From one month to one week

This was an in-house role, not a client engagement, across two titles between 2017 and 2020.

The situation

HelloFresh operations ran across multiple markets on weekly delivery cycles. Answering a question about those operations took roughly a month.

That number is the whole case study. A month is longer than the decision cycle it was meant to inform, which means the analysis arrived after the decision had already been made without it. The business had adapted, sensibly, by not asking. People made calls on instinct and the BI team stayed fully occupied producing answers to questions that no longer mattered.

The constraint

A live business with weekly delivery cycles, so there was no window to freeze anything. The BI team was entirely consumed by ad hoc requests, so there was no slack to rebuild with. And the rebuild had to be justified against a roadmap that was already committed.

What I did

I founded an internal centre of excellence and grew the BI team, which sounds like an org chart move and was really a scheduling one. It created a group whose week was not spent answering tickets, which is the only way anything structural gets built inside a busy function.

I overhauled the data architecture. The old shape had accumulated the way these do, one urgent exception at a time, and most of the maintenance cost was in the exceptions rather than the core.

We built the company’s first data product. Not a report about a process, a piece of automation that replaced a recurring manual process. That distinction turned out to be the important one culturally: it was the first time the data organisation shipped something that did work rather than describing work.

We built supply chain models into the operational decisions and put machine learning into core operational processes, where it contributed measurably to margin.

The result

Time from question to answer came down from about a month to about a week, measured on the recurring operational questions rather than on one-off analyses.

The first data product removed more than two thousand hours of manual work a year.

Putting machine learning into the operational decisions moved group profit contribution margin by more than a percentage point. That is a small-looking number and it is the largest one on this page. HelloFresh had listed in November 2017, so the denominator was the whole company by then, and a point of contribution margin at that scale is a material absolute figure.

I have deliberately not converted it into euros here. Doing that requires stating the revenue base I was measuring against, and I would rather give you the percentage with its denominator named than an absolute number you have to take on trust.

The change that mattered most is not in the numbers. People started asking again. When the answer arrives inside the decision cycle, the volume and the quality of questions both go up, and the function stops being a reporting service.

What I would do differently

I shipped the architecture before the ownership model. I spent the following year retrofitting owners onto systems that already had users, which is a much worse job than assigning them at the start. Every table that went live without a name against it became a negotiation later, and I lost some of those negotiations.

I underestimated how much of the month was queueing rather than computing. I went after the technical latency first because it was the part I knew how to fix. A meaningful share of the elapsed month was a request sitting in a queue. Fixing intake would have been cheaper than fixing architecture, and I did them in the wrong order.