Beyond accuracy

AI forecasting often plateaus because organisations optimise for statistical accuracy instead of building a trusted, purposeful, and embedded planning capability. Real value comes from improving decisions, not just accuracy. In this article, we present three fundamentals and six dogmas for successful AI implementation in forecasting.

The accuracy plateau

Companies invest in AI forecasting and hit a ceiling

Machine learning models can process vast quantities of data, detect non-linear patterns, and react to demand shifts faster than any human planner. Companies across industries have invested significantly in AI-driven forecasting over the past five to seven years and many have seen genuine improvements in statistical accuracy.

Accuracy gains flatten after the initial uplift

Yet a pattern has emerged. Organisations refine their models, add new features, and connect more data sources, but the accuracy needle barely moves. Worse, the people who need the forecast most – sales teams, demand planners, supply chain leaders – often do not trust the numbers the AI produces.

Organisations treat AI as a technology project, not a capability shift

Most AI forecasting initiatives underdeliver not because the models are inadequate. They underdeliver because organisations optimise for statistical accuracy without ensuring the forecast is trusted, purposeful, and embedded in the planning process.

To combat this, we approach AI in forecasting as a capability transformation, not a technology project. Based on over a decade of implementing AI forecasting solutions, we have distilled our experience into three fundamentals and six dogmas. Together, they form a practical framework for organisations that want AI in forecasting to deliver real business value, not just better accuracy scores.

Three fundamentals for AI in demand planning

All demand planning solutions – whether with or without AI – should support a transparent, robust, and efficient planning process. We believe this rests on three fundamentals: trust, purpose, and accuracy. These are not technical requirements but design criteria that determine whether an AI forecast will be used.

Trust

Planners override forecasts they do not trust

If planners and sales teams do not trust the AI forecast, they will override it. And at that point, the investment is wasted.

Transparency earns trust, not just accuracy

Organisations earn trust by limiting black box calculations, making the model’s logic explainable, and demonstrating forecast stability. An AI model that produces highly accurate but volatile forecasts will destroy trust faster than a simpler model that delivers consistent, directionally correct numbers.

Forecast that are not robust create distrust

For S&OP forecasts it is important that the numbers are not changing too much from cycle to cycle if the underlying assumption about the plan remain valid. That is why best-fit approaches are bad as the output can change significantly if another model is automatically selected. Over time the demand planners will start overriding it to avoid trying to explain the changes from cycle to cycle, when the underlying sales have not changed.

Purpose

Different forecasts serve different decisions

Not all forecasts serve the same function. A forecast that drives production scheduling has different requirements from one that informs strategic capacity decisions.

Organisations need to clarify what the forecast is for, who will use it, and which decisions it will support before building an AI model. These considerations are key design decisions that determine at what level the forecast should be stored and calculated.

Understand the demand signal before you model it

Purpose also means understanding the demand signal itself. What are the demand elements? How much is base volume versus promotional, seasonal, or project-driven?

An AI model trained on aggregate shipment data without understanding the underlying demand composition will produce forecasts that are statistically adequate but practically misleading.

Accuracy

Bias causes more damage than random error

Accuracy matters, but bias – the systematic tendency to over- or underforecast – is often a bigger problem. Therefore, it is important to evaluate both accuracy and bias when evaluating model performance.

Evaluate models on their ability to signal change

A genuinely useful AI forecast should help the organisation anticipate and act on changes in the market. Organisations should evaluate models not only on point accuracy but on the ability to signal turning points, demand shifts, and emerging patterns. However, a forecast supporting the short-term planning should be allowed to be more reactive whereas a forecast for S&OP should be more robust.

This requires both the right features and the appropriate governance to act on what the model detects.

Six dogmas for implementing AI in forecasting

The three fundamentals define what a good AI forecasting implementation should achieve. The six dogmas define how to get there. And we call them dogmas deliberately – not because they are beyond question, but because we have seen what happens when organisations ignore them.

#1 Ensure the foundation

A successful AI implementation requires a demand planning process, clear governance, and the right capabilities

AI cannot fix a broken planning process

This is the most important dogma, and it has nothing to do with AI. A machine learning model cannot compensate for a broken demand planning process. If there is no clear forecast ownership, no structured review cadence, and no defined roles, AI will simply automate chaos.

AI cannot compensate for a broken demand planning process. If the foundation is not in place, you are automating chaos.

Governance comes before algorithms

Before investing in AI, organisations should ensure they have clear governance: who owns the forecast, how is it reviewed, and what is the escalation path when forecasts and actuals diverge?

Pair data science skills with planning skills

Organisations also need the right capabilities – not just data science skills, but planning skills. The most effective AI forecasting teams are those where data scientists and demand planners work side by side, each bringing expertise the other lacks.

Start with short-term planning to build organisational readiness

We recommend starting with AI in short-term, operational planning. The foundational requirements for demand sensing are more contained: cleaner data, tighter feedback loops, and more straightforward governance.

Building competence and trust at this level creates the organisational readiness to extend AI into more complex planning domains.

#2 Spend time on internal data

Cleanse and engineer what you have before dreaming about exotic external data

Organisations chase external data before cleaning their own

There is enormous appetite for external data in demand forecasting: weather data, social media sentiment, macroeconomic indicators, point-of-sale feeds. Some of these can be valuable. But we consistently see organisations chase external sources before they have properly cleansed and exploited their own internal data.

Internal data drives the vast majority of forecast accuracy

Customer orders, shipment history, pricing records, promotional calendars, lead times – this data drives the vast majority of accuracy. Getting it clean, consistent, and properly structured is not glamorous work, but it delivers the highest return in any AI forecasting project.

Master data management remains the top operational AI use case

Master data management may not be exciting, but it is consistently cited as the number one operational AI use case – and for good reason. Start here. Get the fundamentals right. Then explore external data with a clear hypothesis about what incremental accuracy it might deliver.

#3 Benchmark against simple models

Calculate forecast value add to prove AI earns its place

Many data science teams cannot prove their model beats a simple alternative

We have seen data science teams spend a year or more building sophisticated forecasting models – only to be unable to answer a simple question: how much more accurate is this than a naïve model?

Simple models set a surprisingly high bar

Every AI forecast should be benchmarked against simple alternatives: a seasonal naïve model, a moving average, or even last year’s actuals. In many product categories, simple models are surprisingly hard to beat. If your machine learning model cannot demonstrate meaningful forecast value add over a method that takes two hours to implement, the investment is not justified.

Measurable improvement builds credibility with leadership

Benchmarking also serves a critical communication function. When you can show leadership that the AI model reduces forecast error by a specific, measurable margin, you build the credibility needed to sustain investment and expand the programme.

#4 Build trust and measure AI adoption

Models are only useful if the results are used and trusted by the organisation

Putting your AI model in production is the starting line, not the finish line

An AI model in production is not the finish line but the starting line. The real challenge begins when planners start interacting with the forecast.

Track planner behaviour, not just model accuracy

Measuring AI adoption means tracking override rates, forecast touch rates, and the accuracy of the human-adjusted forecast versus the untouched AI forecast. If planners are systematically overriding the model and making the forecast worse, that is a governance problem, not a model problem.

Co-develop the forecast with planners to drive adoption

Building trust requires regular communication about model performance, transparency about what the model can and cannot do, and giving planners a voice in how the model is improved.

An AI forecast imposed from above will be resisted. An AI forecast co-developed with planners will be adopted.

#5 Build explainable features

Only do feature engineering you can explain to sales in one sentence

Features determine how well your model learns

Feature engineering is where much of the value in AI forecasting is created. The features you feed the model, e.g., price changes, promotional flags, and seasonality indicators, shape what it can learn.

Data scientists get creative, but creativity can erode trust

The temptation is to let data scientists explore freely: interaction terms, polynomial transformations, embeddings, and latent variables. Some of these may improve accuracy on paper. But if you cannot explain to a sales manager in one sentence why the model uses a particular feature, that feature will erode trust rather than build it.

If you cannot explain a feature to sales in one sentence, it does not belong in your model.

Keep the math sophisticated but the inputs intuitive

This is not about dumbing down the math. The underlying algorithms can be as sophisticated as needed. But the inputs must make intuitive sense to the people who rely on the forecast.

A feature like “price relative to the 12-month average” is powerful and explainable. A feature like “third principal component of a customer-product interaction matrix” is not.

#6 Start with gradient boosting models

Universal, robust, and effective across most demand forecasting contexts

Gradient boosting models consistently outperform in demand forecasting

The AI landscape offers a crowded field of model architectures, from deep learning to transformers to ensemble methods. Our experience is clear: gradient boosting models – XGBoost, LightGBM, and their variants – consistently deliver strong results. They handle mixed data types well, train quickly, and degrade gracefully when data is sparse or noisy. This is what we see from both projects and the literature.

Start with what works, then innovate from a strong baseline

This is not an argument against innovation but rather an argument for pragmatism. A well-tuned gradient boosting model will outperform a poorly implemented neural network every time – and it will be far easier to explain, debug, and maintain.

If a more advanced architecture is warranted, you will know – because you have a strong baseline to beat and can easily gather more refined data, e.g., if you are in e-commerce.

From accuracy to decision quality

The six dogmas solve a real problem today

The six dogmas are deliberately practical. They help organisations get value from AI in forecasting today, with the data, processes, and people they actually have. They are a corrective to the hype cycle that promises transformation but delivers pilot projects that never scale.

Forecast accuracy may not be the right north star

But a larger question is emerging behind the accuracy plateau. A forecast that is five percentage points more accurate but does not change any supply chain decision has zero business value. A forecast that is slightly less accurate but arrives faster, at a more granular level, and with clearer confidence intervals may drive fundamentally better decisions.

Decision quality connects forecasting to integrated business planning

The shift from forecast accuracy to decision quality connects directly to integrated business planning, where the forecast is not an end in itself but an input to cross-functional decisions about supply, inventory, capacity, and commercial strategy.

Start with the foundation, then expand what AI can do

That shift is the next frontier. But it depends on getting the foundation right first. Organisations that follow the six dogmas – grounded in trust, purpose, and accuracy – will build AI forecasting capabilities that are not just statistically sound, but also organisationally useful.

They will move from chasing accuracy to improving decisions. And that is where the real value of AI in supply chain planning begins.