Time moves fast in the world of AI. In 2019, UiPath founder and CEO Daniel Dines was dubbed by Forbes magazine as “the first bot billionaire,” after his company — which automates processes for enterprise clients — landed a $568m Series D at a $7bn valuation.
Fast forward five years, and the commentariat has changed its tune. The launch of GenAI startup Anthropic’s “computer use” product, an AI model that controls a computer in a similar way to a human, prompted the penning of multiple articles heralding the “death” of UiPath’s core automation offering, known as robotic process automation (RPA).
The argument: that powerful GenAI systems like Anthropic’s will render RPA, which uses more traditional software techniques to automate narrow, defined business processes that follow clear rules, obsolete.
But Dines — still running the now publicly-listed UiPath, which employs 4k people and serves some 10k business customers including British Airways, Canon, EY and Uber — seems entirely relaxed, and in fact excited, about how powerful new AI will impact his company.
Last week he was in London, for the opening of a new office and AI “Centre of Excellence” in the UK, where he says his growing team of data science and machine learning experts are cooking up some “fun toys” to stay ahead in the automation race.
Dines sat down with Sifted to share his vision on where enterprise use of AI is going, why he’s investing in London and why he believes generative systems aren’t going to kill RPA any time soon.
“More reliable, faster and cheaper”
While pundits enjoyed casting tools like Anthropic’s computer use as a threat to RPA, Dines sees these kinds of industry advances as an opportunity; UiPath has partnered with Anthropic and uses its models (LLMs) in some of its products.
“The use cases for computer use are very complementary to the use cases for RPA. It’s almost zero overlap,” he says.
Dines says RPA — which accounts for the vast majority of UiPath’s revenues — is very good at carrying out repetitive but important processes, where mistakes are not acceptable, like automating invoice processing.
This is because RPA relies on writing specified software scripts to follow defined rules, making their outputs “deterministic” (you know it will produce the right answer, like an Excel sheet formula), as opposed to LLMs which are “probabalistic”, meaning that they produce their outputs based on statistical best guesses and, therefore, can make mistakes.
“Imagine an LLM driving all of this [invoice processing] in a non-deterministic fashion. No sane customer will run such a thing,” says Dines. “What’s very important for our customers is that if there is uncertainty, they prefer not to use any [artificial] intelligence. The risk appetite is very low.”
He says that generative AI would have to improve by “orders of magnitude” to start making any inroads into the RPA market, and that using a software script will remain “more reliable, faster and cheaper” for the foreseeable future.
Enter the agents
This isn’t to say that UiPath is uninterested in generative AI, and it is investing heavily in using the technology to operate different types of business processes.
“We call it the ‘unstructured’ parts of the process that involve decisions,” says Dines. “For the parts of the process that cannot be described in rules, this is where I think there’s a lot we can do [with LLMs].”
For these kinds of processes, UiPath is building so-called AI “agents” — tools that can automate processes without needing step-by-step prompting. He gives the example of a travel agent tool for enterprise, where employees can book travel and accommodation, with the agent navigating different variables like tiered company travel policies, hotel discounts or special deals with airlines, to suggest the best options to a human user.
Dines says that agents are best used when limited to making suggestions, rather than carrying out jobs that involve money changing hands.
“Agents initially make recommendations. People will validate the recommendations, and I can say, ‘Okay, give me option one.’ I’m not going to an LLM to make the booking, because booking should be more precise and should be reliable,” he tells Sifted.
Dines says that while UiPath isn’t building frontier models, it has built a number of specialised models on top of third-party LLMs that are more accurate for their given task than general purpose systems.
He adds that his team is now working on building its own version of Anthropic’s computer use tool, in combination with Paris-based AI startup H, which UiPath invested in earlier this year.
“H is a very ambitious company, and they have the talent that puts them in the top tier of talent in terms of people building LLMs. They’ve built their own LLM from scratch and are competing directly with Anthropic’s computer use,” says Dines.
Why London
While Dines is in town for the opening of UiPath’s new centre of AI excellence, the company’s history with London’s AI scene isn’t new — the company acquired natural language processing startup Re:infer in 2022, a company that had strong links to University College London (UCL), one of the UK’s top universities for AI talent. Access to this pool of bright AI minds was the main reason to base the new R&D hub in the UK.
“UCL is the centre for us. We have David Barber, who is the head of UCL’s AI centre and was one of the founders of Re:infer,” Dines tells Sifted. “David works for us part time and of course his name attracts talent. The founders are really well known in the community in the UK, so I’m really pleased with the level of talent that we have attracted here, and we keep growing.”
UiPath’s current UK team totals around 40, and the company will be growing that number by hiring more data scientists and ML engineers. Whether UiPath will help lead the next era of automation as it did the last may depend heavily on how its efforts from its new London office play out.
Read the orginal article: https://sifted.eu/articles/ui-path-agents/
