Daniel Dines on Why AI Agents Need Workflow Automation

Harry Stebbings asks why, at this stage of AI, product matters more than innovation. UiPath founder and CEO Daniel Dines says his team spent two years fine-tuning and building around large language models, with some success. He was struck by Cursor, which uses multiple models but delivers an experience developers love.

Dines recalls an early UiPath product built with OpenCV. The library could find a small image within a larger one; UiPath used it to find a button in an application and click it. Users could record a sequence by pointing at the screen, and the software generated simple instructions and stored the images needed to replay the flow.

In a 2013 demonstration to Blue Prism specialists, Dines says he built and ran a flow in minutes that would have taken them roughly two days. Image recognition opened a niche in remote-desktop environments such as Citrix, where customers could not otherwise access applications. UiPath expanded from there.

Dines expects ease of use and product design to drive adoption. In his view, models have reached a degree of maturity and are unlikely to improve materially in a reasonable time; that is his judgment, not an established limit. UiPath uses Alibaba's Qwen for semi-structured documents because, he says, it was the best model for that particular job at the time.

The surrounding product helps customers tag documents and retrain the model as they work. Dines says UiPath can switch models as the balance of cost, speed and accuracy changes. He expects a few frontier models alongside many specialized models, often built on open models, rather than a market resembling concentrated cloud infrastructure.

The lesson for UiPath is to build AI-first experiences from the ground up, rather than only adding incremental features. Dines says the company replaced a longstanding workflow engine with a more modern one designed to connect agents, people, robots, models and APIs in what it calls agentic orchestration.

Stebbings asks why robotic process automation, or RPA, is compatible with agents. Dines says RPA is strongest in structured, rule-based processes spanning multiple systems, potentially with hundreds of steps. Those rules encode company knowledge, and automations remain reliable until the underlying systems change.

He sees language models as more useful for unstructured parts of a process and tacit knowledge that is difficult to express in rules. They are less suitable, in his view, for repetitive deterministic steps. Agents could imitate a user's work and reduce human input, although he cautions that fully eliminating a task is rare.

Stebbings asks whether businesses will buy rule-based and non-rule-based tools from separate vendors. Dines argues that both belong in a single end-to-end process such as order-to-cash or procure-to-pay. UiPath's commercial ambition is to extend its existing robot orchestration platform to agents and manage both within a common framework.

One automation is not enough, Dines says: companies must deploy and monitor thousands, gather analytics, and control who can run them and access applications. RPA and agents both imitate people; either can require retries and exception handling. His claim is that UiPath's experience managing robots is relevant to the harder problem of making agents work reliably at scale.

Dines says some customers prefer a workflow to fail rather than behave unpredictably. He expects agents to begin by making recommendations for a person to validate before an action is taken. Models can appear exceptionally capable in one instance and make basic mistakes in another, he argues, making it difficult to know when to trust them.

Asked how long full trust will take, Dines compares the wait to achieving fully autonomous vehicles, without offering a precise date. In the meantime he expects semi-autonomous agents to do much of the work while people monitor requests, validate outputs and provide missing information inside rule-based workflows.

UiPath wants to supply its own agents and integrate agents from other platforms, Dines says. In his model, a Salesforce agent could be called like an API, its response sent to a person, and the validated result used to trigger another action or agent across systems. He calls the desired orchestration layer a neutral Switzerland.

Stebbings challenges the idea that Salesforce will confine its agents to its own platform, noting its ambition in support and marketing. Dines responds with an anecdote: a healthcare customer told him it would not move Epic data into Salesforce merely to create an agent. He expects customers to favor connectors that give agents only the information needed, with orchestration across repositories.

Dines recommends starting with a business process rather than a job title: map procure-to-pay, for example, and distinguish the deterministic steps from those suited to agents. In healthcare he mentions denial processing and prior authorization as possible focused uses, connected to robots performing rule-based work. These are proposed applications, not proof of clinical or operational outcomes.

His warning about using agents for long rule-based chains is compounding error: even an illustrative 99% success rate at each step produces a much lower chance of completing a 100-step chain without error. The example is a thought experiment, not a measured performance figure for a particular product.

Stebbings asks why Wall Street has not valued UiPath more highly if it already has distribution and a position as an orchestrator. Dines says agentic automation is still early and UiPath must deliver its roadmap; its agent builder had just entered private preview. He expects more enterprise use cases in the following year, a forecast from a CEO with a direct commercial interest.

He attributes limited enterprise success with generative AI so far to unpredictability. His proposed remedy is to put models inside workflows, constrain them with rules and use people to validate results. For a mortgage application, an agent might make a recommendation, a person approve or reject it, and a robot update the banking system; this describes a possible workflow, not an autonomous lending recommendation.

Customer conversations changed Dines's emphasis from identifying isolated tasks for a quick return to understanding the entire process. He expects connected enterprise workflows to be more useful than standalone agents with which a business user simply chats.

Dines predicts that workers will increasingly oversee technology and validate its outputs. Stebbings worries that this turns people into 'validation monkeys.' Dines replies that, as trust grows, people could focus on exceptions: a travel agent might book automatically below a $10,000 limit and seek approval above it.

He is unsure whether agent builders immediately open the small-business market to UiPath. Effective prompting requires skill, slight changes can alter results, and testing is harder than for a predictable script because input data varies. UiPath is working on prompt suggestions and evaluation sets to help developers, he says.

On employment, Dines draws an analogy to agriculture: fewer people work in farming today while machines do more of the work. He expects some jobs to change and new ones to emerge, arguing that productivity is needed as populations age. His historical figures and optimistic employment outlook are his own estimates and opinions.

Stebbings counters that AI adoption might move much faster than the decades-long agricultural transition. Dines points to corporate inertia and estimates RPA itself is still less than 10–20% penetrated, while acknowledging that penetration is difficult to measure. Deployment requires a program, not just a tool handed to a worker.

With current models, Dines estimates that widespread deployment of agents plus automation will take five to ten years. He adds that a move to AGI could change the picture. He defines enterprise AGI as consistently performing at roughly the level of a person with an IQ of 120, rather than alternating between exceptional and poor performance.

Dines believes another major leap is needed to reach that consistency. He questions whether current language models reason as people do, citing their ability to solve very hard mathematics while making simple logical errors. His characterization of their intelligence is a contested interpretation, and his view of a training plateau is a prediction rather than a demonstrated ceiling.

On Nvidia, Dines says it is hard to imagine the company retaining its existing level of dominance indefinitely as major cloud providers build chips, although he also recognizes its hardware and software advantages. On pricing, he expects a mixture of seat-based and consumption-based models rather than an abrupt switch to paying only per unit of work.

Dines identifies his next 12 to 24 months' main challenge as making UiPath an AI-first company and re-energizing staff after a rocky public-market period. Looking back on the IPO, he says he would plan finance and go-to-market growth less aggressively and more consistently. His comparison of steady 30% growth with a declining sequence of higher growth rates is illustrative.

Discussing founder mode, Dines says he misjudged UiPath's maturity when he hired a CEO. He remained involved in product and engineering, and says the relationship required a difficult dance that could have benefited from more communication. During a technological shift, he believes a founder CEO can connect product, go-to-market and marketing quickly.

His approach to motivation is candid communication about the company's position, mistakes and work ahead. As a company grows, he worries people feel smaller and unable to influence decisions. He wants employees to drive changes and be heard, while reducing bureaucracy, giving regions more control and bringing customer feedback closer to product teams.

Dines doubts the value of scheduled one-to-ones with his direct reports if they cannot already call each other and speak honestly. He has about 12 direct reports and admires leaders who can handle more. He admits his own weakness: he can be very direct, or avoid telling someone what is wrong; he regards straightforward feedback as a sign of respect.

Looking back, Dines wishes he had not compromised chemistry for experience in hiring. He also wishes UiPath had moved into agentic technology six months earlier, while insisting the market remains early. Rather than hire additional engineers, he says the company repurposed some from products it de-emphasized.

Asked whether AI coding tools have dramatically improved engineering productivity, he expects some improvement but not a gigantic one for UiPath's complex work across multiple technologies. That assessment reflects his experience and should not be taken as a measurement across the industry.

Stebbings describes comparing his comfortable home with a nicer one and feeling guilty about always wanting more. Dines says he spent much of his late twenties, thirties and forties thinking similarly, and now sees it as wasted energy that could have gone toward learning about the world, people and AI.

Dines says having possessions did not make them worth the mental cycles spent wanting them. He feels better than in his twenties and finds freedom in doing his best without making the outcome the source of peace. Stebbings observes that wealth makes this easier; Dines acknowledges that the stock price still matters for employee morale, talent and what UiPath can build.

For a new parent, Dines's advice is to enjoy the journey: he found it difficult at first but increasingly rewarding over time. In the quick-fire round he also defends a degree of lack of discipline as important to his creativity and to what he has built.

The hardest part of being CEO, he says, is managing unhappiness: routine good news stays with teams, while problems travel upward. He describes himself as a lone wolf who spends much of his life thinking and reflecting, and says the loneliness predates being CEO, including his time as a Microsoft software engineer.

Both men discuss finding it hard to switch from work to ordinary conversation. Dines says explaining his day may first require hours of context, making it easier to talk about television instead. He worries that absorption in his work leaves him increasingly disconnected.

Asked what unresolved decision weighs on him, Dines wonders whether he should have enjoyed life more in his twenties and thirties. He describes years of anxiety and an internal conflict between seeking pleasure and doing what he felt was right. At Microsoft, he says, the final years felt lonely and depressing, but he viewed the work as a compromise that would enable him to build later.

He most doubted UiPath's success during the ten years before raising outside capital, which he calls the hardest period. The confidence others showed by investing changed his own mindset. Another frightening moment came in March 2020, when he feared the pandemic could force layoffs at a company still losing money; he says the immediate fear passed quickly.

Dines writes poetry as a way to cope with stress. Finding words and metaphors for pain, he says, engages his creativity and helps relieve it. Asked to choose among hypothetical investments in OpenAI, Anthropic and xAI at the valuations Stebbings gives, he picks Anthropic for its perceived upside; this is a personal hypothetical, not investment advice.

For a day as another company's CEO, he chooses SpaceX. Looking five years ahead, he hopes UiPath can achieve a second act in agentic technology. He says even capital and good developers do not guarantee a successful move into a new market, which also requires luck; his optimism is tied to his role as UiPath's founder and CEO.