Peak
2024–25Product Designer · AI Supply Chain
Overview
I led design strategy for Peak's AI-powered Production Planning platform, a decision intelligence tool used by Tesco, Nissan, and Adidas. The interface translated complex ML demand forecasts into decisions supply chain managers could actually act on. The work cut overproduction by 50%, improved production efficiency by 30%, sped up response time to disruptions by 90%, and reduced retail stock-outs by 18% across major accounts.
The challenge
Peak's data science team had built an AI that could forecast demand at the SKU level, not the broad-category level most supply chain tools used. The forecasts were accurate. Planners didn't use them.
Instead, they ran their own spreadsheets on the side, adjusted plans manually after disruptions had already happened, and treated the AI dashboard as a thing they logged into every morning to keep the IT team off their back.
The reframe
Forecasting accuracy wasn't the bottleneck. Trust was.
Talking to planners at Tesco and Nissan, I heard the same thing every time: the AI was making a recommendation, but the interface was telling them what to do without telling them why. They couldn't see the signal behind the prediction. They couldn't override it when their on-the-ground knowledge contradicted the model. And they couldn't act on a single SKU-level decision without breaking the procurement plan downstream.
The redesign wasn't about better forecasts. It was about giving planners enough visibility and control to trust the AI in the first place.
Key decisions
SKU-level interface, not category
Most decision intelligence tools aggregate to category to keep dashboards readable. Planners need SKU-level precision because that's the level they actually order at. I designed dense interfaces that treated SKU-level data as a feature, not a problem to abstract away. Tradeoff: dashboards got harder to skim. Win: planners stopped running their own spreadsheets on the side.
Confidence scores and manual overrides
Usability testing kept telling me the same thing: planners wanted to argue with the AI. So I added confidence scores on every forecast and a manual override that didn't just replace the number but logged the planner's reasoning. The model used those overrides as feedback. Tradeoff: more friction on each decision. Win: trust went up, and the AI got smarter from human contradiction.
A unified view across sales, procurement, and production
Planners couldn't act on a SKU decision without breaking procurement. I redesigned the dashboard around a single shared view where every team (sales, procurement, logistics, finance) saw the same numbers in their own context. One source of truth instead of five conflicting versions. Tradeoff: we had to fight for ERP integration to land the data right. Win: cross-team conflict on planning calls dropped sharply.
Insight cards over complex charts
Data scientists love charts. Operations managers need decisions. I replaced most of the dashboard's traditional viz with prioritised insight cards like “X SKU is trending 20% above forecast at this site, suggest reordering by Friday”. Tradeoff: less raw data on screen. Win: non-technical users actually used the product.
Impact
50% reduction in overproduction and inventory holding costs.
30% increase in production efficiency.
90% faster response time to supply chain disruptions.
18% reduction in retail stock-outs across Tesco, Nissan, and Adidas.
Reflection
Peak taught me that AI products fail at adoption, not at accuracy. By the time a forecast hits a planner's screen, the question isn't “is the model right?” It's “do I trust this enough to act on it?” That's a design question, not a data science one.
The most senior thing I did at Peak wasn't a screen. It was reframing the team's view of the problem from “better forecast” to “enough visibility for a human to act”.