You can generate a hundred times more, but you cannot review a hundred times more

Broad adoption, rare transformation

Two measures must be distinguished:Adoption of an AI tool, which is now widespread, and the transformation of workflow which surrounds her, which remains rare.

Nearly 90% of the CMOs surveyed experiment with AI, but less than 10% report capturing value in end-to-end marketing workflows.

McKinsey · Reinventing marketing workflows with agent AI · 2026 · CMOs surveyed — View source

Only 28 per cent of the organizations surveyed are undertaking a real overhaul of their teams and processes.

McKinsey · From campaigns to continuous growth · 2026 · organizations surveyed — View source

The same difference appears on the marketing side: 96% of those surveyed say AI is transforming their function, but 42% remain confined to an AI that assists humans on isolated tasks, and only 8% are starting to operate several agents on campaigns.

BCG · Making the agent marketing transformation a reality · 2026 · CMOs interviewed — View source

The scale gap therefore does not come mainly from the access to models. It comes from the difficulty of reconstructing the process around their new capabilities.

Where the bottleneck moves

The mechanism is simple, and this is what makes it structural: the generation capacity has been multiplied, the control capacity has not.

1. Before AI
   Slow production. The validation capacity — brand review, legal, professional expertise, arbitration — is dimensioned for this volume and does not constitute the constraint.
2. With AI
   The validation capacity remains substantially the same: it depends on a number of people, a time of attention and a level of expertise.
3. Consequences
   The bottleneck moves from generation to control. Upstream gains are resumed downstream, and the full cycle does not shorten.

Human involvement should be targeted, not removed

Putting one person behind each result makes the control cost proportional to the volume generated: that is precisely what had to be avoided. The viable model sorts decisions according to the consequence of an error.

• Low risk, frequent decision

  Automatic control over prevalidated actives. Humans define the rule, not every application.

• Intermediate risk

  Validation triggered by rules and thresholds, not in principle.

• High risk or exception

  Human expertise, arbitration, engagement. This is where judgment produces enough value to justify its cost and delay.

The issue is not "human versus AI". It is to decide where human judgment actually produces enough value to justify its cost and delay.

The cycle to be measured then becomes: generation → automatic controls → review of exceptions → publication → measurement.

Redesign the workflow, not just its tools

Embedded a model on a process designed for a small volume produces traffic jam rather than gain. Organizations that achieve measurable results are those that redefine work rules, responsibilities and metrics—not those that add a tool.

A sectoral example converges with this finding, apart from marketing: in the Bain study devoted to the life sciences organisations studied, only 20% deploy AI on a large scale with measurable value, and 93% of those who achieve it have put in place significant skills development programmes, compared with 61% of those still in exploration.

Bath · The human imperative: scaling AI across life sciences · 2025 · life sciences organisations studied — sectoral example, not generalizable to marketing — View source

Technology therefore creates value only when it is accompanied by new working rules, explicit responsibilities and metrics of the accepted result.