Skip to content
Vision · Analysis 2 of 2

Why generating at scale is not enough to execute at scale

AI is able to produce in volume. What it does not produce alone is a reliable, governed and connected execution to the context of the company. Five documented obstacles explain where the constraint really lies.

Back to the new standards

What this analysis says

The cost of generation collapses. The cost of an accepted, consistent, disseminated and measured result does not collapse on its own.

Five obstacles, in the order in which they appear

They do not form a list but a progression: the volume moves the bottleneck, reliability forbids trusting the model alone, the scale transforms a brand error into a systematic incident, dispersion prevents capitalizing, and adding agents in turn creates a need for orchestration.

  • 01

    Volume increases and the bottleneck moves

    Generation is no longer the constraint. Review, trademark conformity, legal control and human validation become so.

    The gains obtained during the generation can be absorbed during validation: this is the complete cycle to be measured, not the only stage of production. The Human-in-the-Loop can itself become the new bottleneck.

    Explore this obstacle
  • 02

    Capacity does not mean reliability

    The performance of the models follows a serrated border — the jagged technological border — : very strong on some tasks, unstable on others, without the limit being visible to the user.

    A very high performance on some tasks does not imply stable reliability on all. The model cannot therefore be the only unit of trust.

    Explore this obstacle
  • 03

    AI can also scale up brand errors

    This makes AI useful — producing quickly and in volume — also makes a brand error reproducible at the same speed.

    • 63% / 80%

      large brands already use the generative AI in marketing, but 80% are concerned about how their agencies use it; the risks cited are legal, ethical and reputational

      Source: WFA, Annual Report 2024 · WFA member advertisers.

    BCG defends the idea of a brand intelligence layer : purpose, rules, identity, context and decision criteria encoded in the systems themselves. Without it, agents without organizational identity spread the same generic response on a large scale. On a scale, the brand can no longer remain a brand book for humans: rules, identities, constraints and validations must become executable by the system.

    Explore this obstacle
  • 04

    Companies must regain ownership of their marketing machine

    When data, content, rules, workflows, technology and learning are dispersed between teams, platforms and providers, each campaign recreates part of the knowledge rather than enriching a common asset.

    • 12% / 36%

      Internal agencies interviewed consider AI to be fully integrated; 36% consider integration into business systems still difficult

      Source: WFA, How AI is changing in-house agencies · Internal agencies interviewed.

    The real asset owner is not just the data. It's the learning loop: data → knowledge → content → activation → measurement → new learning. Internalizing without a system means resuming complexity without being able to absorb it — which also involves controlling data, permissions, security and intellectual property.

    Explore this obstacle
  • 05

    Connecting agents does not automatically create a system

    Complexity does not disappear with the addition of agents: it moves towards orchestration.

    Multiplying the agents reveals new classes of failures, additional costs and issues of permissions, observability and validation. More models and agents increase the capacity, but also the need for an orchestrator.

    Explore this obstacle
The principle

Don't trust the model. Trust the system.

A probabilistic model will remain imperfect. What makes its industrial use acceptable is the system that selects, contextualizes, limits, verifies, observes and climbs its decisions.

Operational impact

Change the unit of measurement

The right performance unit is no longer the number of content generated or the cost of a token. the cost and time required to achieve an accepted, compliant, disseminated and measured result.

This shift changes the discussion: it no longer relates to the performance of a generation model, but to that of an execution system.

What this implies

The transition to scale does not come from the model alone

It comes from the ability to connect proprietary context, rules, data, tools, validations and measurement in an operable and governed system.

See how CAIAC addresses this problem

Sources

Sources and methodology

The figures cited come from works published by third parties' firms, institutions and laboratories, and we take them back to the area of origin — the population interviewed, category, year, forecast or observed result — without broadening it. The analyses and convictions that accompany them are ours and are reported as such.

  • McKinsey, Reinventing marketing workflows with agent AI — www.mckinsey.com
  • Dell'Acqua et al., Navigating the Jagged Technological Frontier — doi.org
  • METR, Time Horizons — metr.org
  • WFA Annual Report 2024 — wfanet.org
  • BCG, Making the agent marketing transformation a reality — www.bcg.com
  • WFA, How AI is changing in-house agencies — wfanet.org
  • Google Research, Towards a science of scattering agent systems — research.google
  • Gartner, forecast of AI projects — www.gartner.com

Let's talk about your context.

Let's identify the current bottleneck in your marketing execution process.