? So that's there.

We launched as a marketplace, but we found that the capability or the tooling by itself wasn't enough and what people really wanted to do is they wanted to automate it like a specific process that they had. So we covered about 30 % of our usage came from our catalog, but everything else was custom. So that was like another sort of unlock that we looked at. And then we found, okay, it's people that care about a process, right?

They didn't care about 150 things. They care about this one thing that if it just happened a lot more often, right, without consultants, right, their business would be better in specific ways, right? So for example, ⁓ one US bank uses us for ⁓ tracking interest rates for all of their competitors, right? So that they, and loan parameters, right? as they change.

So, and they have it set on autopilot, so it runs every week, right? So they can compare it. So, and that's like a use case that's worth a lot of money for them, because previously they would do it every once in a while or they would be too late and then we surprised us, right? So we found, okay, so it's like stuff like this, but then we found more, there's like also an explanation layer.

The fact that it happens, you still want to talk to a human, right? So that's another unlock. And then the last unlock we found is like, majority of our users just want to know hey What is everybody else doing and what should I do given what everybody else is doing? Like what is the impact on me?

Right recommend it and then will be found at the beginning of these years like okay and do it ⁓ EU AI act is coming out. It's important for you because you do AI stuff Run regulatory impact analysis. Don't even ask me just run and tell me right and also I want you to do the 60 page report But I don't want to read it Right? So I want you to do the diligent work that I know that the platform can do, but then I want a summary.

And if I get to talk to someone about this, even better. Right? So this is the full circle kind of that, that we've kind of gone to. So now we're at the stage that the tooling is done.

The platform works. If a new model comes out amazing, makes it like simpler for us to deliver the work. But now basically the platform is moving a lot more from hey, we're like a consulting a software marketplace to know we're like a consulting company ⁓ that just delivers end to end the outcomes, right? So it's fully in line with what Sequoia is writing.

Well, I think there are a couple of things. One of the things is changing people's behavior is the hardest thing to do. So this is going to be the draggiest issue that you can have. I think what we try to, like one of the mistakes that we've made is we try to jump too quickly to that.

It's like, okay, you don't need a consultants now. It's just like, here's the platform, you can do it by yourself. But I don't know how to do it by myself. I always ask someone, right?

And they would send me slides and then I would show the slides to my boss, right? And the boss would say, the slides are crap. So I would go to a consultant and I like, the slides are crap. And then the work would get done, right?

So, and then I would be like, if I do it now, I'm at risk, right? For my career. So, and this hasn't changed, this hasn't changed, right? I think there's a lot more pressure on people now with like, hey, but AI exists.

Like you should somehow be able to do it with AI and nobody knows what AI is. And we have an agent for that, right? But like, what is an agent? So I think even through all that hype, the inner workings of companies haven't really changed.

So what we're trying to do is, one of the things is, think one to five or one to 10 of users on our platform are super users, and then everybody else in that team benefits from them. This is one of the things that we've learned. And those are the system thinkers, those are the people that... We'll spend the hours to automate it once and then just click a button and run it all the time, right?

And they will be the stars and those are the people that you want to have in your team. Now the question is what do you do with the rest of the, that's more like an organizational question. What do you do with the rest of the team? It's not like they're bad or anything.

Is it like, do you want them to own these systems or should they do something else? Ultimately, my belief is even in our company, even with me, like, I am less and less the person who does the work and our developers as well, but we're organizing a system that does the work for us. And we're setting the parameters and setting the guard rails and all of that stuff, right? So we have an army of bots now that do our ⁓ bug fixes and small feature requests.

And then a human checks it, right? Like, but there's a system, like somebody designed the system of how these bots get spun up and what they do and what context they have, etc that's a, that becomes a job that becomes. honestly the job of our CTO and our head of QA, which previously that was not their job at all, right? So the jobs are changing dramatically and startups is of course going to be quicker. So like that's kind of what's on the ground. like one to five, one to 10 as a super user.

Okay, so what do you do about it? The other people want the outcome as well, but it's too daunting. So that's why we're turning on this proactivity. It's like, I know who you are, Utsav I know what your company's mandate is for fifth row.

And I know... what you care about, right? So given the company mandate, given your profile and your usage and everybody else's usage, let me look through the fifth row knowledge on you and just like recommend what the next action might be. And then just go and take the next section and give you the benefit. And through a feedback loop of like, have you clicked on it, right?

Have you interacted with the content that the platform has sent you? Then it is desire like like button, right? Like it's like, okay, this is more of what you like. Okay, we're going to send you more of what you like, right?

But this is like, proactive kind of consulting that you don't have to deal with. That's one layer of how we can induce a lot more usage and a lot more value. But the other one is going a half step back and just saying, we're just gonna have people that are gonna be at the front that you can talk to. They're not gonna do the analytical work, but they're gonna do the change management work.

They're gonna explain things. they're not going to be blockers for you or upsells suddenly for like, you want this next analysis? ⁓ this is another two weeks and however many junior man days, right? It's going to be there. Like let's, let's start at the platform can do it. We're going to, we're going to do it.

And then it's much more senior check-ins, right? So we see that we went full platform and now we had to build attack, but people still want to run projects. And somebody like two quotes from, from the past week kind of stand out for me. Somebody said, we left millions of dollars of services revenue on the table for hundreds of thousands of dollars of SaaS revenue.

And it feels like a lot of the AI consulting discussion is very much this, right? And the Sequoia articles came at the right time into this discussion. Everybody else is saying as well, this is the third one, not the second one, everybody else is saying, agentic AI platforms like ours. never work out of the box in enterprises. You always need to set them up and always need to work with these people.

It's a misnomer that some, it just like works like that. Um, This is a Silicon Valley thing, like constantly discussing with people like this. It never just works and it's not the platform can't do it, but it needs to be plugged into, um, specific data sources and blah, blah, blah. So actually really useful.

The second, the second quote was, um, deliveries now faster than discovery. and I see it for us, I've seen it, this has really changed over the past three months, that we can build any feature we want, like literally any feature we want. I yesterday asked, we have a potential Japanese client, and like, hey, what would it take to translate everything on the platform to Japanese? And the answer was like a lot of cursor credits, right? And probably a day of somebody's time.

Right. To because we have only English, right? So if we did Japanese, then we can do any other language. So any other language takes one day.

Right. So that's like, that's the barrier right now versus previously, this would be like two, three weeks, maybe. Right. Certainly not one person.

So, um, we start, we, we had to like change our behavior of like, how does information get to the developers when writing the code is no longer a barrier. So this is organizational redesign. I'm going a little bit further from what you were asking because you were asking about our customers. But I feel like we are on the forefront of how the organizations change.

So I think we need to do it step by step and introduce it step by step. giving them a human face at this moment right now is probably the easiest fix for actually getting value and utilizing also the benefits of the developments over the past three years.

I think it's very difficult for 2035, but I can tell you a framework of how I think about it. So consulting for me has three layers. It's data analysts and the human layer on top of the toxic customers. Right now, the analysis is already automated.

It's what we do, right? So that layer goes away. The problem with that is that is what drives most of the revenue. Like the portion of the revenue on the proposal that you're sending to a customer.

Right. So there's a massive innovator's dilemma on how to change it. And we're in discussions with all, all of the large consultancies. And right now they're looking at efficiency, ⁓ sort of how to enable their, their internal people, etc. ⁓ There is going to be a massive reorganization at one point.

And what fifth row is trying to build is like, what does a consultancy, what is the biggest consultancy you can build without any consultants upfront or any analysts upfront. Right. ⁓ But deliver the value of what the consulting should bring. right? Something that you don't have internally that requires external cross-market understanding, needs to be always saying, we need to be blamable, right? We need to have a mandate where somewhere in the boardroom is like, why did we do Because Fifth row said it, right?

And we trust Fifth row right? So that has to be that. So the analyst layer is not a moat. the human touch layer a moat in terms of the customer relationships that you already have. The only biggest long-term moat, ⁓ and the human touch is gonna stay being a moat ⁓ as long as humans buy from humans, right?

What happens in the next 10 years, I'm not necessarily sure because you see things like OpenClaw, like rising right now, and there's MoltBook and things, Facebook literally just today or yesterday bought ⁓ right? So you see that these marketplaces, they will be bots buy from bots, right? So then the human layer is not as important, at least the subset of market, the market's gonna be small and then it's gonna grow rapidly, right? ⁓ then the third is data and primary, ⁓ primary data and always give an example of like ⁓ canvassing company in India, right? Where people go door to door, the data from that is super valuable, right?

I would buy that company right now, right? because that data doesn't exist yet, right? And you will need to ⁓ keep getting new data to get and better information about that market, right? But it's very physical right now how to get this information. So source of primary data, that's constantly new, journalism as well could be one that you look at it, is going to be a real moat.

So I would look at... My prediction for consulting is that it's it's data pools and established workflows ⁓ that do the analysis as if, you know, it was consulting before, but delivers it as much more ⁓ piecemeal value or like I can buy, like basically what we launched, the marketplace that we have on the website right now, where you could pay, I don't know, 50 bucks to get like a really good investment memo. I think it's gonna look a lot more like that, but not now, at some point in the future.

I would look at the, would look at two places. I would look at the data. think there are two big opportunities right now. Like, can you find a source of data that, like, for example, if you're doing secondary market analysis, ⁓ but that's cheap now, right? So that's not a, if you were like, how big is the market of X?

Unless you have proprietary data, everybody has the same information. And previously the human labor to get to the number was like determining the value of the thing, right? And then the size of the market of people who would buy it would determine the price. That's no longer like you can like on fifth row, can do it now.

And it's like a fraction of a, you know, not a dollar, but like a couple of tens of couple of dollars, probably to get like a really good analysis on something. So that's not it. What are, what are new sources of data that you could own and base your consulting around that? And the other, so the other area, and then I'm, discarding the layer completely because there's so many things you could build by yourself even with Claude code or cursor or whatever. ⁓ You have to look at the fact checking because that really when this all breaks down very, very quickly and there's so many examples of that. ⁓ So find a tech partner or something for that. ⁓ And then it's the opportunity right now is I think retraining ⁓ process management, right?

Of like you have all this AI here like what should the organization look like, right? And go sell that. So some combination of these things is probably where the opportunity for net new things ⁓ lies right now.

Thank you, Utsav