Recently, I’ve been hearing a familiar claim:
“We’re rolling out AI-based forecasting, so DDMRP is no longer necessary.”
I’m not an AI expert. What I am focused on is how planning assumptions translate into operational reality. Forecasting and execution sit right at the heart of that, so I wanted to step back and examine what is actually changing — and what isn’t.
To do that, I reviewed published research, consulting papers, and large-scale forecasting benchmarks, asking a simple question:
Has AI fundamentally altered the nature of forecasting?
What AI genuinely improves
There’s solid evidence that AI can enhance forecast performance.
Across multiple studies and case examples, AI-driven approaches show meaningful reductions in forecast error — often cited in the 20–50% range under the right conditions. In practical terms, this means forecasts can become less inaccurate, particularly when supported by good data, thoughtful segmentation, and disciplined planning processes.
But how those improvements are described matters.
When organizations talk about “20–30% better forecasts,” they are almost always referring to relative error reduction, not a dramatic jump toward perfect accuracy. A forecast that improves from 70% to roughly 80% accuracy is an improvement — but it is still far from certainty.
And that gap is structural, not technical.
Uncertainty remains embedded in real-world demand: shifting customer behavior, promotions, substitutions, new products, one-off events, supply disruptions, variable lead times, and broader volatility. These factors don’t disappear just because the algorithm improves.
What forecasting research consistently shows
- No single model consistently dominates
- Combined approaches outperform isolated ones
- Gains are incremental, not exponential
- Accuracy plateaus quickly at granular, operational levels
Forecasts are inherently probabilistic.
The longer the horizon, the wider the error band.
The more detail you demand (SKU, location, short-term execution), the more fragile accuracy becomes.
AI changes how we forecast — not the fundamental limits of forecasting itself.
Does that make DDMRP obsolete?
In my view, clearly not.
The critical difference is this:
Forecasting aims to reduce error.
DDMRP assumes error will persist and designs the system to cope with it.
Forecasting — including AI-based forecasting — is about anticipation.
Execution systems must function when anticipation proves imperfect, which it inevitably does.
A forecast can be statistically strong and still generate operational problems:
- Shortages where demand actually materializes
- Excess inventory where it doesn’t
- Constant expediting
- Unstable schedules and priorities
This is why I don’t see DDMRP and AI forecasting as alternatives. They address different layers of the problem.
Decoupling, buffering, and pull-based execution are about absorbing variability and preserving flow, not predicting demand flawlessly.
A useful analogy is investing. AI doesn’t succeed by predicting markets with certainty. When it adds value, it does so through better pattern recognition, faster scenario evaluation, and risk management. The market remains uncertain — and systems are built with that assumption in mind.
A position I’m confident standing behind
- AI can meaningfully improve forecasts, depending on context
- Those gains reduce error; they don’t eliminate uncertainty
- Because uncertainty remains, operations must be designed to handle it
So the most important question isn’t:
“If we use AI forecasting, do we still need DDMRP?”
It’s:
“How does our operating model deal with the forecast error that remains?”
Because that error doesn’t vanish. It shows up as inventory imbalance, service failures, instability, or firefighting — unless the system is explicitly built to absorb it.
How this connects to our work at Kiira
This is exactly where we focus at Kiira with our clients.
Not on choosing between advanced forecasting or demand-driven execution, but on building operating models that remain resilient when reality diverges from the plan — which it always does.
Forecasting and DDMRP are not competing ideas.
They solve different problems.
Used together, they address both sides of uncertainty: estimation and execution.