How I read the media industry.

Media dashboards are good at telling you how efficiently money was spent — CPM, CTR, CPA, ROAS. They are much worse at telling you whether it created growth.

In my experience the biggest gains rarely come from buying the same plan more cheaply. They come from deciding differently where the money goes. At ALDI Suisse, media performance rose by more than ten percent on a smaller budget — primarily through how TV was allocated across channels and dayparts, and how the tactical out-of-home communication was adapted. That is not a negotiation result. It is an allocation decision.

So I start with the business objective, choose the KPIs that show progress towards it, and only then optimise the media KPIs underneath.

When I started building a media planning agent, the first scaffold let a language model guess the channel percentages. A guess can’t be reproduced twice, and nothing can calculate net reach or frequency from it.

So I turned it round. The model reads the brief, names contradictions and writes the hand-over. A deterministic engine does every number — allocation, reach, frequency, waste — and then attacks its own plan. Same brief, same plan.

I think that division of labour holds well beyond my prototype: use models where language and interpretation matter, use arithmetic where accountability matters, and keep a human signature at the end.

I’ve worked inside three holding groups and on the sell side of AdTech. Every one of those seats comes with incentives: volume commitments, preferred partners, inventory that needs to be filled. None of that is sinister. All of it shapes what gets recommended.

Independence, for me, isn’t a virtue to claim — it’s a structural position. With no network to fill and no inventory to sell, the recommendation can follow the objective. Sometimes the right answer is an established partner, sometimes a specialist, sometimes no partner at all.

The network is a resource. It is never the strategy.

When I started in 2008, a media plan was largely a list of channels and a budget split. Today it is a system: platforms, data, formats, partners, retail media, content, measurement — and increasingly automated tools inside each of them.

The strategist’s job moves from choosing channels to orchestrating that system so its parts work towards one objective — and from producing the plan to deciding which questions the plan has to answer. That takes people who understand the brand question and the technology underneath it, and who are comfortable saying which parts should be automated and which should not.

There is a lot of talk about autonomous agents buying media. I expect a slower and more useful sequence.

First, agents that research and recommend — reading briefs, comparing options, showing their working. Then agents that act inside narrow, approved boundaries: write access to buying platforms behind approval gates, every action logged. Full autonomy comes only where governance, measurement and trust allow it.

The interesting design question isn’t how much an agent can do. It’s where the human approval sits — and what a person needs to see to make it a real decision rather than a rubber stamp.

People increasingly ask AI systems for recommendations, comparisons and answers. Those systems don’t just return links; they interpret, evaluate and recommend.

That moves part of brand building from being visible to being understood: is what’s known about a brand consistent, current and trustworthy enough for a machine to recommend it? I don’t think this replaces media or search. I think it adds a layer that marketing teams will need to measure and manage — and that, today, few of them own.