The iceberg of truth in AI projects

In our last article we wrote about why so many AI initiatives in marketing produce inconsistent results, and how the temptation to lead with tools often gets in the way of building something that lasts. This article picks up that thread, because once you stop looking at AI as a tool problem, the next question becomes: where does the work actually sit?

The iceberg is a common metaphor when describing almost any process that is more complex than they first appears, but we have decided to stick with is as the shape of it maps neatly onto where AI development tends to break down inside an enterprise. 

What sits above the surface

Above the water sits everything most people now picture when they think about AI. Generative AI as a category. The chat interface. Prompting. Image generation, text generation, document drafting, marketing copy, summarisation. The surfaces where a user types something and a model returns something usable, often within seconds.

It feels immediate, intuitive, almost magical. This is the part of the iceberg the entire industry is competing to design.

It is also, in our experience, the smallest part of what actually makes AI reliable inside an enterprise.

The prompt is the final delivery to the model that defines the output, but by the time a prompt is written, hundreds of decisions have already shaped and determined the quality of the output. Whether the model has access to the right brand context. Whether the campaign brief is structured in a form the model can read. Whether anyone has defined what “on brand” means in a way that holds up across teams. Whether the output is something one person reviews, or something that flows into a downstream system without a check.

A polished prompt cannot fix any of that. It can only make the surface look smoother.

What sits below

Beneath the prompt sits the part of the iceberg that determines whether AI produces something a brand can actually use.

It includes the data: campaign assets, brand guidelines, buyer personas, product information, prior approved content. It includes the context architecture: how that information is structured, indexed, and made available to a model at the moment it needs it. It includes the workflows: the sequence of steps that a piece of work moves through, with the right model called at the right time, against the right inputs, and the right human involved at the right point. And it includes the governance: who owns the quality of each input, who signs off on each output, and how that accountability is recorded.

Most enterprise organisations have some version of all of these things already, but spread thinly across people, tools, and document libraries. Bringing them into a form that AI can use reliably is not a small integration task. It is the central work of getting AI to behave consistently.

Why this matters for enterprise teams

We see two consequences of ignoring what sits below the surface.

The first is that pilots stay as pilots. A team produces an impressive demo. The output looks good, the model is capable, the prompt is well-crafted. But when the same approach is asked to operate at scale, against real brand standards, on a real campaign timeline, it falls apart. The visible part of the iceberg cannot carry the weight on its own.

The second is harder to see, and we think more damaging in the long run. Teams that lead with prompts and tools often end up rebuilding the same foundations several times, once per platform, once per use case, once per vendor. The investment is real, but it accumulates as cost rather than as capability. The structures underneath never get built, so each new AI initiative starts from the same place as the last one.

What we are continuing to explore

Through our research and conversations with people doing this work inside real organisations, we have a good sense of the shape of what below-the-surface work looks like in a B2B enterprise context. What is becoming increasingly clear is that the details are specific to each organisation, and they are not something we can wait for foundation model providers to deliver. The model providers are experts in general intelligence. Only you have the knowledge of what your own organisation needs.

The next articles in this series go deeper into that system. We will look at what structured marketing context actually means in practice, how briefs can be turned into inputs that an AI can work from, and where workflows replace prompts as the unit that enterprise teams should be designing around.

For now, the observation we want to leave with you is straightforward. If your AI outputs are inconsistent, the prompt is almost never the place to start looking.