Most AI portfolios in mid-cap companies are not underperforming because the technology is immature. They are underperforming because nobody is allocating capital as if returns matter.

That is the dominant failure mode. Allocation indiscipline.

Boards are applying venture exploration logic to AI capex when they should be applying capital allocation logic. The two look similar from a distance. They are not. And conflating them is what produces AI capex that grows while EBITDA does not.

McKinsey’s April 2026 AI Transformation Manifesto confirms the pattern. The companies they identify as leaders reach breakeven in one to two years and concentrate their efforts on one to three business domains rather than scattering across dozens of pilots. The message is simple. Winners focus. Most do not.

The dispersion pattern

The pattern is usually easy to spot. Two or three initiatives carry real economic logic and a credible business owner. Several others were initiated after persuasive vendor conversations, conference enthusiasm, or executive curiosity, with sponsorship but without operational ownership, milestone discipline, or explicit exit criteria. And then there are the projects that nobody owns, nobody kills, and nobody scales. They drift.

The board reads the resulting dispersion as ambition. It is not. It is the visible output of a mental frame that confuses portfolio breadth with strategic engagement.

A recurring governance pattern helps explain why. Bpifrance Le Lab 2025 documents that 73 percent of AI projects in French mid-caps are launched by the CEO directly. This concentrates initiation at the top while leaving the operating layer without distributed ownership of the resulting initiatives. The CEO names the project. The organization waits.

The mental frame is the problem

Venture portfolios tolerate low hit rates because upside is asymmetric and downside is bounded. Mid-cap hold periods offer neither condition.

Here is the part that gets missed. The direct capex on an AI initiative may be bounded. The organizational cost rarely is. Each initiative consumes managerial bandwidth, attention, and political capital that cannot be recovered if the initiative is killed late or never killed at all. The asymmetry that justifies the venture frame does not exist. What exists instead is a deadline, an EBITDA expectation that must be defended quarter after quarter, and a buyer at exit who will discount obvious dispersion.

The right frame is standard capital allocation discipline. Every line of capex must point to a P&L outcome, a named owner, and a decision date. Initiatives that cannot defend those three are not investments. They are leaks.

The four criteria

A grid to restore allocation discipline before the next quarterly review. It is not complicated.

No P&L line, no portfolio entry. Every initiative must map to a specific line in the financials where its impact will be measured. Not a productivity metric. Not a satisfaction score. A revenue line, a cost line, or a margin line. If it cannot be located in the financials, it is not ready for portfolio inclusion.

The owner is a P&L leader, not a function head. The owner has to be a P&L leader sitting one to three levels below the CEO. Not the CIO. Not the head of innovation. Executive sponsorship helps. It does not replace operating ownership, and confusing the two is one of the most common allocation mistakes I have observed.

Concentration over breadth. If the active portfolio carries more than three serious transformation bets in parallel, the burden of proof should rise sharply on the marginal initiative. The question to ask is not “could this work?”. The question is “should this be one of our three?”.

Every initiative carries a 90-day kill or scale date. At that date, the initiative either receives confirmed budget and ownership extension, or it is closed cleanly with the freed capital reassigned. Continuation cannot be the default.

Two no’s out of four, and the burden of proof shifts from continuation to closure.

The reallocation logic

Closing an initiative is not destruction. It is reallocation. The capital, the managerial bandwidth, and the political capital invested in the closed initiative become available for the priority bets that can absorb them. That is the point. Closure releases. It does not waste.

The winners will not be the companies with the best AI stack. They will be the ones whose boards solved the allocation problem early. Technology becomes tractable once allocation discipline exists. The reverse is not true, and waiting for the technology to settle before fixing allocation is one of the most expensive mistakes a board can make right now.

The 90-day window

If the pattern is recognizable in a company you are watching, the next quarterly review is the moment to apply the grid. Identify the initiatives that pass all four criteria. Reassign the others. Communicate the closures explicitly, with named decisions and a clear rationale, so the organization learns that continuation is earned, not assumed.

Boards that complete this exercise within 90 days are better positioned at exit. Those that postpone it will discover that impressive AI spend does not read as ambition in diligence.

It reads as undisciplined allocation. And the multiple reflects that.

Concentration is not managerial preference. It is capital discipline.