Sebastian Schrötel is Senior Vice President of Product Management at software company UiPath [PATH], which specializes in robotic process automation (RPA) for those laborious repetitive tasks previously performed by humans — think data entry, file transfers and form completion.
An expert in Integration Platform as a Service (iPaaS) models, Schrötel leads UiPath’s Autopilot project, which combines generative and specialized artificial intelligence (AI) to help companies enhance their productivity.
He previously spent 16 years at SAP, including as Vice President, Head of Product for Application Development, Automation, and Integration, and was a founding member of SAP’s Machine Learning Initiative.
Schrötel joined OPTO Sessions to discuss where automation is heading next. He kicked off with today’s hot topic, agentic AI.
“Agentic automation adds goal-based and intent-based agents to [classic automation],” Schrötel explains. “[Agents] can react more flexibly to the context and actual problem in space. They can make decisions better and they can also interact with humans in the loop.”
However, for complex models to work efficiently, Schrötel says “orchestration” of human and non-human elements is crucial. “We are orchestrating AI agents, robots as classic deterministic automations, and humans.”
Low-code Load
Founded in Romania in 2005 by Daniel Dines and Marius Tîrcă, UiPath has grown into a multinational company active in 100+ countries worldwide, with more than 10,800 customers.
When it comes to strategy, Schrötel says, “the task for the next two to three years is to unlock as much value as possible with agentic orchestration and agentic automation in our customer space.”
UiPath’s enterprise solutions platform follows a subscription model which is “very typical for cloud platform providers… it depends on the use cases and on the scoping.”
For ease, the company focuses on low-code tooling, like drag-and-drops or pre-built elements.
“Why do we bet so hard on low-code? Because we believe low-code is visual programming, and visual tools are an elementary part of reducing the barrier to doing something new in the platform, and a big catalyst of developer efficiency and speed of implementation.”
The Agentic AI Boom and Hallucination
With AI generally, one ongoing worry is unreliability.
“The prevalent term is that AI models ‘hallucinate’,” says Schrötel. “If I want to automate a finance process or sensitive business process in my company, of course I don’t want any hallucination on top of my automation. I don’t want to have an AI agent that takes a decision I don't trust.”
The solution is integrating robotic and agentic AI automation to offer “control agency”. “This gives [agents] an enterprise-grade framework to govern, control and still have the deterministic part of the platform. The traditional automation part is still very relevant, especially for highly repetitive and sensitive processes.”
Retaining an old-school human element is important for UiPath, he says, rather than letting agents “roam free”: “if an AI component is not 100% sure what to do, they bring humans into the loop”, giving rise to a model where “humans, robots and agents are orchestrated together in one platform, working closely together.”
Real World Use Cases
UiPath already has supported numerous companies with custom solutions, among them insurance company Hiscox [HSX], which employs around 3,000. “They’re using what we call communications mining: a capability that understands large inputs of unstructured communication: email inboxes, chats, and so on.
“By having the capability to process and interpret and automate and extract sentiments… they could reduce the repetitive workload by 28%.”
Another client was betting firm Flutter [FLUT]. “They implemented Autopilot, our agent that interacts as a digital assistant directly with humans. With its help they were able to reduce… customer transfer escalations from 20% to 6%. It was a good example where AI and technology do not just create cost efficiencies, but directly impact customer satisfaction.”
There’s also global chip maker Intel [INTC]. “They combine our gen AI approach with our specialized AI models. They use it for classifying products and different material codes.
They were able to reach a 99%-plus accuracy in product classification. It just took them just 16 weeks to implement, creating huge value and reducing a lot of manual steps.”
All AI is not the same, and Schrötel underlines the difference between generative AI for public use — like ChatGPT — and the specialized models UiPath builds for specific tasks using a company’s specific sets of data.
“A specialized model can achieve way higher accuracies in a certain task than any generic model. It also solves the problem of ‘I’m training this model on my private and confidential data’.”
The Need for Speed
One of the startling things about AI is its speed.
“You can implement an automation and an agentic automation on the UiPath platform in a few hours. Is it something that then solves all the edge cases? Maybe not.” But if everything is in place, the implementation time can be just a few days.
From there, maturing time is key. “This is something our customers often do. They create a version one, they test it over time, then they create a version two… Both the deterministic parts and non-deterministic parts get better.”
Ultimately, Schrötel stresses, “if the input is not good, the output can never be better. For AI, this matters even more, because AI models can easily take in way more data than a deterministic model. Especially as they can take a lot of unstructured information.”
For real efficiency, orchestration helps, with humans introduced into a support system via certain rules. “You could say, I’m implementing thresholds. My agent can automatically clear incoming invoices up to an amount of, let’s say, £10,000 pounds. If it’s above £10,000, I need to have a human in the loop.”
Biological intelligence may not be totally obsolete, after all.
Continue reading for FREE
- Includes free newsletter updates, unsubscribe anytime. Privacy policy
