Moving AI from pilot to production starts with the data

AI adoption has moved quickly. McKinsey’s 2025 global survey found that 88% of respondents say their organizations regularly use AI tools in at least one business function, up from 78% the previous year, yet only 7% say AI is fully scaled across the organization, with 30% remaining stuck in the piloting phase.

For marketing teams, this gap becomes even more obvious when an AI pilot meets the complexity of the existing data stack. A pilot can demonstrate what an AI model is capable of when it has a defined task and carefully selected data, but moving that capability into a live enterprise environment is often considerably harder.

The issue often sits beneath the model itself. Marketing data is spread across platforms, warehouses and internal systems, with different structures and definitions making it difficult for AI to establish what the numbers actually mean.

The pilot works until it meets the real world

There is a reason AI pilots can look so convincing at the beginning. The scope is usually narrow, the data is easier to control and the questions being asked are relatively straightforward.

Production however introduces a new and different level of complexity. An AI system working across an enterprise marketing stack for example may encounter several versions of the same metric, with different field names and different rules for how they should be calculated. Without a governed understanding of that environment, the model can act like a black box and make its own assumptions.

Cost is a simple example. One platform may store it under one field name while another uses something completely different. To a human who understands the organization's data, the distinction may be obvious, but to an AI model operating without that knowledge layer, the first plausible match can look like the right answer.

This lack of clarity creates a problem because the output you then receive can sound completely credible. AI can produce a confident answer from an incorrect assumption, leaving businesses with insights that looks useful until it reaches a business decision.

Data foundations need a knowledge layer

Improving data quality is therefore only part of the challenge. AI also needs to understand the meaning attached to that data.

For marketing teams in particular, this means establishing canonical definitions for important concepts and mapping those definitions across the different platforms where they appear. It also means capturing the rules the organization uses to measure performance, so that AI is working from the same framework as the people responsible for interpreting its outputs.

A knowledge layer provides that foundation between the data warehouse and the AI working with the data and encode how metrics should be interpreted, which fields correspond across platforms and which business rules need to be applied before an analysis is considered reliable. Such layers sit on top of existing data warehouses, meaning that organizations don’t have to make costly investments in replacing their infrastructure.

The key is giving AI a defined starting point. Instead of selecting whichever version of a metric it encounters first, the system has a governed understanding of what that concept means within the organization and can interpret it correctly.

Context needs to reflect what is happening now

Static definitions alone are not enough though. Marketing performance changes alongside the business, which means AI also needs access to the context surrounding a particular investigation.

A campaign may look as though performance has deteriorated, but actually the underlying reason for is a recent budget change. A sudden movement in results could also coincide with a new promotion or a live market test. Without that information, an AI system can identify the change accurately, but still reach the wrong conclusion about what caused it.

This distinction can be reflected by separating persistent marketing knowledge from context that is resolved for each investigation. This is where collaboration between marketing and data teams becomes important. Marketing leaders understand the objectives behind campaigns and how performance should be interpreted, while data leaders understand the systems, structures and controls through which that information is managed.

Bringing those perspectives together creates a more useful foundation for AI. Definitions can be agreed before they become embedded in automated workflows, while governance requirements can remain part of the analytical process rather than being introduced after an output has already been generated.

Production requires confidence, not another pilot

The next stage of enterprise AI adoption will depend on making these foundations operational. McKinsey's research shows that most organizations are still somewhere between experimentation and early scaling, despite the rapid growth in adoption.

The priority for marketers should therefore be to examine what sits around the model before adding another one. Are the relevant metrics consistently defined? Can AI understand how those metrics relate across systems? Does it know which business decisions are currently shaping performance?

A well-structured knowledge layer can provide the missing connection between enterprise data and AI reasoning. It gives models a governed understanding of the environment in which they are operating, reducing the assumptions they need to make themselves.

That changes the production question. Instead of asking whether a newer model is capable enough, organizations can focus on whether the AI has been given the knowledge it needs to make it work reliably.

The models will continue to evolve, but for businesses trying to turn AI investment into measurable value, the more immediate action is making sure those models have something dependable to reason from.

We've featured the best AI chatbot for business.

This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.

The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit