Daniel Dines, UiPath CEO & Founder: Why Agents Do Not Mean RPA is F*** | E1240

Intro (0s)

• Daniel Dines has a personal story that he has never shared with anyone, which involves wasting a significant part of his life thinking in a certain way, specifically his late 20s, a big part of his 30s and 40s (9s).
• He considers himself a "lonely wolf" and finds life to be pretty lonely, not just related to his job, but also due to spending most of his time thinking, analyzing, and reflecting (17s).
• Daniel Dines is excited to be on the show and thanks the host for having him, mentioning that he was invited back due to the success of their first show (42s).
• The first show with Daniel Dines did well, receiving many emails from young entrepreneurs who were inspired by his journey, prompting the host to invite him back to capitalize on that success (52s).

Why Does Product Matter More Than Innovation in AI? (1m3s)

• At this stage of the AI cycle, product matters more than innovation, as ease of use, simplicity, and great UI drive adoption rather than technological advancement or sophistication of models (1m3s).
• The story of UiPath began with using a library called OpenCV, which provided a feature to find a smaller image within a bigger image, and repurposing it for automation (2m9s).
• This image-based recognition allowed UiPath to automate processes without having access to the applications, other than by remote desktop via Citrix, and became their first niche (4m41s).
• The initial product created a magical experience by letting users record flows on the screen, generating simple statements, and storing everything, making it easy for users to record entire flows (3m5s).
• In 2013, a demo of this product was shown to Blue Prism experts, who were impressed by its simplicity and reliability, and it took only a few minutes to complete a task that would have taken two days in Blue Prism (3m44s).
• UiPath found its first niche in automating processes without access to applications, and from there, they expanded into what they are today (4m48s).
• The company is using various models, including Gwen, a fantastic open-source model built by Alibaba, for understanding semi-structured documents, due to its product quality (5m36s).
• The choice of model depends on product quality, and UiPath uses the model that best fits their needs, rather than relying solely on technological advancement or sophistication (5m58s).
• The current model used for a particular job may change, and the experience around the product will be crucial in making it simple to use and retrain the model on the fly (6m2s).
• The entire product experience, including helping people tag documents and retraining the model, makes it extremely simple to use and allows for the exchange of models if a better one is found (6m14s).
• The choice of model depends on a cost versus speed and accuracy equation (6m41s).
• It is unlikely that there will be one or two monolithic models, similar to cloud development with AWS or Google Cloud, and instead, there will be multiple models specialized in different areas (6m47s).
• The human brain development is an example of having multiple models, with both general cognitive models and specialized models that perform specific tasks better (7m8s).
• Even simple tasks, like drinking from a cup, require dedicated models that are trained from a young age (7m25s).
• There will be a world with a few Frontier models and many dedicated models, with dedicated models likely being built on top of Open Source models rather than closed-source Frontier models (7m50s).

What’s Next for UiPath with Product as the Priority? (8m8s)

• The recognition that product is more important than technological advancement has a significant impact on the next chapter for UiPath. (8m19s)
• This realization led to a big shift in the company's approach to building software, with a focus on creating products from the ground up. (8m29s)
• UiPath's product, Cursa, was built from scratch with an AI-first approach, serving as a model for how the company should build software in an AI-driven world. (8m45s)
• The company is now building its agentic AI approach from the ground up, giving up on some of its existing RPA technology to adopt new frameworks and build from scratch. (8m56s)
• The goal is to create an AI-first experience, which is a key priority for UiPath's future development. (9m12s)

Why Is RPA Compatible with Orchestration & Agents? (9m15s)

• The company had a workflow engine called Warlow, but it has been replaced with a more modern technology that works better for agentic orchestration, which is specifically designed to facilitate connections between agents, human users, and other robots or models (9m28s).
• RPA is compatible with agentic orchestration and agents because it automates tasks that span multiple business systems, are of medium to complex complexity, and involve multiple steps, usually with rule-based inputs and steps (10m52s).
• RPA captures company knowledge within rules, making it reliable until the underlying system changes, and is particularly useful for tasks that involve structured inputs and rule-based steps (11m29s).
• Agentic Large Language Models (LLMs) are not good at following repetitive steps and are better suited for dealing with unstructured parts of business processes, where enterprise knowledge is difficult to express in rules (12m0s).
• LLMs work well when there is tribal knowledge involved, and human users are supposed to have certain knowledge, allowing agents to mimic user actions and reduce human input, but not eliminate tasks entirely (12m44s).
• RPA and agentic orchestration are not mutually exclusive, and RPA is still valuable for automating tasks that involve multiple business systems and complex rules, while agentic LLMs are better suited for tasks that require human-like judgment and decision-making (10m52s).

Will Enterprises Split Vendors for Rule-Based vs. Non-Rule-Based? (13m16s)

• The distinction between rule-based and non-rule-based automation is the easiest way to differentiate between the two, but this does not necessarily mean that enterprises will buy from two different vendors for each type (13m17s).
• Rule-based and non-deterministic parts often sit within the context of a long business process, such as order to cash or procure to pay, and it makes sense to have the same technology and framework for both (13m52s).
• Having the same technology and framework for both rule-based and non-rule-based automation is important for orchestration, and it is powerful to have a platform that can manage both low-skilled and high-skilled tasks (14m12s).
• The platform should be able to automate thousands of tasks, manage them, deliver, deploy, monitor, and provide analytics, as well as control access and ensure security (15m6s).
• The key differentiator of the UiPath platform is its ability to orchestrate robots and manage them in a rule-based environment, and this is being expanded to non-rule-based automation (15m34s).
• Both RPA and agentic automation imitate people and processes, and there are similarities in deploying and managing them, but non-rule-based automation requires more exception handling and retries (16m0s).
• The experience with robots is being used to build agents that can work reliably thousands of times, and this is important for enterprises that want to deliver autonomous workflows in production (16m37s).
• Enterprises prefer workflows to fail rather than be too smart, and this risk appetite will be the same for delivering agents, which will make recommendations rather than taking actions directly (17m10s).
• Agents will make recommendations, and then a human user will validate and call an action, because enterprises are scared of agents being too smart and making wrong decisions on their behalf (17m31s).
• This is similar to human workforces, which can make mistakes and errors of judgment, and this is why enterprises create rule-based workflows and precision-type enterprise workflows (18m12s).

How Long Until Users Fully Trust AI Agents? (18m38s)

• The development of fully trusting agents in automation is expected to take a long time, similar to the development of fully autonomous self-driving cars, but semi-autonomous agents will be used in the meantime to reduce human input (19m8s).
• In the future, semi-autonomous agents will do most of the job, and humans will monitor their inboxes, validate tasks, and provide feedback to the agents and the Enterprise workflow (19m28s).
• The Enterprise workflow will be rule-based, with an orchestration layer on top, and will connect different people to do their jobs (19m47s).
• UiPath aims to be a provider of agent-based, non-rule-based actions, but will also integrate with agents built on other platforms (20m16s).
• UiPath will act as an orchestration technology that is agnostic, providing equal access to different platforms, and will be the "Switzerland of the platform" (21m3s).
• Other platforms, such as Salesforce, will focus on building agents that work specifically for workloads within their platforms, rather than providing connections to other platforms (21m24s).
• Customers will prefer to use UiPath to have connectors that feed agents only with the data they need to make a decision, rather than migrating data between platforms (22m17s).
• The future of automation will involve specialized agents within data repositories and an orchestration layer on top, as well as agents on UiPath's platform (22m46s).
• Most RPA tasks today require connecting to multiple platforms, working across platforms, and gathering data from two or more systems simultaneously to make a decision (22m53s).
• In the next 1 to 3 years, game-changing results are expected in tasks that require agents, such as processing denials or prior authorization in healthcare systems (23m31s).
• To identify tasks suitable for agents, companies should analyze their current processes, such as procure-to-pay, and identify rule-based and non-deterministic parts (23m45s).
• Agents will be created specifically for tasks like processing denials or prior authorization and will be connected with robots that perform rule-based tasks in an orchestration layer (24m28s).
• The biggest misnomer about the non-rule-based agent layer is that agents will be good at doing rule-based tasks, which is incorrect as they are not fundamentally good at tasks like multiplication due to their creative nature and error rate (24m47s).
• Large Language Models (LLMs) have a high error rate for rule-based tasks, with a success rate of 0.99 resulting in a significant error rate after multiple steps (25m10s).
• LLMs will often provide different results for the same question, working similarly to the human mind, which makes them unsuitable for tasks requiring consistent and accurate results (25m36s).

Why Doesn’t Wall Street Value UiPath’s Position More? (25m43s)

• Having a strong position in Switzerland and a well-established distribution network is crucial for success, and distribution is more important than product innovation in this context (25m49s).
• Wall Street does not fully appreciate UiPath's position because it's still early in the development of the company's agentic technology, and there's a need to deliver on the roadmap, including the agentic orchestration workflows and the agent Builder (26m7s).
• The agent Builder has been launched in private preview, marking the early stages of a significant movement to adopt agents and AI in the enterprise space (26m26s).
• Geni AI has not been successful in the enterprise so far due to its unpredictability, but this will change when it's integrated into agentic workflows with rules and human validation (26m54s).
• The future of AI in the enterprise will involve using agents as part of larger workflows, triggered by external factors, and will require human validation and oversight (27m31s).
• UiPath's conversations with customers are focused more on education and learning than selling, with an emphasis on understanding how customers see the world and educating them on UiPath's vision (28m19s).
• One key lesson learned from customers is the importance of thinking end-to-end and considering entire processes, rather than just individual tasks (28m44s).
• For agenting to be successful, it's essential to have a comprehensive understanding of the end-to-end process and to integrate agents into larger Enterprise workflows (29m9s).

How Will Agents Reshape Roles & Functions in Enterprises? (29m27s)

• The structure of companies and internal roles will change with the adoption of agents, as people will become more productive and roles will shift from doing tasks to overseeing technology and validating its output (29m28s).
• The validation process will primarily focus on difficult cases, and as agents become more autonomous, people will become comfortable with their output, allowing for more efficient decision-making (30m26s).
• The progression to full autonomy will involve a gradual increase in trust in agents' recommendations, with users initially validating most decisions before eventually only intervening in exceptional cases (30m37s).
• The use of agents will enable users to set limits and automate routine decisions, such as booking travel arrangements within a certain budget, reducing the need for human input (31m3s).
• The democratization of agents through agent builders and other elements may allow Enterprise companies to access new markets, including small and medium-sized businesses (SMBs), but this will depend on the development of necessary skills (31m22s).
• Building agents requires specialized skills, including creating effective prompts, which can be more challenging than building scripts due to the unpredictability of language and the impact of slight changes in prompts on agent performance (32m7s).
• To address this challenge, efforts are being made to help automation developers build better prompts, suggest prompts, and create evaluation sets to test and refine agent performance (32m42s).

Will AI-Driven Verification Reduce Company Size? (33m7s)

• The future of companies will involve a shift towards verification and approvals-based functions for humans, potentially leading to a reduction in job numbers, but also creating new jobs as society evolves (33m10s).
• The example of the agriculture industry is cited, where 100 years ago 50% of the UK or US population worked in agriculture, but nowadays it's only 2%, with most workers supervising machines (33m28s).
• The speaker believes that the only way to keep up with economic growth is to increase productivity, especially with population aging and reducing in many First World countries (34m6s).
• RPA is seen as essential to increase productivity, and the focus should be on the benefits it brings rather than the "doomed day scenario" (34m27s).
• The speed of progression through the technology cycle will likely be much faster than any other technology adoption cycle, with the farming industry's adoption of machinery taking multiple decades (34m48s).
• The speaker estimates that RPA technology is less than 10-20% penetrated, and it's not an easy technology to deploy, requiring a whole program behind it (35m27s).
• The deployment of Gen K, a type of automation, will take the next 5-10 years with the current state-of-the-art LLMs (Large Language Models) (36m4s).
• The possibility of AGI (Artificial General Intelligence) is discussed, with some people thinking it will happen soon, but the speaker's definition of AGI is different, requiring an LLM with the capabilities of a person with an average IQ of 120 (36m49s).
• When AGI is achieved, the jobs landscape will change completely, and it's not just about RPA or automation, but a big change for every industry (37m23s).
• A new giant leap is needed to achieve AGI, and the speaker doesn't believe that current LLMs reason in the same way as humans (37m39s).
• The speaker mentions Sam Altman, who predicts that AGI will happen in 2025, but the speaker's definition of AGI is different (36m36s).
• The discussion revolves around the concept of stochastic engines and the limitations of Large Language Models (LLMs) in certain tasks, highlighting the existence of a different type of intelligence that is not equipped to work in the context of business operations where reliability is needed (38m4s).
• The transition of labor forces and the deployment of AI agents are expected to create significant economic gains, with an estimated $9 trillion in GDP gains per year, but this would require a substantial investment of $n trillion in capex (39m5s).
• The effectiveness of simply adding GPUs and training existing algorithms to achieve Godlike intelligence is questioned, with concerns about the training plateau and the need for alternative approaches (39m39s).
• The dominance of Nvidia in the chip market is discussed, with the possibility of Amazon, Meta, Google, and other hyperscalers building their own chip capabilities and potentially disrupting Nvidia's monopoly (40m12s).
• The future of business models is explored, with a potential shift from seat-based pricing to units of work or cognition, but it is acknowledged that this transition will not be binary and will likely involve a combination of pricing models (41m7s).
• The concept of inertia and the speed of prices is mentioned, with the suggestion that it is underestimated and will play a significant role in the evolution of business models (41m22s).
• The idea of paying for transactions rather than seats is discussed, with the possibility of consumption-based pricing mechanisms being combined with traditional seat-based models (41m35s).
• The importance of Nvidia's position in both hardware and software is highlighted, with the company's ability to maintain its dominance in the market being uncertain (40m45s).

UiPath’s Biggest Challenge in the Next 2 Years (42m1s)

• The biggest challenge for UiPath in the next 12 to 24 months is transforming the company to be an AI-first company and re-energizing its people (42m2s).
• The company had a rocky ride into the public markets, particularly in 2021, which may have been partially unavoidable for a company going through an IPO (42m33s).
• If given the chance to do a second IPO, the approach to finance and going to market would be done differently, with a focus on planning for consistent growth rather than aggressive growth (42m49s).
• The goal would be to achieve a consistent 30% year-over-year growth, rather than experiencing large fluctuations in growth rates (43m17s).
• This approach is considered better because it allows for more organic growth, which is often rewarded in the public market, as noted by Paul Graham (43m43s).
• Aggressive growth can sometimes make a company miss out on its future, as it may not be able to keep up with its own rapid expansion (43m32s).

Thoughts on Founder Mode (43m47s)

• Founder mode is beneficial for companies, especially during certain stages, and it's essential for companies that are still growing and not yet established. (43m49s)
• Hiring a CEO and then returning as CEO can be a good approach, but it's crucial to understand the company's stage and whether it needs a founder's leadership or an experienced CEO. (43m56s)
• The decision to hire a CEO, Rob, was based on the assumption that the company was at a more established stage, but in hindsight, it was still in a stage where founder mode was essential. (44m26s)
• The revenue of a company, such as $1 billion, is not necessarily an indicator of its establishment, and what's more relevant is the technology cycle. (44m50s)
• Founders like Larry and Sergey from Google, and Bezos from Amazon, have returned to their companies, not necessarily as CEOs, but to contribute to the company's growth and direction. (45m8s)
• The experience of running product and engineering directly while having a CEO, Rob, was challenging, and it required a delicate balance and communication between the two roles. (45m30s)
• In times of significant technological change, the CEO's role is crucial in connecting product, go-to-market, marketing, and other aspects of the business to create a powerful and efficient system. (46m7s)

Daniel’s Way To Motivate His Team (46m26s)

• Corporate words like "empowerment" and "alignment" are often used but have little meaning, and instead, a more transparent approach is needed to reenergize a team by acknowledging the current situation and the work ahead (46m27s).
• As a company grows, people can start to feel smaller, and it's essential to create an environment where people have joy and feel empowered to make decisions (47m24s).
• To make a big change, individuals should be encouraged to drive that change and have a voice to fight for it, as being shut down can affect morale more than the stock price (47m44s).
• To reduce bureaucracy and empower teams, efforts should be made to give regions more control and get closer to the customer, taking cues from the customer to inform product development (48m7s).
• A transparent and supportive approach can help motivate teams, allowing them to take big risks and make decisions, which is essential for building a successful company (48m2s).

Biggest Management Rules Daniel Thinks Are BS (48m26s)

• Some management rules that are often emphasized, such as being disciplined and having regular one-to-one meetings, are not considered important. (48m27s)
• One-to-one meetings are not believed to be effective for honest and candid discussions, and instead, direct reports should be able to call their superiors as needed. (48m42s)
• The ideal number of direct reports is not limited, and having as many as one can handle is considered a good model, with examples of successful managers having 45-50 direct reports. (49m4s)
• The current number of direct reports is around 12. (49m14s)

Which Part of the CEO Role Daniel Struggles With Most? (49m17s)

• The role of a CEO involves handling people in different ways, and there are two approaches: a direct style that may be perceived as rude or tough, and an indirect style that avoids confrontation (49m18s).
• The direct style is actually a sign of appreciation, as it shows that the CEO is invested in the person or issue, whereas the indirect style can be a bad sign, indicating that the CEO is not fully committed (49m52s).
• As a leader, the CEO has not changed much in essence since being 17 years old, but has gained a better understanding of what it means to run a company (50m11s).
• The CEO believes that their core personality and leadership style have remained the same, and they would have run the company in the same way even at a younger age, despite lacking experience (50m21s).
• The CEO's leadership style is shaped by their personality, and they have learned to work with people in a way that is authentic to who they are (50m27s).

The Recent Decision Daniel Wishes He Could Undo or Do (50m45s)

• A decision that is regretted is hiring for experience and compromising on chemistry, as it is believed that chemistry should not be compromised for experience (51m7s).
• Something that should have been done earlier is landing into a "Gen 6" six months earlier, which could have made the company even faster (51m25s).
• The company does not feel like it is behind, as it is still very early, but acknowledges that it could have been even faster (51m33s).
• Unlike Benel, who stated that they are not hiring any more software engineers, the company has repurposed a lot of engineers from products that they deemphasize into Agent (51m51s).
• The existing software engineer cohort has been made more efficient from the tools they are now using, mostly AI tools (51m59s).
• Internally at UiPath, the efficiency within development teams is expected to change, but not as drastically as in companies like Salesforce, due to the complexity of the technology being built (52m11s).
• The company expects to get some productivity improvement, but does not think it will be gigantic (52m36s).

How Daniel Balances Gratitude with Ambition? (52m41s)

• A person reflects on their life, feeling that their apartment is not as nice as others, and constantly thinking about what would make it better, but realizes that this mindset is a waste of time and energy (52m42s).
• This person understands that having material possessions does not bring happiness and that what truly matters is how one thinks day by day, and that focusing on a bigger kitchen or other material things takes away from more important pursuits like reading, understanding the world, and understanding people (53m51s).
• It is acknowledged that the world is an amazing place where one can build incredible things with almost zero capital, and that there is no single material possession worth spending time wanting (54m12s).
• The person believes that this is the best time in history to be alive and that they feel better now than they did in their 20s, both physically and mentally, due to not wanting things and feeling free (54m43s).
• This feeling of freedom comes from having peace and not wanting anything, not in a Zen-like state, but in the sense that one is trying to do their best without being attached to the outcome (55m20s).
• However, it is also acknowledged that as a leader, one has responsibilities and that the outcome of certain events, such as a company's stock price, can affect others, like employees, and the ability to attract talent (55m52s).
• When asked for advice for a new father, the response is to enjoy the ride, as it starts off difficult but becomes more enjoyable with time, and to take time off to spend with the child, even if it's just two weeks in the first two years (56m30s).

Quick-Fire Round (56m57s)

• Lack of discipline is believed to be essential in stimulating creativity, and it has empowered the individual in their personal and professional life (57m4s).
• The hardest part of being a CEO is managing the unhappiness of people, as good news doesn't typically reach the CEO, but problems do (57m39s).
• Being a CEO can be a lonely experience, but for this individual, it's not just about the job - they have always felt like a "lonely wolf" and spend most of their time thinking, analyzing, and reflecting (58m12s).
• Switching context can be challenging, whether it's going from work to spending time with family or discussing different topics, as the individual's mind is always focused on their work and thoughts (58m43s).
• Providing context for daily tasks and decisions can be difficult, making it hard to have casual conversations with others who may not understand the complexity of the work (59m11s).
• Unmade decisions can weigh heavily on a person, and for this individual, the biggest decision that weighs on them is whether they should have pursued a career in San Francisco at a younger age (59m32s).
• Another decision that weighs on the individual is whether they should have enjoyed life more, as they didn't have a hedonistic life and experienced anxiety and self-doubt in their younger years (59m56s).
• Despite the option to give up on their current pursuits and focus on more pleasurable things, the individual feels driven by a powerful aspect of themselves to continue on their current path (1h1m12s).
• Working at Microsoft was a difficult experience, feeling lonely and depressed, but it was seen as a necessary step to build something later, and this internal conflict was a constant struggle (1h1m19s).
• The journey of UiPath, from zero to billions, was not without its challenges, and there were times when success was questioned, particularly before raising any money, during the 10 years of bootstrapping the company (1h2m6s).
• Raising capital and gaining people's trust was a turning point, and the fear disappeared, allowing the company to move forward successfully (1h2m13s).
• The most terrifying moment was in March 2020, with the onset of COVID-19, when it seemed like the world might stop, and the company would have to fire almost everyone to survive, but this moment passed quickly (1h2m48s).
• Dealing with moments of intense stress requires coping mechanisms, such as writing poetry, which helps to forget the source of stress and tap into creativity (1h3m16s).
• When it comes to investing in AI companies, the choice between OpenAI, Anthropic, and X.ai would depend on the upside, with Anthropic being a potential choice due to its bigger upside in the Enterprise space (1h4m3s).
• Being the CEO of SpaceX for a day would be an incredibly fun experience (1h4m44s).
• In 5 years, UiPath's success would depend on its ability to have a second act, which is a challenging task for companies, requiring not only capital and good developers but also a high degree of luck, and the agentic space is seen as the biggest opportunity for this second act (1h5m1s).
• Daniel Dines expresses gratitude towards the host for having him on the show again and mentions that he has learned from the host in many ways, reflecting on their previous conversation (1h6m13s).
• The host reciprocates the sentiment, stating that it's been a pleasure having Daniel Dines on the show and that they have enjoyed each other's company (1h6m21s).
• Daniel Dines also mentions that he enjoys being on the podcast and values the time spent with the host (1h6m26s).