Keelvar AI (Kai)

Designing agentic AI to solve real user pain points in procurement.

Problem

Autonomous Sourcing users in Keelvar, an e-sourcing platform, aren't always procurement professionals; they often make ad-hoc, tactical purchases—like sourcing football shirts for an event, hiring a marketing agency for a campaign, or ordering raw materials to meet a surge in demand. While leading design on the product, usage increased by 10x in the last year. However, analytics and customer feedback showed that in some cases, customers were never completing their sourcing requests because configurations were too complex, or forms were too complicated to fill in.

Solution

From these issues, we identified our initial goals: reduce the time taken to launch autonomous sourcing requests and improve the drop-off rate for launching a request. With the rise of agentic AI, Keelvar saw an immediate opportunity to achieve these goals by generating requests from a simple prompt given by customers.

What does real value mean to Keelvars users when it comes to AI? When should, or shouldn't, you use AI to help your users?

We needed to answer these questions, and firstly, understand where would hit hardest in terms of the user experience improvement in the shortest space of time, so we could experiment and get feedback.

After creating initial concept designs, I facilitated an in-person workshop combining competitive analysis and sketching exercises, based on Google Design Sprint methods. This allowed me to rapidly align the team around a shared UX direction.

Ultimately, the team decided that a 'Canvas' view that separates AI output from the chat, enabling both AI and manual edits where needed, as the best starting point for a solution, as pages already existed in-app we could fork and modify.

After designing the initial UI using our design system guidelines, handing over to developers and establishing the end-to-end flow for the newly named “Kai” (Keelvar AI), I ran 13 user tests with customers and internal users. Below is a summary of the findings, generated by Google Notebook LM using session transcripts.

'This will really make it so nobody has to be trained on how to do anything. That's what it needs to get to to be super, super successful.'

They appreciated the speed and how vague prompts became concrete requests, but details mattered that were at times surprising. For example, we expected that 'thinking' messages for Kai would be deemed as noisy and background information. However we found the opposite was true, with users wanting visibility into Kai's 'thinking', for transparency into it's conclusions, so we kept these messages in an accordion.

This is only the beginning for Keelvar and Kai, but early results show that both initial aims are on track, with times to launch reduced and drop-off reducing. To further remove friction from the user experience, prompt suggestion and further methods for inputting information into Kai such as documents will be key to adoption. Utilising emerging AI prototyping tools such as Figma Make will be key to rapid ideation for future Kai instances.

This work was created at & is property of Keelvar.

UX research - helped the team define the problem and align on a shared design goal

User testing - validated ideas and identified areas for improvement before launch

UI design - established a look and feel for AI interface.