Why 95% of Brand AI Pilots Fail, and What the 5% Do Differently

By Linda Cereda, former Global VP Marketing Data, Nike

According to a 2025 MIT study, 95% of enterprise AI pilots fail to produce measurable P&L impact. That number is not a rounding error. It represents the largest gap between investment and return in the recent history of enterprise technology, and it is getting harder to ignore.

The instinctive explanation is technical. The models are not ready. The data infrastructure is not right. The use cases need refinement. Some of that is true. None of it is the actual problem.

The businesses that are generating real AI value - the 5% - are not running better technology projects. They are running business transformation efforts that happen to use AI. The rest are doing the opposite. I have seen this pattern from both sides: building AI capability inside one of the largest brands in the world, and watching mid-market and PE-backed companies spend significant budgets on work that never moved from pilot to production.

The failure modes are consistent. Here are the five that show up most often.

Failure one: treating AI as a technology project

The most common version of AI failure looks like this: IT builds it, the business watches, leadership sponsors from a distance. The pilot works, but nothing scales, because no one in the business owns the outcome nor was the work designed with business context and people in mind.

The 5% do not run AI projects. They run business transformation mandates that use AI to get there. The starting point is a specific, measurable business outcome - a revenue target, a cost reduction, an improvement in a KPI that the business actually cares about, not a technology capability looking for a use case. That outcome is agreed before the first line of code is written. It is signed off by the business owner and one named leader owns delivery.

The question is never "where can we apply this technology?" It is "what business outcome are we buying, and who is accountable for delivering it?" Technology sits inside that mandate, not ahead of it.

At Nike, the AI work that generated results - for example a 44% reduction in forecasting error on SNKRS, a $26 incremental demand lift per engaged member - started with aligned outcomes: improve demand forecasting, drive fairness in member access, grow LTV. A commercial owner held the P&L accountability and drove the operating change.

This is not a statement against technology teams. It is a statement about where accountability has to sit. AI is not a tech project. It is a business transformation mandate with a critical technology component.

Failure two: the purgatory of bottom-up pilots

Walk into most large brands today and you will find dozens of AI pilots running in parallel, sometimes hundreds. Different teams, different tools, different data, different definitions of success. One client I worked with had three separate synthetic persona projects running across three business units. None of them knew the others existed. All three were built on different data, produced different personas, and consumed resources that could have driven one initiative to scale.

This is the purgatory of bottom-up AI. Lots of activity. Very little progress.

The pattern it creates is predictable. Duplicated work. Inconsistent data foundations. No single initiative with enough resources behind it to reach production. Two-thirds of businesses are stuck in exactly this state, in pilot phases with no clear path out.

The 5% do something different from the start. They orchestrate. Instead of thirty pilots at low intensity, they identify three or four company priorities that AI can materially move, and they pursue those with full conviction, real budget, real data, real ownership. Johnson & Johnson reviewed over 900 AI use cases and found that roughly 10% of them delivered 80% of the value. The answer was not to run all 900. It was to identify the ten that mattered and build toward them properly.

Think big. Start small. Scale quickly. But decide what you are scaling before you start.

Failure three: signing an AI tool and calling it an AI strategy

This failure mode is the quietest and the most expensive. A leadership team approves an AI tool. The contract is signed. The tool is deployed. Eighteen months later, the business is asking why nothing has changed.

What happened is that no one did the 70% of the work.

BCG's research puts the ratio at roughly 30% technology, 70% people and process. That number matches what I have seen in practice. A tool deployed on top of an old workflow produces a slightly faster version of the old outcome. The AI is doing the work the human used to do, but the work itself has not changed.

The 70% breaks into two things that most AI rollouts underinvest in equally.

The first is process redesign. The brands that generate real AI value do not deploy tools onto existing workflows. They rebuild the workflow around the AI capability. At Nike, AI-driven triggered communications moved from 4% to 13% of total send volume, not because a tool was added to the existing batch-and-blast process, but because the lifecycle workflow was rebuilt around AI-driven next-best-action from the ground up.

The second is people. A tool the team does not trust, understand, or know how to use will not get used. The AI champions, the training, the mental model shift from "I run this process" to "I direct the AI that runs this process", that takes active investment. It does not happen by default. The most common outcome of skipping it is a tool that sits alongside existing ways of working, used by a few enthusiasts, ignored by everyone else.

Signing the tool is easy. Rewiring the process and bringing the people with you is the work. Most companies do the first and underestimate both of the others.

Failure four: optimising for cost savings and missing the bigger opportunity

There is nothing wrong with using AI to save time and money. Automation has real value; Klarna's AI agent now handles two-thirds of customer service chats. That is a genuine result. The problem is when cost savings become the ceiling, not the floor.

Three things go wrong when AI strategy is primarily about headcount and cost reduction.

The first is that the full cost of AI gets underestimated. Saving ten people is straightforward to model. The cost of the tokens those ten people were not consuming is not. Unexpected compute bills have become a real problem for companies that sized their AI investments against one metric and ignored the others.

The second is that the savings themselves are not deployed strategically. Is the efficiency reinvested in growth? Does it fall straight to the bottom line? Or does it quietly disappear into the organisation without generating lasting value? Most companies have not answered that question before the savings arrive.

The third, and most consequential, is the opportunity cost. Companies laser-focused on automating what they already do are missing the question that actually determines long-term competitive position: what could we do that we have never been able to do before?

The companies building durable AI advantage are not just cutting costs. They are using AI to transform what the consumer experience looks like, e.g. real-time 1:1 personalisation, AI visibility in search, agentic commerce that anticipates what a customer needs before they search for it. L'Oreal's Beauty Genius does not save money. It creates a consumer experience that did not exist before AI.

Automation will soon be table stakes. The brands that will lead this decade are the ones also asking what transformation looks like.

Failure five: building on a weak data foundation

AI is only as good as the data underneath it. This sounds obvious, but it is consistently underinvested in.

The question I use to make it concrete is this: if a consumer types "running shoe that works for beginners with knee injuries" into an AI-powered search or shopping agent, can your business return a good answer? That requires the right product data, structured in a way AI can read, connected to content that addresses that specific question. Most brands cannot do it today. Not because the AI is inadequate, but because the underlying data is fragmented, inconsistently structured, or simply missing the attributes the question requires.

EY's research found that 86% of organisations identify data silos as their top integration challenge. The sequence matters: data governance first, then platform modernisation, then AI deployment. Most companies are attempting that in reverse, deploying AI tools on top of data infrastructure that was not built to support them.

This applies to consumer data and product data equally. Comprehensive, clean, agent-ready data is not a prerequisite you clear once and move on from. It is a continuous investment that compounds in value as AI capability scales around it.

The brands that are winning on AI have not necessarily deployed the most tools. They have built the most usable data foundation. That is the moat.

What the 5% look like in practice

They start from business outcomes, not technology. They orchestrate a small number of high-conviction initiatives rather than spreading effort across dozens of pilots. They redesign the workflow, not just the tool. They think about AI as both an efficiency and a transformation lever. And they treat data readiness as a strategic investment, not a cleanup task.

None of that is complicated. Most of it is not happening.

Before any AI initiative begins, there is one question worth sitting with: if this works technically, what is the specific operating change in this business twelve months from now, and who is accountable for delivering it?

If that question does not have a clear, owned answer, the initiative is not ready to start.

Linda Cereda leads The Convened's AI Transformation practice within Capability Infrastructure. To discuss an AI strategy, prioritisation, or transformation engagement, contact us.