There is a photo of me from 2008 in which I am eating Guglhupf and editing an Excel sheet at the same time. The photo is remarkable because it sums up fairly precisely how performance marketing worked back then: type with one hand, eat with the other, and hope the formula holds.
At the time I was the first and only employee at the Austrian offshoot of blueSummit, one of the largest SEA/SEO agencies in Germany. On paper that sounded like a big stage: Lufthansa, BMW, international campaigns. In practice, a good 40-60% of my working day consisted of something I now call “data yoga”. Export a CSV from Google Ads, export a CSV from Analytics, build a VLOOKUP, set up a pivot table, hunt for errors, replace umlauts, start again. Anyone doing serious campaign analysis back then needed the patience of a watchmaker and the frustration tolerance of a Viennese taxi driver.
In those years I built international reports with 15 tabs, with formulas chained into one another like a Swiss clockwork, and a contents tab you could use to jump between the sheets. I was proud of every single link. I also cursed them regularly, whenever a comma was missing somewhere and the whole construction collapsed in on itself. That was craft in the most literal sense: time-consuming, error-prone, and oddly satisfying when it worked.
A word nobody understood
In 2014 I joined SlopeLift, a performance marketing agency in Vienna. We called ourselves a “workshop” from the very beginning, and I remember the reactions well. Performance marketing and craft? The industry was in love with scale, with volume, with the next tool that would automate everything. Everyone wanted shortcuts. We built granular campaign structures anyway, wrote individual ad texts instead of cloning them across hundreds of ad groups, and made manual bid adjustments per device and time of day. We even had rule-based scripts that adjusted bids according to weather data.
Clients never paid for the formula. They paid for the interpretation.
Was that AI? Strictly speaking: no. Honestly speaking: the underlying idea was identical. Read data, recognise patterns, react automatically. The only difference was the tooling. We used Google Sheets formulas where neural networks work today. But the logic underneath, the conviction that granular data work enables better decisions, was already the same in 2014 as it is in 2026.
The value shifts
What happened during those years without anyone saying it out loud: the real value of our work shifted, gradually and without warning.
Clients, whether at blueSummit or later at SlopeLift (BAWAG, dm, Hervis, Austrian Airlines), never paid for the formula. They paid for the interpretation. For the moment when someone says: “ROAS drops every year in week 12, because that is payday week for this audience.” For the context that no spreadsheet and no dashboard could deliver. But 60% of our working time went into the vehicle that carried that interpretation. Into the formula itself, into the preparation, into the back and forth between export and import. The ratio was grotesque.
When reporting then moved to Looker Studio, something fundamental changed, and it was not the tool as such. What changed was the way we approached problems. Before: sitting on the same formula for hours, trial and error, asking yourself for the third time whether it is the old minus the new over the new or the other way around. After: describe the problem, get a solution, correct first time in 90% of cases.
That switch from “building a formula” to “describing a problem” sounds like a small step. It was tectonic. Because it raised a question nobody was asking out loud at the time: if syntax becomes a commodity, where does the value of the human lie?
That question stayed with me for the next five years. It has not been fully answered to this day.
Continue with Part 2: The point where AI stopped being a tool