The Coalition of the Embarrassed - Vincent Tauzia
Vincent Tauzia
SubscribeSign in
The Coalition of the Embarrassed<br>The barrier to AI is not technological. It is strategic.
Vincent Tauzia<br>May 06, 2026
Share
I feel like we have reached a point of no return for the AI hype cycle. The market, the big tech, the large corporations, the consultants, the financiers, even the press all adhere to one narrative: ‘the AI revolution is right now’ and ‘there will be many losers’. The losers are those who don’t embrace AI, those whose job will be replaced. FOMO often works, but they seem to ignore that AI has not proven it has had much impact on the real economy. I am pretty sure they know, but they have too much capex, too much narrative equity, too many quarterly forecasts riding on the story to admit it, or to nuance their stance. They are all part of what I call “the coalition of the embarrassed”, people who privately know the productivity case is weak but are too committed to say so publicly.<br>Inputs everywhere, outputs nowhere
We hear a lot about inputs when it comes to AI: the trillions of tokens used by Meta or Visa, the adoption of AI tools, the surveys showing it is the main conversation in board rooms. But we don’t see much in terms of outputs. Last year, a widely circulated paper from MIT showed that 95% of pilot projects using generative AI failed to produce measurable economic value. It all looks like as if Robert Solow’s famous paradox “you can see the computer age everywhere but in the productivity statistics” in 1987 could be applied to our age of generative AI. Another economist, Erik Brynjolfsson, coined the term ‘J-curve’ to describe the lag, measured in years, the emergence of a new technology (and the related reorganization investment it requires) and measurable output gains.<br>What the rare winners actually do
Yet, some corporations have been able to demonstrate tangible economic gains from using AI. In a recently published report, McKinsey analyzed what 20 companies that succeed with AI actually do. They tend to invest deliberately into highly focused areas they know they have economic leverage, rather than spreading AI across all functions and departments. As a result, they delivered on average a 20% EBITDA uplift, reached breakeven in 1–2 years, and generated $3 of incremental EBITDA for every $1 invested. There are three things these champions excel at:<br>Deeply embed AI in the most critical business processes they have. To do so they give full ownership to business leaders (not IT) who have fully embraced AI and embed software / data engineers in their organization.
The 30 / 70 shifts: more than 70% of tech talent should be in-house, more than 70% should be “doer” engineers (not managers or consultants), and more than 70% should perform at competent or expert level. Small, elite teams consistently beat large armies of lower-skilled staff.
Treat platforms and data as strategic assets. Those companies invest in reusable tech platforms with dedicated teams, roadmaps, and budgets. They make data easy to discover, access, and consume, before shifting to data enrichment for sustained AI performance.
It requires organizational courage and sustained investment over time to get tangible results. And, even more important, it requires focus to avoid ceding to the easy attraction of spreading some “AI magic powder” across all functions.<br>We are compressing tomorrow’s experts
I should also name a short-term vs. long-term tension that could have even larger societal impact if not properly handled by all economic agents. A recent study run by KPMG and the University of Texas at Austin reveals that the most sophisticated AI users, the ones with the most impact, are above manager level, using it for delegating complex, multi-step tasks with clear objectives. These experts know what should be done and know how to do it. They are great at directing the AI and at checking the results, closing the feedback loop quickly when needed. It also means that the space of junior roles is being compressed, as described in a recent Harvard Business Review. Anthropic’s analysis of how Claude is being used shows usage is concentrated in software and analytical work, roles traditionally assigned to entry-level knowledge workers. This should not be managed lightly; the young graduates of today are tomorrow’s experts.<br>Two non-negotiables: observability and human-in-the-loop
We need experienced workers, a lot of them. It may seem obvious to many readers, but still worth articulating. First, there are 2 firm principles I think every AI-practicing organization should strictly enforce: observability and human-in-the-loop. Observability is the precondition of control: practitioners need to understand how AIs behave, how they communicate with each other and how they self-improve. It would be extremely dangerous to let closed loop run without possibility of human intervention. The US military has a rule called 3000.09 on...