Dec 18, 2024 · 1h 6m · news
Daniel Dines, UiPath CEO & Founder: Why Agents Do Not Mean RPA is F*** | E1240 · 20VC with Harry Stebbings
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this deep-dive interview, UiPath CEO and Founder Daniel Dines discusses the strategic integration of rule-based RPA with probabilistic AI agents, while sharing candid reflections on corporate leadership, public market realities, and his personal journey toward finding mental peace and purpose.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Harry holds 20.5% of the talking time here. How this is scored →
Harry as informed peer 4.0 Guest teaching 5.3 Guest disagreement 3.0 Harry pushing back 3.8
speaking balance: gold is Harry, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 52:10 Dismissing Salesforce technology complexity
Daniel directly dismisses Salesforce's engineering challenges when Harry compares their hiring freezes, dryly stating that UiPath's underlying technology is vastly harder to build.
Hardest push from Harry ▶ 21:40 Challenging Marc Benioff's agent strategyHarry refuses to accept Daniel's premise about enterprise platforms, interjecting to assert that Benioff is building a full suite of cross-functional agents that directly compete with UiPath.
Biggest teaching moment ▶ 25:00 Compounding mathematical error rates in LLMsDaniel educates Harry on the inherent flaw of non-rule-based LLMs for multi-step tasks by walking through compounding probabilities (0.99^100) across sequential enterprise workflows.
Harry holds his own ▶ 45:03 Citing returning tech founders to challenge revenue thresholdsHarry uses deep domain knowledge of tech history to push back on Daniel's revenue stage thesis, citing Larry Page, Sergey Brin, and Jeff Bezos returning to active management regardless of company size.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Harry as informed peer | Guest teaching | Guest disagreement | Harry pushing back | Why |
|---|---|---|---|---|---|---|
| Welcome and the AI Product Cycle | 2 | 4 | 1 | 1 | Harry introduces a thesis on product experience over pure model innovation in the current AI cycle. Daniel agrees, sharing an untold backstory about how UiPath used OpenCV image recognition to beat Blue Prism through superior user experience. | |
| The Landscape of LLM Models: Specialized vs. Monolithic | 4 | 5 | 2 | 3 | Harry probes why UiPath relies on Alibaba's Qwen model over western alternatives and asks if AI models will mirror monolithic cloud platforms. Daniel politey rejects the cloud analogy, using human brain dedicated motor functions to explain why open-source specialized models will dominate. | |
| UiPath's Shift to Agentic AI and Rebuilding Workflow Engines | 3 | 5 | 1 | 2 | Harry asks what legacy technology UiPath had to abandon to become AI-first. Daniel candidly describes abandoning their long-standing Windows workflow engine in favor of modern agentic orchestration. | |
| Coexistence of RPA and Agentic AI in the Enterprise | 5 | 6 | 3 | 6 | Harry aggressively asks why enterprises wouldn't buy non-rule-based systems from new vendors instead of RPA incumbents. Daniel explains enterprise workflows naturally contain both deterministic and non-deterministic steps, requiring a unified orchestration layer. | |
| Enterprise Risk Appetite and LLMs as "Idiot Savants" | 5 | 6 | 4 | 5 | When Daniel labels LLMs as idiot savants and notes enterprises fear models that are too smart, Harry pushes back by noting human employees also make mistakes. Daniel reframes customer risk tolerance, noting enterprises prefer predictable failure over unpredictable smart behavior. | |
| The Timeline for Autonomous Agents: A Self-Driving Analogy | 6 | 6 | 5 | 7 | Harry directly challenges Daniel by quoting Marc Benioff's ambition to build all enterprise agents. Daniel forcefully rejects Benioff's premise, citing a major healthcare CIO who refused to move Epic data into Salesforce just to deploy agents. | |
| Near-Term Agentic Use Cases and Debunking the Rule-Based Misnomer | 3 | 6 | 2 | 2 | Harry asks for specific agentic use cases and the main misnomers around them. Daniel breaks down mathematical probability in LLM steps, demonstrating how error rates compound across sequential rule-based tasks. | |
| The Early Stage of Enterprise Agentic AI and Customer Education | 5 | 5 | 3 | 5 | Harry bluntly asks why Wall Street undervalues UiPath if it holds the primary orchestration platform. Daniel admits it is still early innings, explaining enterprise GenAI stalled previously due to unpredictability. | |
| The Future of Work: Transitioning Humans from Doers to Validators | 4 | 5 | 4 | 6 | When Daniel argues humans will shift from doers to validators, Harry forcefully rejects this outcome as soul-crushing validation monkey work. Daniel defends the transition by showing human input will only be required for high-level exception handling. | |
| Corporate Inertia and the Realities of Productivity Gains | 5 | 6 | 5 | 6 | Harry contends AI transition speed will be drastically faster than multi-decade shifts like agricultural mechanization. Daniel explicitly tells Harry he underestimates corporate inertia, pointing out that UiPath's legacy RPA is still under 20% penetrated. | |
| Defining AGI for Enterprises and the Stochastic Reasoning Debate | 6 | 7 | 5 | 5 | Harry brings up Sam Altman's claim that AGI arrives in 2025. Daniel rejects Altman's framing, defining enterprise AGI as predictable 120 IQ performance and arguing current LLMs are purely stochastic engines incapable of human logical reasoning. | |
| AI Capex, Training Plateaus, and the Nvidia Moat | 6 | 6 | 4 | 4 | Harry cites Masayoshi Son's $9 trillion AI capex prediction and asks if Nvidia can retain its monopoly. Daniel questions GPU scaling efficiency, citing Sundar Pichai on training plateaus while analyzing Nvidia's hyperscaler revenue concentration risks. | |
| The Future of SaaS Business Models: Seats vs. Consumption | 4 | 5 | 2 | 3 | Harry explores whether consumption pricing will replace seat-based SaaS models. Daniel offers a pragmatic view that enterprises will adopt hybrid models rather than purely eliminating seats. | |
| Reflections on Founder Mode and Leadership Transitions | 6 | 5 | 4 | 6 | Harry cites Paul Graham's Founder Mode essay and points to returning founders like Larry, Sergey, and Bezos to argue revenue size is irrelevant to founder leadership. Daniel clarifies he never actually left the company while serving as co-CEO. | |
| Driving Team Culture: Transparency vs. Corporate Fluff | 4 | 5 | 3 | 3 | Harry and Daniel bond over hating corporate buzzwords. Daniel shares his unconventional management approach of eliminating formal direct-report one-to-ones in favor of ad-hoc transparent communication. | |
| Learning from Mistakes: Strategic Hiring and the Agentic Timeline | 5 | 6 | 5 | 4 | Harry asks about past leadership mistakes and cites Benioff freezing developer hiring. Daniel dryly dismisses the Salesforce comparison, stating UiPath's underlying technology is far harder to build. | |
| Living in the Best Era: Peace of Mind, Freedom, and the True Purpose of Success | 1 | 5 | 1 | 1 | Harry asks personal questions on ambition and envy. Daniel reflects on wasting his earlier decades desiring material outcomes, explaining the profound freedom of wanting nothing. | |
| Quick Fire: Leadership Realities, Loneliness, and Selective Discipline | 1 | 4 | 2 | 1 | In a quickfire round, Harry asks about CEO loneliness and discipline. Daniel explains how intentional lack of discipline fosters his personal creativity. | |
| Quick Fire: Personal Regrets, Stress Relief, and UiPath's Second Act | 2 | 4 | 2 | 2 | Harry asks quickfire questions on stress, stock picks, and regrets. Daniel shares that he writes poetry to process intense stress and chooses Anthropic over OpenAI due to valuation upside. |
Statements from this episode (44)
UiPath CEO Daniel Dines says he lives mostly inside his own head
“I am a lonely wolf. I find life pretty lonely, man. That's not only about this job, but I live mostly in my head. Thinking, analyzing, reflecting. This is how I spend my life.”
Dines: UiPath spent two years fine-tuning LLMs with moderate success
“So over the last two years, we we've spent a lot of time really trying to fine tune LLMs, build around them, and certain degree of success.”
UiPath CEO Daniel Dines praises Cursor AI as a beautiful product
“But I've got really inspired by stories like Cursor AI. And my development team loves that product. It's a beautiful product built on the top of multiple LLMs, but just, it just works.”
Dines: AI models have reached maturity and won't innovate materially soon
“I personally don't believe that the models can innovate in a material way in a reasonable amount of time. I think they reach maturity in a way.”
UiPath uses Alibaba's open-source Qwen model for semi-structured documents
“We are using Gwen, which is a fantastic model built by Alibaba, which is totally open source. We are using it into understanding, like a lot of our semi-structured documents”
Dedicated AI models will be built on open-source, not closed frontier models
“So obviously there will be a world A few frontier models and a lot of dedicated models, and dedicated models will be built on the top of, I think is more likely to be built on the top of open source models that, you know, close source frontier models.”
Dines: UiPath is rebuilding its software from scratch for agentic AI
“So we are building our agentic AI approach from the ground up, and we gave up on some of our, you know, RPA stuff in order to come on the new technology, New frameworks, building from scratch, because we want to build an AI first experience.”
Dines: LLMs Are Ineffective at Executing Repetitive Rule-Based Steps
“LLMs are actually not good at following repetitive steps. You are not gonna have LLMs multiply to numbers. No, you are gonna follow an algorithm, and you are gonna use script language, or you will program it, right?”
Dines: RPA Bots Are Low-Skilled Workers, AI Agents High-Skilled Workers
“Think about as a metaphor, like, robots are more like low-skilled employees, while agents are high-skilled employees. But you manage them within the same platform.”
Dines: Enterprise clients prefer automation workflows to fail rather than be 'too smart'
“Look, one story that I keep hearing from our customers is they prefer our workflows to fail than to be too smart. Because their risk appetite for this type of workloads, it's low.”
Dines: Enterprise AI agents will recommend, not act directly without human validation
“Agents will make recommendations. Agents are not gonna take actions directly. There will be a progression from agents making recommendations, going to a human user for validating, and then calling an action.”
Dines: LLM-based AI agents act like 'idiot savants' with unpredictable reliability
“Not because we are scared, because they are like idiot savants. Sometimes they can be extremely smart, sometimes they can be extremely dumb. And you have no idea right now how to distinguish between these two scenarios.”
Dines: Fully autonomous AI agents will take as long as self-driving cars
“As long as it's gonna take for you know, the nice self-driving cars that we have today to be fully autonomous and to drive, you know, asides with people on the streets and they will just work.”
Dines: Salesforce and SAP Will Limit AI Agents to Own Platforms
“What's the interest of Salesforce to provide amazing connections to SAP and vice versa? No, it's not. So they will focus on building agents that work specifically for workloads that stay within their platforms.”
Dines: Enterprise Clients Won't Move Epic Data to Salesforce For Agents
“I was talking to the CIO, which is close to us, and he said, I will never put data from Epic into Salesforce in order to create an agent. Never. There is no chance to do this. So, It's as simple as this. They will prefer to use us to have connectors, feed agen…”
Dines: Generative AI has not been successful in enterprise so far
“AI, Gen AI was not so successful in enterprise so far.”
Dines: Human validation of AI will focus only on complex edge cases
“These validations will be actually mostly on the difficult cases in, in the longer term.”
Dines: Building a good AI prompt is harder than writing code
“Building an agent requires creating a prompt, and building a good prompt, it's actually more difficult than building building a script.”
Dines: RPA market penetration remains under 10% to 20%
“Yeah, it's probably less than 10, 20% penetrated.”
Dines: Wide-scale agentic AI deployment will take 5 to 10 years
“I think it's gonna take next five, 10 years to, with the current state-of-the-art LLMs, it's gonna take next five to 10 years to get to very wide scale deployment of agentic plus automation.”
Daniel Dines defines enterprise AGI as predictable 120 IQ multitask capability
“AGI for enterprises would be when I have an LLM that has the capabilities of a guy with an average IQ like a 120 points. But predictably. Not a 180 in some fancy math jobs and 60 into, you know, other type of job. Predictably. This is AGI.”
Daniel Dines says current LLMs cannot reason and require architectural leaps
“Every industry will be subject for a big change, but I also believe that we need a new giant leap in order to get there. I think, I don't believe that actual LLMs reason in the sense that I expect a guy with a 120 IQ reason. It's still, it's a stochastic engin…”
Dines: Scaling GPU compute with current algorithms will not yield godlike AI
“Look, I think it's, it would be an easy investment if there is a predictable outcome, but I don't, do you really think that just adding GPUs and with the existing algorithm to train, they will suddenly become, become godlike, intelligent? I don't understand. A…”
Dines: Nvidia will struggle to maintain current chip dominance against hyperscalers
“It's hard to see it at existing levels, and think about, I think half of NVIDIA revenue today is from these five hyperscalers, so, and if they are all in the business of building, it's kind of, I think it's hard, but obviously NVIDIA has such an important, you…”
Dines: Future SaaS pricing will combine seat-based and consumption-based models
“I think the pricing models in the future will be both seat based and some kind of, they will have some kind of consumption based mechanism. Combine.”
Dines: UiPath's main challenges are becoming AI-first and re-energizing staff
“My biggest challenge is is transforming the company to be an AI-first company, and re-energizing our people that you know, we had...”
Dines: Public markets reward steady 30% YoY growth over hypergrowth deceleration
“I think it's better to plan and execute a growth of 30% year over year rather than 80, 60, 3020, 10, you know, because in a way doing such a huge aggressive growth makes you maybe sometimes steal from the future. Not knowingly, but you discover in time. I thin…”
Dines: $1B revenue companies still require founder mode
“I think for One billion in revenue company, it's actually, it's not it's not that established in a way. You need to get to a critical mass.”
Daniel Dines managed UiPath product directly while Rob Enslin was CEO
“No, I was, I, actually, I never left the company. Ok, so, I ran product and engineering directly, while Rob was co-CEO and CEO.”
Dines: Unified CEO leadership is vital during major technological shifts
“It's in a time when there is such a huge Change in technology. I think the CEO baton is so important, because you connect instantly product and go to market, and marketing, and this is, it's a powerful, you know, flying wheel that has to work.”
Dines: Shutting down employee initiatives hurts morale more than stock drops
“If you are shut down, then it affects the moral more than the stock price, I think.”
UiPath CEO Daniel Dines does not conduct scheduled one-on-ones with direct reports
“People put too much importance on being disciplined, and having, you know, regular one-to-ones. I don't believe in one-to-ones with my directs.”
Dines: Direct feedback is appreciation, indirect communication is a bad sign
“Actually what I'm telling people that where we, that I work with in a very direct style, that's a sign of appreciation. The moment you see that I I'm, you know, working around, and I'm finding my words. That's not really a good sign.”
Daniel Dines says never compromise personal chemistry for executive experience
“I hired for experience. And I should have never done compromise on chemistry for experience.”
Daniel Dines admits UiPath should have moved into agentic AI earlier
“We should have landed into a genetic six months. Six months earlier, yes, that, that's for sure.”
UiPath has frozen software engineering hiring to repurpose staff for AI
“No, we are not hiring as well. We have repurposed a lot of engineers from products that we de-emphasize into Agenda.”
Daniel Dines claims UiPath's technology is much harder to build than Salesforce's
“Look, with all respect to Salesforce, our technology is much harder to build.”
Daniel Dines says AI productivity gains for software engineers won't be gigantic
“You will get some productivity improvement that I don't think will be Gigantic.”
Dines: Material desires waste cognitive bandwidth better spent on AI and learning
“Thinking about a bigger kitchen will take cycles from reading, understanding world, understanding people, understanding AI.”
Dines: Stock price matters for hiring and building, not personal wealth
“So stock price is important. Not for my, you know, wealth, for how much money I can spend in this life. But it's important for what we can build, for the talent that we can attract.”
Daniel Dines believes a lack of discipline is essential for stimulating creativity
“I believe in some certain lack of discipline. It's very important in order to stimulate creativity.”
Daniel Dines says running a company is hard because only problems escalate
“The hardest part is to manage the unhappiness of people. I guess, look, good news don't go to me. Really. They are in day by day, in their day by day jobs, but everything that is not working well goes up to me, so I have to deal. But that's not only being CEO …”
Dines: Anthropic at $40B valuation has bigger upside than OpenAI at $160B
“Probably anthropic... I think the upside is bigger.”
Dines: Achieving a tech second act requires luck, not just capital and talent
“Very few companies actually get to have a second act, and what I realize, it's actually kind of very hard. You know, as a startup, you become successful understanding a space very, very well, right? Getting into another space for a second act, it's kind of ver…”