Since Anthropic’s MCP protocol opened the way in November 2024 for direct connections between Claude and business tools, LinkedIn has transformed into a catalog of promises…
“I connected the AI to my Meta account, I save three hours a day.” The reality on the ground is less flattering. Not because the AI is bad. But because what we give him to eat often is. And between a correct figure and a false figure displayed with confidence, the line is thinner than you might think.
The illusion of the automatic analyst
The LinkedIn posts didn’t wait. In just a few months, there has been an avalanche of tutorials, “magic prompts” to extract insights from your Meta, Shopify, Google Ads campaigns. The narrative is well established: no more need for an analyst, no more need for training, the AI understands everything and tells you what to do.
We put this narrative to the test. Several times. On real accounts, with real data, with several tools. The verdict is clear: almost all of these setups produce unusable analyses. The AI calculates an LTV without excluding refunds. It adds VAT and VAT. It includes test commands that everyone forgets to mention. She ignores the channel she was asked to exclude three exchanges ago. And each time, it produces a clean, well-formatted figure, displayed with perfect confidence. Without warning. Without nuance. Just a number.
This is exactly where the problem begins.
AI amplifies but does not filter
There is a fundamental misunderstanding in the way the market presents AI applied to data marketing: that of compensation. The implicit idea is that AI will compensate for flaws in the data, fill in the gaps, correct inconsistencies. She doesn’t do any of that. She amplifies what we give her. Clean, contextualized, structured data, it becomes formidable. Raw and chaotic data, it produces more noise, faster, with more conviction.
This is the paradox that the prevailing discourse refuses to address: a false figure displayed with confidence is infinitely more dangerous than a figure which we do not really know if it is correct. In the second case, we know that we do not know. In the first, we think we know.
There is a second, even deeper blind spot. The AI responds to what is asked of it. She doesn’t know what question to ask. Understanding that a platform ROAS of 4 on Meta means nothing until you have calculated the real ROMI, margin included, logistics costs integrated, attribution model explicitly chosen, is a business skill. The AI will not spontaneously tell you that you are asking the wrong question. She will simply answer it.
It is a distinction that the current excitement systematically blurs. AI is an expertise multiplier, not a substitute. If the expertise is there, it saves considerable time. If it is not, it creates unfounded certainties.
What it really costs
The consequences are not abstract. An overestimated LTV, and we validate an overly optimistic target ROAS, we scale a channel which destroys margin without seeing it. A poorly understood attribution, and we overinvest in last clicks capturing demand already acquired. A ROMI calculated on an incomplete scope, and the loss-making operation remains invisible for three months, until the treasury reports it instead of the dashboard.
In paid acquisition, a wrong decision quickly costs tens of thousands of euros. The economic equation of “no more need for analysts, AI does it” does not hold up to this calculation. Gartner has documented it in 2024: data quality remains the primary factor in failure of AI deployments in an analytical context. What these studies measure in abstraction, the field teams experience in operational reality.
What policymakers should do now
These recommendations are aimed at marketing directors, acquisition managers and executives who manage growing media budgets.
Define metrics before automating anything. What is the real business compass? Platform ROAS is a tactical indicator. The ROMI, calculated incl. VAT, margin included, on an explicit attribution scope, is the only figure that allows you to decide. This definition must exist in writing, shared, before connecting any AI.
Centralize and clean the data upstream. As long as data is scattered across multiple platforms with different logics, no AI will produce actionable analysis. A centralization tool with an integrated semantic layer is a prerequisite, not an option.
Invest in critical team training. Mind-blowing probabilistic models are a structural property, not a bug to be fixed. The real challenge is to have teams who reflexively challenge outputs, not who validate them out of confidence.
AI in media performance management is indeed a paradigm shift. Not for the reasons that LinkedIn posts state, not because it replaces expertise, it doesn’t do that, but because, plugged into clean data and questioned by someone who knows, it compresses several hours of analysis into a few seconds. This is real leverage, not an empty promise.
But the order of operations is not negotiable. Clean data first. AI next. Reversing it means automating mediocrity more quickly. And in an industry where every driving decision counts, going faster in the wrong direction is not progress.
