AI supply chain · decision intelligence

Getting planners to trust an AI they didn’t build.

Peak’s demand forecasts were accurate. Nobody used them. I redesigned the product around the real bottleneck, which was never accuracy. It was trust.

Role: Design strategy, product design
Timeline: 2024–25
In production at: Tesco, Nissan, Adidas

Peak's decision feed of plain-language insight cards with confidence and actions — The decision feed. Plain-language calls a planner can act on, ranked by impact and paired with confidence, not a wall of charts.

SKU-2841 · Sports drink 500ml · Reading DC

Demand forecast

90% interval

OctNovTodayDecJan

Reorder 400 units by Friday

Holds service level at 98%

−18%: retail stock-outs
−50%: overproduction
+30%: production efficiency
90%: faster disruption response

Context

Peak builds AI for the supply chain. Their Production Planning platform sits between a machine-learning demand model and the planners at Tesco, Nissan, and Adidas who decide what to make and order each week. I led design strategy: turning model output into decisions a human would actually act on.

A risk board ranking every SKU across sites by where service level is about to break — The board a planner opens first. Every SKU across every site, ranked by where service level is about to break, with the exceptions that need an order today surfaced to the top.

The challenge

The data science team had built something genuinely good: demand forecasts at the SKU level, not the broad category level most tools stop at. The forecasts were accurate. Planners didn’t use them.

They ran their own spreadsheets on the side. They adjusted plans by hand after a disruption had already happened. The dashboard was a thing they logged into each morning to keep IT off their back. A more accurate model was landing on screens and changing nothing.

In the field

I sat with planners at Tesco and Nissan and watched the workaround happen. The model would surface a number. The planner would glance at it, distrust it, and reopen the spreadsheet they actually trusted. The same decision, made twice, the second time by hand.

Affinity wall from planner interviews: printed forecast sheets covered in handwritten and sticky notes — Affinity mapping from planner interviews. The reframe came out of these rooms, not the dashboard.

The reframe

It was never an accuracy problem. It was a trust problem.

Better forecasting wasn’t the bottleneck. The interface told planners what to do without telling them why. It gave them no way to push back when their on-the-ground knowledge disagreed. And it couldn’t move a single SKU without breaking the plan downstream. The job wasn’t a better number. It was enough visibility and control for a human to act on the number at all.

How we rebuilt trust

Four moves, each aimed at one reason planners didn’t trust the model.

Move 01

From a category black box to a SKU-level decision

Most decision tools aggregate to category to keep dashboards readable. Planners order at the SKU. I designed for SKU-level density on purpose, and replaced raw charts with insight cards that say what changed and what to do. Dashboards got harder to skim. Planners stopped running spreadsheets on the side.

Before and after: an old category-level line chart next to a SKU-level insight card — The shift, side by side. A category aggregate you cannot act on, versus a SKU-level call with the reason and the action attached.

Move 02

From a black box to visible confidence

Every forecast now carries a confidence level and the signals behind it: promo, weather, baseline, seasonality. A high-confidence call surfaces to act on. A low-confidence one asks for a human before anything ships. Planners could finally see why the model believed what it believed, and the uncertainty around it.

A SKU demand forecast with a shaded confidence interval and a driver breakdown — Confidence as a band, not a single line, with the drivers exposed. Uncertainty widens into the future instead of hiding.

Move 03

From no say to a logged override

Usability testing kept telling me the same thing: planners wanted to argue with the AI. So the override does not just replace the number, it captures the planner’s reasoning and feeds it back as training signal. More friction per decision, by design. Trust went up, and the model got smarter from human contradiction.

A planner overriding an AI recommendation and logging the reason, which feeds the model — The override is the money moment. The planner changes the call, logs why, and that reasoning trains the next forecast.

Move 04

From five versions of the truth to one

A planner could not act on a SKU without breaking procurement. I rebuilt the product around a single shared forecast that every team reads in its own context: sales, procurement, production, finance. One number, no re-keying. Cross-team conflict on planning calls dropped sharply.

One shared forecast read by sales, procurement, production and finance — One forecast, four teams, one number. Each reads it in their own context instead of keying their own version.

What changed

Adoption was the metric that mattered, and it moved because trust did. Planners stopped keeping their own spreadsheets. The numbers followed.

“By the time a forecast reaches a planner, the question isn’t is the model right. It’s do I trust this enough to act on it.”

Retail stock-outs dropped 18% across Tesco, Nissan, and Adidas. Overproduction and holding cost fell by half. Production efficiency rose 30%, and the team responded to disruptions 90% faster, because a planner could see the signal, override with reason, and move one SKU without breaking the plan.

Reflection

Peak taught me that AI products fail at adoption, not accuracy. By the time a forecast reaches a planner, whether the model is right matters less than whether they trust it enough to act. That’s a design question, not a data science one.

The most senior thing I did at Peak wasn’t a screen. It was moving the team from “make the forecast better” to “give a human enough to act.”

Back to work