The Honest Assessment
Most practitioners I speak to sit in one of two camps: the believers who think AI will eliminate demand planners within five years, and the sceptics who dismiss AI forecasting as a black box that their ERP already does better. Both camps are wrong in ways that are costing their organisations money.
AI meaningfully improves demand planning in exactly three places. Where the hype outruns reality is in the assumption that better algorithms eliminate the need for human judgment. Here is what I have observed building and deploying these systems in the GCC market.
- Non-linear seasonal pattern detection
- Processing external signals at scale
- Anomaly detection before propagation
- Reducing forecast cycle time by 60–80%
- Consistent application of methodology
- Replacing commercial intelligence
- "Zero-touch" autonomous planning
- Black-box models without interpretability
- Generic models on GCC-specific patterns
- Eliminating the need for S&OP governance
Where AI Genuinely Changes the Game
1. Non-Linear Seasonal Pattern Detection
Classical statistical methods — ARIMA, Holt-Winters exponential smoothing, even seasonal decomposition — are fundamentally designed to detect patterns that repeat predictably. They are powerful tools for stable seasonal demand. They struggle with Ramadan, because Ramadan moves 11 days earlier every year — which means the same calendar month has radically different demand profiles depending on whether Ramadan falls in it or not.
ML-based approaches, particularly gradient boosting methods like XGBoost with Islamic calendar features encoded explicitly, handle this significantly better. In the ML forecasting engine I built, incorporating Ramadan day-count as a feature — rather than calendar month — reduced MAPE for affected SKUs from 18.4% to 11.2% over an 18-month validation period. That is a meaningful improvement that statistical methods cannot match without manual seasonal adjustment.
2. Processing External Signals at Scale
A demand planner reviewing 500 SKUs in a planning cycle cannot simultaneously monitor commodity price trends, competitor stockout signals from retailer EPOS, weather forecasts, and social media sentiment for every category. An ML pipeline can monitor all of these continuously and surface the signals that exceed a threshold — essentially acting as a triage layer that directs human attention to the inputs that actually matter.
At ANDS, we integrated commodity price feeds from the Dubai Mercantile Exchange as a leading indicator for certain categories. The model detected a correlation between palm oil futures and demand for a specific product line three weeks before the commercial team had manually identified it. That lead time advantage in a market with 60-day supply lead times is worth real money.
3. Anomaly Detection Before Propagation
The most expensive forecast errors are the ones that persist across multiple planning cycles before anyone identifies them. A systematic upward bias in a new product introduction, or a model that has not updated since a major range rationalisation, can generate purchase orders that compound the error month over month. ML anomaly detection identifies these drift patterns at the SKU level and flags them for human review before the inventory impact materialises.
The combination that actually works: Holt's Double Exponential Smoothing as the statistical engine, with structured commercial intelligence overlay from the S&OP process, with ML anomaly detection running in parallel. The statistical engine provides the baseline. The commercial overlay adds the information that no algorithm can know. The anomaly detection catches the errors that both components miss. That combination consistently outperforms either pure algorithmic or pure intuitive approaches.
Where the Hype Outruns Reality
AI Cannot Replace Commercial Intelligence
The most fundamental limitation of AI forecasting is that it can only learn from historical patterns. It cannot know that a key account is planning to double their order next month because they are opening a new store. It cannot know that a competitor just went out of stock and you are about to see a demand transfer. It cannot know that the sales director just made a commitment to a volume discount that will pull forward three months of demand into one.
This information lives in the heads of people who have relationships with the trade. The planning function's job is to extract it, structure it, and fold it into the baseline before the model locks. No AI system in production today can substitute for that.
Generic Models Fail on GCC-Specific Patterns
The AI forecasting tools sold to GCC businesses are almost universally trained on North American or European retail data. They handle Christmas seasonality elegantly and fail on Ramadan systematically. They understand Black Friday and have no model for White Friday or National Day. The GCC operator who buys an off-the-shelf AI forecasting tool and applies it without GCC-specific feature engineering is getting a worse forecast than a well-calibrated Holt-Winters model, not a better one.
The Planner of the Future Is Not the One Replaced by AI
The supply chain planners I watch being displaced are not the good ones — they are the ones who never developed beyond maintaining a spreadsheet. The planners who are thriving in an AI-augmented environment are the ones who can interrogate a model's assumptions, identify where human judgment should override the algorithm, and communicate both the forecast and its uncertainty to commercial and financial stakeholders.
The future is not AI replacing planners. It is planners who understand AI replacing planners who don't.
See the ML forecasting models live
The ML Demand Forecasting Engine compares XGBoost and Prophet against classical methods on real GCC data — with full methodology and interactive controls.