Thought LeadershipRetail & FMCG

AI-Driven Demand Planning in FMCG: Why Most Programmes Fail

Why most FMCG AI demand planning initiatives fail to reach production — and the three factors that separate sustainable capability from expensive pilots.

Cairn Novaris·7 min read·2025

The promise and the reality

AI-driven demand planning is one of the most compelling technology applications in FMCG. The business case is straightforward: better demand forecasts mean less safety stock, fewer stock-outs, less waste and more accurate production planning. In a sector where margin is tight and supply chain complexity is high, the value of even modest forecast accuracy improvements is significant.

And yet the majority of AI demand planning programmes in FMCG do not reach production. They produce encouraging results in controlled trials, generate positive presentations to leadership, and then quietly fail to scale. Understanding why is more useful than celebrating the occasional success.

Three reasons programmes fail

1. The data problem is underestimated

AI demand planning requires clean, timely, granular demand signals. Most FMCG organisations do not have them. Sell-in data is available but reflects distributor ordering patterns, not consumer demand. Sell-out data exists but is incomplete, delayed and inconsistently formatted across retail partners. Internal data is held in systems that were not designed to talk to each other.

The honest answer is that most FMCG AI demand planning programmes are not AI problems. They are data infrastructure problems with an AI application on top. Organisations that invest in the AI before fixing the data foundation are building on sand.

"The data is never as clean as you think it is. And the gap between the data you have and the data you need is almost always larger than the initial assessment suggests."

2. The process change is treated as a technology implementation

Demand planning is not a technical function — it is a human one. Planners have developed judgement, intuition and informal processes over years. An AI system that produces better forecasts statistically will not automatically be adopted by planners who do not understand how it works, do not trust its outputs, or whose performance metrics still reward the behaviours the old system incentivised.

Prosci's ADKAR framework is instructive here. Awareness of why the change is needed, desire to participate in it, knowledge of how to use the new system, ability to demonstrate the required behaviours, and reinforcement to sustain the change — each of these must be explicitly addressed. A technology implementation that skips this work will produce a system that is used defensively or abandoned. The Kotter 8-step model is equally applicable: without a guiding coalition of commercial and supply chain leaders who genuinely believe in the new approach, the change will not hold when it encounters the inevitable resistance.

3. The operating model does not change

AI demand planning changes the role of the planner. It does not eliminate the need for human judgement — it elevates the quality of decision it requires. Planners move from number generation to insight interpretation. The skills required are different. The interface between commercial planning and supply planning must be redesigned. The S&OP process must be adapted to make use of the new information the system provides.

Organisations that implement AI demand planning without redesigning the operating model around it find that the system produces better forecasts that nobody acts on. The forecast accuracy metric improves. The business outcomes do not.

What sustainable programmes look like

The FMCG organisations that build genuine AI demand planning capability share three characteristics. First, they invest in data infrastructure before they invest in algorithms — acknowledging that the data foundation is the hard problem, not the modelling. Second, they treat adoption as a change management challenge from the outset, applying structured frameworks to build individual and organisational capability rather than assuming the technology will sell itself. Third, they redesign the planning operating model around the new capability — changing how decisions are made, not just the tool used to inform them.

Value Stream Mapping is a useful diagnostic: mapping the end-to-end demand planning process — from demand signal to production plan — typically reveals the steps where quality is lost, delay is introduced and human judgement compensates for system inadequacy. Those are the steps that drive both the data requirements and the process redesign priorities.


This is the kind of challenge our Operations & Performance and Technology practices are built to address. We do not implement demand planning AI in isolation — we work across data infrastructure, process redesign, change management and technology to build the capability that delivers lasting commercial improvement.

Subscribe to Cairn Novaris Insights

New thinking from practitioners — direct to your inbox. No noise.

More from Cairn Novaris
Thought Leadership
Browse all insights
View full library →
How We Work
Illustrative engagements
See our approach →
Get In Touch
Talk to us about this topic
Start a conversation →