Walk into any major retailer’s technology review meeting this year and you will find a familiar scene: an AI pilot that performed brilliantly, a business case everyone believes in, however a project that has quietly stopped moving. This is not a new trend, in fact research shows that only around a third of retail and CPG organisations ever get an AI project beyond the pilot phase.

The wider industry research is even less forgiving. RAND found that more than 80% of AI projects fail, roughly twice the rate of conventional IT projects, while MIT’s Project NANDA reported that 95% of Gen AI pilots deliver no measurable return.

Now here is the uncomfortable part; almost none of these projects died because the AI was not good enough.

The models forecast demand accurately, the computer vision reads the shelf correctly, the recommendations are sound. What kills retail AI projects is everything underneath the AI, and the causes of death are remarkably consistent, as SymphonyAI‘s Senior Director of Strategic Business Growth, Jonathan Tye-Walker, explains.

1. Data feeds that break on contact with reality

The quietest killer of AI projects is the data feed. A pilot runs on a carefully prepared extract: one banner, one category, one clean feed. It works. Then the project moves towards production and meets the real world, where a grocery retailer changes its sell-out file format without warning, a distributor renames half its product hierarchy or a promotional calendar arrives in a structure nobody has seen before.

Every one of those changes breaks a custom feed and every broken feed pulls engineers away from progress and into repair.

Data integration that held up beautifully in a demonstration turns out to be brittle under operational load. The project does not fail dramatically, it just spends its budget standing still.

2. Agents multiplying without shared context

A newer, and if anything more dangerous, failure mode is emerging. As agentic AI arrives in retail, agents are multiplying. One monitoring shelf compliance, another optimising promotions, another flagging availability gaps.

Deployed without a shared view of products, stores and promotions, each agent operates on its own version of commercial reality. The result is a field team receiving contradictory instructions, a category manager unsure which recommendation to trust and a governance headache nobody planned for.

The industry has a name problem here. This looks like an AI problem, so organisations respond with more policies and more oversight meetings. It is actually a context problem.

Agents that do not draw on the same commercial foundation cannot agree, no matter how well each one performs individually.

3. The internal build that never ends

The most seductive killer begins with a perfectly reasonable decision; the internal build.

It starts as a targeted project – a forecasting model here, a compliance dashboard there. Eighteen months later it has become a sprawling integration programme with its own roadmap, its own maintenance burden and no clear route to commercial return.

MIT’s research found that internal builds succeed at roughly a third of the rate of partnerships with specialised providers, and retail is particularly exposed.

Product hierarchies, promotional mechanics, space constraints and retailer data relationships are deep domain problems that general-purpose tooling was never designed to understand.

Architecture is the project

Notice what connects all three? None of them are modelling failures and none of them are fixed by a better model.

Brittle feeds are solved by a shared commercial data foundation that removes the need for point-to-point integration. Agent sprawl is solved by giving every AI component the same understanding of a product, a promotion and a store. The custom build trap is solved by starting from a foundation built for retail rather than assembling one from scratch.

The retailers reaching production are not the ones with the most ambitious pilots. They are the ones treating architecture as the project: consistent, commercial context first, orchestration across workflows second and governance built in from day one, so every recommendation can be traced, measured and improved.

Get that sequence right and the AI, ironically, becomes the easy part. Get it wrong and you join the two thirds whose projects die of causes that never involved the AI at all.