Many organisations were quick to jump on the AI train in the hopes of leading innovation in their respective fields. However, businesses are now stuck with the expenses without fully reaping the benefits of what the technology promised.
According to Graham Sheldon, Chief Product Officer of UiPath, one area where AI makes perfect sense is automation, addressing numerous business challenges related to operational efficiency. Sheldon sat down with Frontier Enterprise for an exclusive interview, where he discussed emerging use cases for AI and automation, his storied career in Microsoft, and how his experiences there have influenced his current focus at UiPath.
What was your transition like from Microsoft to UiPath?
I was at Microsoft for two decades. I started out in product management and met Satya (Nadella) fairly early on; I was his technical advisor for four years. I spent a couple of years in Dynamics, then moved over to search ads, MSN, and we launched Bing. I worked on various teams within Bing, from monetisation to algorithms, marketplace design, and relevance. Then I led an Applied Research incubation team before coming back to start a new project, which became Microsoft Teams. We launched Teams and grew it from zero to 300 million users.
During my time in the applied research team, I witnessed the first inflection point for deep neural networks. We used convolutional neural nets for tasks like language translation and speech-to-text conversion, and achieved perception breakthroughs in image processing. Things quietened for about 10 years, and then with the next set of transformers and advancements in hardware, it was clear that more advanced algorithms were emerging.
With GPT-3, I saw a new inflection point coming. I wanted to get back into AI, in a place where I had a good shot at changing the way we work again. I found a kindred spirit in Daniel (Dines, founder and CEO of UiPath). His focus on accelerating human achievement and productivity at the heart of the company appealed to me.
A lot of our hypotheses from a few years ago about how automation could enhance AI have proven correct. Around this time last year, we began discussing how automation is like the arms and legs, while AI is the brain, creating transformative experiences for knowledge workers. By bringing together the context that is critical for making the right decisions or recommendations to a user with an automation platform, we can perform meaningful actions. These bets, such as context grounding and building an integration service to support human-in-the-loop systems for critical actions and key decisions, have paid off for many customers. We’ve heard many stories of customers who were able to achieve significant results.
On the flip side, many customers have come to us after investing in AI without a clear purpose, ending up with a solution in search of a problem. They wrote a big check and now don’t really know what to do with what they bought — it’s just a hammer in search of nails.
How did the AI evolution that you mentioned change the way UiPath designs products?
Our strategy starts with a customer focus. We design from the outside in, focusing on the core value that the customer wants to achieve. Usually, if you talk to the right person, such as the CTO, CIO, CHRO, or COO, they will describe their objectives in terms of a customer journey. For example, they might want to improve customer satisfaction and loyalty. It could also be an employee journey, where they aim to retain, develop, and upskill the best talent. Additionally, it could be a supplier journey, striving for a seamless and consistent experience for their suppliers.
When looking at these journeys, whether anticipating business growth, driving employee engagement, or enhancing customer satisfaction, we ask ourselves, ‘What really drives that?’ Often, it’s the interactions with a customer. A great customer experience typically happens because the person they interact with has the right tools to answer questions confidently and quickly, leveraging their best skills. They’re not wasting time searching through a 300-page document, navigating seven different applications, or manically typing in everything you’re telling them. Similarly, doctors shouldn’t spend their nights and weekends transcribing medical notes into another system. They should focus on what they went to school for and are passionate about.
If you approach design from this outside-in perspective, you identify what humans add demonstrable value to, what they love doing, and what they excel at. From there, you can determine what AI can handle.
Where we’re going, and the way that we now think about this in terms of building a product for UiPath, is really about optimising the use of human and digital agents. The people are the customers, and we’re focused on creating a great end-to-end experience for them. The evolution of AI has significantly changed the way we think about building a product, as we now consider where humans, AI, or workflows are best suited, since AI isn’t the right solution for every problem.
Rules-based workflows remain essential, especially to ensure compliance with government regulations, like fair and equitable loan distribution. These regulations must be built into the system, as you can’t rely solely on AI, which might produce unpredictable or hallucinated results. When proving compliance, you need to attest that policies and processes were followed correctly. Therefore, it’s about using the right tool for the right problem.
There’s a definite role for both agentic AI and deterministic business rules in solving such problems. Instead of making final decisions, AI can make suggestions to users while adhering to policies. Our customers are realising that there’s a time and place for AI and for policy. By bringing them together in an automation platform, you can achieve outcomes in a way that is not possible in many other places.
When it comes to product development, how do you measure customer interaction with the product?
We collect all sorts of signals, both directly and indirectly. We run insider programs with preview customers and have a community product that’s free, where people are self-selected from our partner group who want to be on the leading edge. They agree to give us more feedback and allow us to see some of the data. We have specific sharing agreements to obtain this information. My experience working in Bing showed me that product management is still very similar — it’s hypothesis- and experiment-driven. In the enterprise world, sometimes it’s hard to do proper A/B testing due to the lack of volume and the high variability and noise, but you can still conduct reasonable testing.
Last year, we developed Autopilot for developers and released several features. Initially, we thought the key aspect of designing the workflows would be incorporating natural language to generate a scaffold for the workflow. However, we observed extensive use of the feature for generating expressions. In our system, if you create a workflow and need to include some business logic, you use a VB (Visual Basic) expression to do it. We saw much more usage of VB expressions than we expected, as remembering the syntax can be quite a pain in the butt. Consequently, we built a feature to help correct expressions that we hadn’t initially planned to develop, but due to its high usage, users demanded it.
Even with poorly written expressions, you can understand the intended purpose, and it’s beneficial because it enforces proper coding practices and patterns.
What are some of the things you have in your development roadmap?
In the very immediate term, we have Autopilot, which has been well-received by the insider preview community. That’s our first wave of innovation related to generative AI. The IDP (intelligent document processing) product has been growing incredibly fast, so building out new specialised models for different industries and functions is a big focus for us. There’s also significant growth around our test and process mining capabilities, where generative AI has been truly transformative. People used to think of test automation software as a sleepy market, but it’s picking up now that generative AI makes these tasks much easier.
Additionally, with our partnership with SAP, a lot of people are starting to pay attention to the test side of process mining because these areas are on the rise. It’s a big lift to transition from an on-premises system to the cloud, so you have to focus on the right processes and ensure they don’t break. Further out, I think large action models and agentic AI are probably the biggest near-term innovations that will be truly transformative for our market.
