Case studies @ Expana
Building a global pricing and forecasting platform
My role
Director, Product Design
Sep 2023 — Current
My Team
4 designers · 2 researchers
Reporting to the CPO, I led design and research through a company-wide transformation — building a new platform from the ground up to replace several legacy products, hiring and growing the design team, and embedding design and research into cross-functional decision-making.
Strategic objectives guiding my team's work over the first 2.5 years:
Migrate customers onto a unified platform
Outcome
58% of accounts migrated in 18 months.
Key accounts renewing with zero churn, versus significant churn on the legacy platforms.
Grow revenue per customer through cross-sell and premium upsell
Upgrade to view forecast
Outcome
Down 16% on pre-migration levels 18 months in, tracking to level by 24 months.
Growth challenged by increased competition and a trend toward smaller, leaner subscriptions
Grow the user base by solving problems for new roles and industries
Outcome
Work started in 2026 — design and research initiatives underway; happy to walk through the approach.
Work shown in the case studies below was produced by designers and researchers on my team, with my occasional hands-on input.
Supercharging Cost Modelling with AI
Why Cost Models got the AI treatment first
Building a cost model from scratch was time-consuming and technically demanding, preventing many users from getting value from them at all.
Cost Models were central to adoption and retention, so cutting that friction had real potential to make a difference.
Defining the problem
Design, product and engineering joined together to review the full cost model journey to identify frictions, pain points, and opportunities where AI could add the most value.
We prioritised model creation: reducing friction there had the clearest path to wider adoption, which was the outcome that mattered most.
We set out to address two pain points in model creation:
Pain point #1
Time consuming
For larger customers who needed dowzens or even hundreds of cost models, the sheer volume made building them time consuming and costly.
Pain point #2
Unknown recipe
Users often didn't know the precise recipe of the product they were modelling, and if they did, mapping it to Expana's data was difficult.
Design and engineering in parallel
While engineering focused on the backend, mapping prompts to Expana's time-series data and model logic, my team explored ways to integrate prompting into the existing UI.
When engineering moved to the frontend, we threw out the usual process of design → handoff → build. I did not want design to be a bottleneck. Not only that, we needed engineering to help uncover some of the possibilities and constraints
We worked without formal handoffs, just lightweight Figma work and real-time problem-solving. Where designs didn't quite fit the implementation, we fed that friction back into the next iteration.
Introducing AI into the Expana platform
Trust had to come first (our customers rely on the accuracy of our data) so AI-generated content needed to be clearly distinguishable from human-written content, and our commitment to AI visible across the product.
We gave AI features in the platform a dedicated colour and used the standard sparkles icon to help users identify AI content.
We kept the AI sub-branding light, and used tooltips to help users discover them.
Structuring the prompting workflow
Early on in POC phase, we learned that our thinking around the user journey for AI cost modelling needed some attention.
How we started: Build a cost model from a single prompt
On paper, this appeared to be the fastest path for a user creating a cost model:
User provides a prompt
eg. ‘build a lasagna cost model’
AI generates cost model
model includes full recipe & weightings / calculations
In practice this resulted in long wait times and low accuracy, leaving users with dozens of fields to correct — each correction triggering another lengthy wait.
Introducing a staged prompting workflow
We agreed that breaking the workflow into stages, allowing the model to focus on narrower tasks at each step, was the right approach. We settled on the following flow:
User provides a prompt
eg. ‘build a lasagna cost model’
AI determines the recipe
The recipe is written in common language and easy to review.
The user modifies or confirms...
AI matches the recipe with Expana data
A short wait is rewarded by results that align with user expectations.
The user modifies or confirms...
AI estimates weightings
The user now has a cost model which generates pricing analysis of their product
The examples below show the AI builder generating a croissant cost model using the staged approach (the video has been sped up)
Accuracy improved and the staged workflow was, counter-intuitively, faster than the single-prompt alternative.
We had to commit some effort before we could validate this approach, but it proved to be the right call.
Continous improvements
Feedback came through multiple channels, including user testing and customer input via our commercial and customer success teams. When prioritising, we considered volume of requests and whether an issue represented a migration blocker or churn risk.
Quick replies
The typical workflow for building a cost model involved some fairly predictable user inputs. To reduce the need for typing, we introduced quick replies: predefined prompt options users could select instead of typing from scratch.
Giving users more granular control
Updating a model by prompting was a blunt instrument: useful when you wanted to hand control to the AI, but risky when you needed a more precise update. We introduced lockable fields, giving users the ability to protect specific fields from AI modification.
Outcomes
We launched version one of AI Cost Models five months after kickoff, measuring success through a combination of usage, customer feedback, and commercial impact.
Usage
Six months after release, the majority of cost models built on Expana were created using the AI integration.
Every cost model a customer creates represents increased stickiness.
AI Cost Models as a sales tool
In an early example of the feature leading directly to customer subscriptions, Aldi signed a €1.5m three-year commitment, with AI Cost Models cited as a deciding factor.
Customer buzz
At a trade show, a high-value customer was found spontaneously demoing AI Cost Models to their team at our unmanned stand, highlighting strong customer interest in the feature.