TL;DR
AI in demand forecasting uses machine learning to predict what customers will buy, learning from sales history plus the drivers that actually move demand: price, promotion, weather, holidays and events. Gartner expects 70% of large organizations to adopt AI-based supply chain forecasting by 2030, yet only 23% of supply chain organizations report having a formal AI strategy today. Gradient boosting remains the workhorse for most SKU-level forecasting, while Temporal Fusion Transformers earn their cost only on rich multivariate retail data. Time-series foundation models such as TimesFM and Chronos now produce useful zero-shot forecasts for products with little or no sales history. Judge accuracy on WMAPE, bias and forecast value added against a naive baseline, never on MAPE alone. Start with one category, run the model in shadow mode for a quarter, and automate release only where planner override rates stay low.
Key Takeaways The adoption gap is the real story. Gartner predicts 70% of large organizations will adopt AI-based supply chain forecasting by 2030, while just 23% of supply chain organizations report a formal AI strategy.Forecasting, planning and sensing are three different jobs with different horizons, inputs, owners and failure modes. Buying one to fix another is a common and expensive mistake.Gradient boosting is the workhorse for SKU-level demand forecasting. Transformers are not an automatic win over boosting or a well-tuned statistical baseline.Foundation models changed the cold-start problem. TimesFM and Chronos forecast products with no sales history zero-shot, the case Gartner singles out as requiring AI.Measure with WMAPE, bias and forecast value added. MAPE alone breaks on low-volume and intermittent SKUs, and a forecast never compared to a naive baseline is an unmeasured forecast.McKinsey reports errors falling 20 to 50 percent with AI-driven forecasting, lost sales from unavailability down by up to 65 percent, and inventory reductions of 20 to 30 percent.Amazon Forecast is closed to new customers. Evaluate SAP IBP, o9, Blue Yonder, Kinaxis, RELEX or Anaplan, or build on Azure Machine Learning or Vertex AI.Watch on YouTube
Optimizing Shipping Costs with AI: Kanerika’s Real-World Success Story
How a logistics operation put AI to work on its own shipment data to cut cost, running on the same data foundation a demand forecast needs.
Two Gartner Numbers That Do Not Line Up In September 2025, Gartner predicted that seventy percent of large-scale organizations will adopt AI-based forecasting to predict future demand by 2030 . Three months earlier, the same firm surveyed 120 supply chain leaders and found that just 23% report having a formal supply chain AI strategy .
Both numbers describe the same decade. One says AI forecasting is about to become the default. The other says most of the teams who will run it have not yet written down how.
That gap, not the algorithm, is where forecasting programs stall. The modeling is largely solved. The hard parts are the decision the forecast is supposed to change, the demand that never got recorded, and the planner who has to trust the number enough to stop overriding it.
This guide is organized around that gap: which model to pick and when, how to tell whether it is working, what the platform market actually offers, and the sequence that moves a forecast into next month’s replenishment run.
What Is AI in Demand Forecasting? AI in demand forecasting is the use of machine learning to estimate future demand for a product, at a location, over a defined horizon. The model learns from sales history and from the external drivers that move demand, rather than extrapolating history on its own.
The structural difference from classical demand forecasting is where the learning happens. A statistical method fits one model per time series. A machine learning method fits one model across thousands of series at once, so behavior learned on established products transfers to new ones.
What AI Actually Changes Scale. One pooled model covers tens of thousands of SKU-location pairs, instead of one hand-tuned model per series that nobody has time to tune.Drivers. Price, promotion, weather, holidays and competitor moves enter as model features, instead of arriving as manual overrides applied after the forecast is already wrong.Cadence. Retraining can run weekly or daily, so the forecast reflects last week’s demand rather than last quarter’s.Cold start. A new item can inherit the demand shape of similar items. Gartner singles this out, noting AI is required for automated predictions on new product introductions and promotions with limited or no historical data.Granularity. Forecasts can run at SKU-store level, where a per-series statistical engine runs out of signal entirely.What AI Does Not Change It cannot see demand you never recorded. Sales censored by a stockout look like low demand, so the model learns the stockout and repeats it.It cannot repair a broken product hierarchy or an inconsistent unit of measure. Those are data governance problems, and they resurface later disguised as forecast error.It does not settle an S&OP disagreement. The consensus number is still negotiated between sales, supply and finance, and the model has no vote in that room.It does not remove the need for a baseline. A forecast nobody compared against a naive rule is an unmeasured forecast, whatever its architecture.
For the wider picture of where forecasting sits among other operational use cases, see our overview of AI in supply chain . If your interest runs outside operations, our guide to AI forecasting tools across finance, HR and marketing covers the cross-domain toolset.
Demand Forecasting vs Demand Planning vs Demand Sensing Three terms get used interchangeably in vendor material, and they are three different jobs. Planning asks what capacity and budget to hold. Forecasting asks how much to buy. Sensing asks where to move the stock you already own.
Buying a demand sensing module to fix a demand planning problem is a common and expensive mistake, and it usually survives a year before anyone names it.
The table below adds the detail the diagram leaves out, including the failure mode each approach produces when it is pointed at the wrong question. Read the last row first if you are diagnosing an existing process rather than choosing a new one.
Statistical demand planning vs ML demand forecasting vs demand sensing
Statistical Demand Planning ML Demand Forecasting Demand Sensing Typical horizon 3 to 18 months, monthly buckets 1 to 12 months, weekly or daily buckets 0 to 6 weeks, daily Primary inputs Shipment and order history History plus price, promotion, weather, holidays and events POS, channel inventory, open orders and other short-term signals Typical methods Exponential smoothing, ARIMA, Croston and TSB Gradient boosting, LSTM, Temporal Fusion Transformers, foundation models Pattern detection on near-real-time feeds, usually layered over a statistical base Granularity Product family, region, channel SKU-location SKU-store and SKU-distribution center Who owns it Demand planner Demand planner with data science support Supply chain operations and replenishment Decision it changes Budget, capacity, long-lead procurement Replenishment quantity, buy plan, safety stock Allocation, expediting, short-term deployment Refresh cadence Monthly Weekly, sometimes daily Daily or intraday Main failure mode Blind to any driver it was never given Overfits promotions when feature discipline is weak Chases noise when the downstream signal is thin
Where S&OP and Integrated Business Planning Fit Sales and operations planning is the monthly forum where the numbers get agreed, and integrated business planning extends that forum to the financial and supply plans on the same cadence. The forecast is an input to that process, not the process itself.
Keep the machine learning forecast and the final consensus number in two separate fields, always. That one design choice lets you measure forecast value added on the human step: whether the overrides improved the number or degraded it.
Case Study
30% Fewer Stockouts With AI Demand Forecasting
A perishable-foods producer serving domestic and international markets replaced a manual planning cycle with LSTM-based demand forecasting and a supply chain collaboration platform. Stockouts fell 30%, order fulfillment time dropped 20%, and profitability rose 12%.
Read the Case Study →
Why Traditional Forecasting Falls Short The case against statistical-only forecasting is narrower than vendors suggest. Three failures are real, and the rest of the usual list is marketing.
History Alone Cannot See the Driver An exponential smoothing model sees one column: past demand. It cannot know that last March’s spike was a promotion, that a competitor was out of stock, or that a price cut is scheduled for next month.
So it reads the promotion as seasonality and repeats it a year later, when no promotion is planned. Planners then correct that by hand, every cycle, on every affected item, forever.
Manual Cycles Cannot Cover the Catalogue A planner can meaningfully review a few hundred items a month, while a mid-size distributor carries tens of thousands of SKU-location combinations. Most of the catalogue is therefore never really forecast.
Those items run on a default profile and get attention only after causing a stockout or a write-off. Amazon’s published account of its own forecasting history describes the same pressure from the inside: a patchwork of specialized components for seasonality, price elasticity and error distributions that became hard to maintain, until the team consolidated on one unified model.
Where Statistical Models Still Win A well-tuned statistical baseline is not a legacy artifact to be retired. On stable, high-volume, long-history series it is fast, transparent, cheap to run and genuinely hard to beat.
It also wins on intermittent demand, where most weeks are zero. Croston’s method and its TSB variant were built for that pattern, while boosting and neural networks drift toward predicting zero everywhere.
Gartner’s own implementation guidance tells planning leaders to benchmark AI forecasts against simple models. That advice exists because the comparison gets skipped, and because the simple model keeps winning on a meaningful share of the catalogue.
AI and ML Services
From Forecasting Prototype to Production Model
Kanerika builds demand forecasting on your own data: feature engineering, backtesting against your existing baseline, integration into the planning system, and the monitoring that stops accuracy decaying after go-live.
Explore AI/ML Services →
Core AI Techniques Used in Demand Forecasting Four families cover almost everything in production today. They differ less in accuracy than in cost, interpretability and the shape of data they need, which is the comparison that decides the choice.
Gradient Boosting, the Real Workhorse Gradient boosted decision trees, in implementations such as LightGBM and XGBoost, produce the majority of production demand forecasts in the enterprise, and the approach has placed near the top of most large public forecasting competitions of the last decade.
The reasons are practical. Boosting handles mixed data types without heavy preprocessing, tolerates missing values, trains on ordinary CPU hardware, and exposes feature importance a planner can interrogate.
Its one real demand is feature discipline. The model knows about a promotion or a price change only if you build the column, and only if that column is not quietly leaking information from the future. Our primer on ML algorithms covers the mechanics.
RNNs, LSTMs and Temporal Fusion Transformers Recurrent architectures, and long short-term memory networks in particular, handle long sequences and learn across many series at once, which suits deep history with complex interacting seasonality.
Temporal Fusion Transformers are the most credible transformer variant for demand work. They accept known-future inputs such as a planned price or a scheduled holiday, return quantile forecasts instead of a single point, and expose variable importance.
Both carry real infrastructure cost: GPU training, longer tuning cycles and an MLOps practice to keep them current. That cost is justified by rich multivariate data, not by ambition.
Why “Transformers Are State of the Art” Needs Qualifying The claim that transformers are the state of the art for time-series forecasting is a 2021 framing that has not aged cleanly, and it still gets repeated long after the evidence moved on.
Zeng and colleagues, in work presented at AAAI 2023, compared transformer baselines against what they described as embarrassingly simple one-layer linear models named LTSF-Linear. The linear models outperformed the transformers on most long-term forecasting benchmarks: Are Transformers Effective for Time Series Forecasting? .
The practical reading is not that transformers are useless. It is that they are one option among several, that they must be benchmarked against gradient boosting and a tuned statistical baseline, and that the burden of proof sits with the more expensive model.
Time-Series Foundation Models: TimesFM, Chronos and TimeGPT The genuinely new development since 2024 is pre-trained time-series foundation models. These train once on very large corpora of time series, then forecast a series they have never seen with no fitting step at all.
Google Research published TimesFM , a decoder-only foundation model for time-series forecasting. Amazon Science published Chronos , which tokenizes time series and trains language model architectures on them. Nixtla’s TimeGPT follows the same pattern commercially, behind an API.
For demand planning the interesting property is zero-shot behavior. A product launched last week has no history to fit, but a foundation model has already learned what launch curves of that shape tend to do.
Two cautions apply. Zero-shot accuracy is competitive, not automatically superior, so it still needs a backtest against your own baseline. And these models were pre-trained on public corpora, which contain neither your promotion calendar nor your channel mix.
Choosing a Model: A Decision Table The choice is driven by the data you have, not the model you admire. Start from history length, series count, intermittency and horizon, in that order.
Which forecasting model to start with, by the data you actually have
Your situation Start here Why it fits Watch out for Under 12 months of history, or a brand new item A time-series foundation model, zero-shot, or an attribute-based analogue Pre-trained on large corpora, so it returns a usable forecast with no fitting step Validate it on items where you do have history before trusting it on the ones you do not One to three years of history, thousands of SKUs Gradient boosting on a pooled panel Learns across series, absorbs drivers cheaply, runs on CPU hardware Feature leakage from price or promotion columns built with future information Intermittent demand, most weeks at zero Croston’s method or TSB Designed for sparse series, where boosting and neural nets drift toward zero MAPE is undefined on zero-demand periods, so the metric itself misleads you Rich multivariate retail data, long horizon, known-future inputs Temporal Fusion Transformer Accepts planned prices and events, returns quantiles, exposes variable importance GPU cost, tuning time and the MLOps practice needed to keep it current Stable, high-volume, long history A tuned statistical baseline Fast, transparent and genuinely hard to beat on this pattern Assuming a machine learning model will automatically improve on it
How to Measure Forecast Accuracy and What “Good” Looks Like Accuracy is where forecasting programs quietly lose credibility. A team reports an improved number, the planners do not believe it, and nobody can reconstruct how it was calculated. Four measures answer four different questions.
MAPE vs WMAPE vs Bias vs Forecast Value Added MAPE is the mean absolute percentage error. It is the default in most planning tools and it breaks on low-volume SKUs, because dividing by a small actual produces enormous percentages. It is undefined when the actual is zero.WMAPE weights the error by volume, so a 40% miss on a fast mover counts more than a 400% miss on one unit. For a mixed catalogue this is the honest headline number.Bias measures direction rather than size. Errors that cancel out can still hide a forecast running consistently 8% high, which shows up as inventory rather than as error. The target sits near zero.Forecast value added compares each step against the step before it. It is the only measure that tells you whether the model beat the naive rule, and whether the planner overrides beat the model.
Track all four on the same grain and horizon, every cycle. A single aggregate accuracy figure with no stated grain is not a measurement, it is a claim.
Accuracy Changes With Horizon and Aggregation Two levers move any accuracy number without the model improving at all. Aggregation is the first: error falls sharply as you roll SKU-store up to category-region, because opposing errors cancel.
Horizon is the second. Next-week and next-quarter accuracy are different problems, and a forecast optimized for one is usually worse at the other. Quote both the grain and the lag, or the number is not comparable to anything.
Why a Naive Baseline Is the Only Honest Comparison Before any model is scored, build the dumbest possible forecast: last period’s demand, or the same week last year. That is the naive baseline, and every subsequent number is expressed relative to it.
If the machine learning model cannot beat last year’s actuals on a category, that category does not need machine learning. Saying so in month two is far cheaper than discovering it in month nine.
Gartner’s five-part plan for touchless forecasting makes the same point, advising leaders to keep results explainable and benchmark AI forecasts against simple models. It is the least glamorous step in a forecasting program and the one most often skipped.
The AI Demand Forecasting Process, End to End A forecasting program runs through six stages. Skipping the first is the most reliable way to produce a model nobody uses.
1. Define the Decision the Forecast Will Change Name the order, transfer or production run that will be different because of this number. If nobody can name one, the project is an analytics exercise, not a forecasting program.
This also fixes the grain and horizon for you. A replenishment decision needs SKU-location at weekly buckets out to lead time plus review period, and nothing coarser will serve it.
2. Assemble and Reconcile the Data Pull shipment or POS history, the product and location hierarchies, the price and promotion calendar, and a calendar of holidays and known events. Reconciling them takes longer than everything else on this list combined.
Two checks pay for themselves immediately: flag periods where an item was out of stock so censored demand is not learned as real demand, and confirm the unit of measure is consistent across every source.
3. Engineer the Demand Drivers Turn raw signals into features: lagged demand, rolling means, price relative to a reference price, promotion type and depth, days to holiday, and a product attribute vector for cold-start analogues.
The discipline here is temporal integrity. Every feature must be computable at forecast time from data that existed then, or the backtest will look excellent and production will not.
4. Train, Backtest and Compare Use rolling-origin backtesting rather than a single random split. Time series have order, and shuffling destroys the thing you are predicting.
Score the naive baseline, the tuned statistical model and the machine learning candidate on the same folds, grain and horizon. Report WMAPE and bias for each, plus forecast value added between them.
5. Deploy Into the Planning Cycle The forecast has to arrive in the planning system, on the planning calendar, in the field the planner already looks at. A forecast that lives in a dashboard is a forecast that gets ignored.
Write the model output and the consensus number to separate fields, and log every override with a reason code. That log becomes the most valuable dataset the program produces. AI and analytics across the supply chain depends on this integration step more than on the model.
6. Monitor Drift and the Override Rate Two signals matter after go-live. Data drift and accuracy decay tell you the model is stale, while the planner override rate tells you whether anyone trusts it.
A model with good WMAPE and a 70% override rate has delivered nothing. Investigate the overrides before retuning, because planners are usually reacting to a driver the model never received.
Forecasting Products With No Sales History New product introduction is the hardest case in demand planning and the one competitors write about least. There is no history to extrapolate, and the launch decision is often the largest single inventory commitment of the year.
Gartner names this case directly, describing AI as required for automated predictions on new product introductions and promotions with limited or no historical data. Three approaches are practical today.
Attribute-based analogues. Describe the new item by category, pack size, price tier, flavor and channel, then let a pooled model borrow the demand shape of items with similar attribute vectors. This works well when attribute data is clean and badly when it is not.Zero-shot foundation models. TimesFM or Chronos can forecast from a handful of observations once the launch begins, improving quickly as early sales accumulate.Hierarchical borrowing. Forecast at category level, where history exists, then allocate down to the new SKU using a planned share corrected weekly against actuals.
Whichever you choose, treat the first eight weeks as a measurement exercise rather than a forecast. Set an explicit review point, size the initial buy for the downside, and let the reorder absorb the upside.
Guard against one specific trap. A launch forecast never scored against actuals teaches the organization nothing, so the same over-optimism recurs at every launch. Score it, record the miss, and feed the result into the next attribute model.
Talk to Kanerika
Walk Your Forecasting Pipeline With Our Team
Bring your current process and we will map it against these six stages in 30 minutes, then tell you which stage is actually costing you accuracy. No platform pitch until that is clear.
Book a Meeting →
Benefits of AI in Demand Forecasting Benefit claims in this category are usually unsourced. Every figure below carries its source, and anything that could not be traced to a primary publication has been left out.
Forecasting error falls 20 to 50 percent. McKinsey (2022) reports AI-driven forecasting can reduce errors by between 20 and 50 percent in supply chain networks: AI-driven operations forecasting in data-light environments .Lost sales fall by up to 65 percent. The same McKinsey (2022) analysis reports reductions in lost sales and product unavailability of up to 65 percent, the number that matters to a commercial director rather than a planner.Inventory drops 20 to 30 percent. McKinsey’s work on distribution operations reports inventory level reductions of 20 to 30 percent: Harnessing the power of AI in distribution operations .It is the top AI use case retail executives report progress on. IBM notes executives identified demand forecasting (88%) as an area where they were already making strides with AI: keys to implementing AI in retail environments .Planner time moves from producing numbers to judging them. Gartner frames the goal as touchless forecasting that eliminates frequent manual inputs and regular human interventions, which is a capacity argument rather than an accuracy one.Service levels and working capital improve together. McKinsey (2021) documented this pairing across supply chain AI adopters: succeeding in the AI supply chain revolution . The two normally trade against each other.
One caveat belongs beside all of these. Published ranges come from organizations that finished an implementation and had the data to measure it, so treat them as an upper envelope rather than a plan.
On-Demand Webinar
Optimizing Supply Chain With AI and Analytics
A recorded session on applying analytics and AI across demand, inventory and logistics decisions, including where measurable value shows up first and what to instrument before you start.
Watch the Webinar →
AI Demand Forecasting Platforms Compared The market splits into planning suites that include a forecasting engine, and cloud machine learning platforms you build on. Most cross-domain “AI forecasting tool” lists mix the two, which is why they rarely help a demand planning decision.
The comparison below covers demand planning specifically. It is a positioning summary drawn from each vendor’s stated focus, not a benchmark result, so run your own evaluation on your own data before committing.
Demand planning platforms compared, by fit rather than feature count
Platform Best fit Forecasting approach Consider carefully if SAP IBP SAP estates running S&OP and integrated business planning on one stack Statistical and machine learning forecasting built into the planning suite Your master data and demand signals largely live outside SAP o9 Solutions Large enterprises wanting one planning data model across demand, supply and revenue Graph-based data model with configurable machine learning forecasting You need measurable value inside two quarters rather than two years Blue Yonder Retail and CPG with heavy promotional and replenishment complexity Machine learning forecasting with demand sensing and allocation modules Your catalogue is small and stable enough not to need that depth Kinaxis Manufacturing and multi-tier supply chains that live on scenario speed Concurrent planning with machine learning forecasting layered on Demand forecasting, not supply scenario response, is your main gap RELEX Grocery, fresh and high-velocity retail at store level Forecasting tuned for fresh, promotions and store replenishment You are not primarily a retailer Anaplan Finance-led consensus planning and S&OP orchestration Connected planning and modeling platform; statistical forecasting is not its core engine SKU-level machine learning forecasting is your primary requirement Build on Azure Machine Learning Teams with data engineering capacity inside a Microsoft estate AutoML forecasting covering statistical and deep learning methodsYou have no MLOps practice to own retraining, monitoring and rollback Build on Vertex AI Teams on Google Cloud wanting managed training and tuning Managed tabular forecasting with hierarchical and quantile supportThe planner-facing workflow still has to be designed and built by you
Buy Versus Build Buy when the planning process, not the model, is the bottleneck. A suite brings the workflow, the approval chain, the audit trail and the S&OP calendar, and those take longer to build than any forecasting model.
Build when your demand drivers are unusual, when you already run a capable data analytics platform , and when someone owns model performance after launch. Without that last person, a built model decays within a year.
A hybrid is common and sensible: keep the suite for workflow and consensus, and feed it a custom model output as a candidate forecast. The planner experience stays intact while the modeling stays yours.
Data Analytics Services
Run a Vendor-Neutral Forecasting Bake-Off
We evaluate planning suites and build options against your own demand data and your own baseline, so the shortlist is decided by backtest results rather than by a demo.
Talk to Our Data Team →
A Note on Discontinued Services Older guides still recommend Amazon Forecast as the quick option for small and mid-size businesses. AWS documentation now states plainly that Amazon Forecast is no longer available to new customers , so that advice cannot be acted on by the exact audience it targets.
AWS points new users toward time-series forecasting in SageMaker Canvas instead, and existing Amazon Forecast customers can continue using the service. The wider lesson belongs in any platform shortlist: managed AI services get retired, so check current vendor documentation for every tool on your list rather than trusting a blog post, including this one.
Real-World Applications by Industry Demand forecasting looks different in each sector because the driver set and the cost of being wrong are different. Six patterns cover most enterprise use.
Retail and E-commerce Store-level and channel-level forecasting drives replenishment, allocation and markdown timing, with price, promotion depth, competitor activity and weather dominating the driver set.
Amazon’s published history shows the trajectory: textbook time series methods in 2008, then a patchwork of specialized components for seasonality and price elasticity, then consolidation onto one unified model. Demand forecasting is also the AI use case retail executives most often report progress on, at 88% in the IBM study cited above.
Manufacturing Forecast accuracy converts directly into raw material purchasing and production scheduling, where lead times are long and changeovers are expensive. The horizon that matters is lead time plus a review period.
Sensor and machine data widen the picture. Our guide to IoT data in manufacturing covers how equipment and process signals feed the same planning stack.
Case Study
85% Accurate Sales Forecasting for a Manufacturer
A manufacturing client moved from spreadsheet-driven projections to AI-driven forecasting and reached 85% forecast accuracy, with granular product and customer insight and a 50% improvement in identifying customer churn risk.
See the Results →
Consumer Packaged Goods and Food Shelf life turns forecast error into waste rather than inventory, because a week of over-forecast on fresh produce is discarded, not carried.
Promotional volatility compounds it, since a large share of volume moves on promotion and uplift varies by retailer, display and competing activity. Kanerika’s work with a perishable-foods producer cut stockouts 30% while improving fulfillment time by 20%, using LSTM models alongside a supply chain collaboration platform.
Healthcare and Life Sciences Hospitals forecast consumable and pharmaceutical demand against admission patterns, seasonal illness and scheduled procedures. The cost asymmetry is extreme: an over-forecast is expiry, an under-forecast can be a clinical event.
Distributors face the mirror problem across thousands of pharmacy and clinic locations, with regulatory constraints on substitution that remove the usual flexibility when a forecast misses.
Financial Services Demand forecasting appears here as cash and volume forecasting: ATM and branch cash demand, contact center volume, and transaction volumes for capacity planning.
The patterns are strongly calendar-driven, which suits models with explicit known-future inputs. Forecasting outside the supply chain has its own toolset, covered in our guide to AI forecasting tools for finance, HR and marketing .
Energy and Utilities Load forecasting is the oldest continuous application of statistical and machine learning forecasting anywhere in industry, and weather is the dominant driver.
Renewable generation added a second forecasting problem alongside demand, because supply became weather-dependent too. Both run on the same modeling stack and the same short-horizon retraining cadence.
Challenges and Limitations of AI in Demand Forecasting Four obstacles account for most stalled programs. None of them is a modeling problem, which is the point.
The Data You Do Not Have Censored demand is the quiet killer. Every stockout produces a record of low sales the model reads as low demand, so the forecast learns to under-order exactly the items that already sold out.
Promotion history is the second gap. Many organizations know a promotion ran but cannot reconstruct its depth, mechanic or store coverage, which removes the strongest single driver from the feature set.
Explainability and Planner Trust A planner who cannot see why the number moved will override it, and an overridden forecast delivers nothing regardless of its backtest score.
The practical answers are ordinary: expose feature importance for the drivers that changed, show the forecast against the same period last year, and publish forecast value added so planners can see when their own overrides helped.
Integration With Legacy ERP and Master Data Most enterprises run forecasting against an ERP never designed to receive an external forecast weekly, and the integration work is routinely underestimated by a factor of two.
Master data is the deeper issue. Duplicate item codes, inconsistent units of measure and a hierarchy that changed mid-year each break a pooled model silently. Standardized context protocols for agentic AI and AI/ML systems are starting to reduce this integration tax, without removing it.
The Cost of Ownership After Go-Live Models decay. Demand patterns shift, hierarchies change and feature pipelines break, so accuracy falls quietly unless someone is responsible for watching it.
Budget for retraining, monitoring and a named owner from day one. A forecasting program without an owner reverts to spreadsheets within eighteen months, and the reversion is rarely announced.
A Phased Implementation Roadmap Sequence matters more than ambition. The plan below assumes one team, one category and an existing planning process to plug into.
Days 0 to 30: Baseline and Scope Pick one category with reasonable data and a real commercial problem. Name the decision the forecast will change, and agree the grain and horizon with the planner who owns it.
Build the naive baseline and today’s process accuracy on that same grain. Owner: demand planning lead with a data engineer. Exit test: you can state current WMAPE and bias at the agreed grain and horizon.
Days 30 to 90: First Model in Backtest Assemble the driver data, engineer features with strict temporal integrity, and train two candidates: a tuned statistical model and a gradient boosting model.
Score both against the naive baseline using rolling-origin backtesting. Owner: data science with planner review. Exit test: positive forecast value added over the naive baseline on the majority of category volume.
Days 90 to 180: Shadow Mode in the Planning Cycle Run the model alongside the existing process without acting on it. The planner sees both numbers, works as normal, and every override is logged with a reason code.
This phase produces the evidence that decides the program. Owner: demand planning, data science on support. Exit test: the model beats the incumbent on WMAPE and bias across a full quarter, and the override log explains the exceptions.
Beyond 180 Days: Scaling Toward Touchless Forecasting Expand category by category, not all at once. Each new category brings its own drivers, its own data gaps and its own planner to convince.
Use the override rate as the automation gate. Where planners override rarely and those overrides do not improve the number, release the forecast automatically. Where overrides are frequent and helpful, keep the human in the loop and find out what the model is missing.
That gate is what Gartner means by touchless forecasting: removing frequent manual inputs and routine interventions, not removing planners. Its five-part plan covers the vision, the change parameters, the data strategy, the technology roadmap and the adoption journey, in that order.
The Future of AI in Demand Forecasting Three shifts are already visible in production work, and none of them is a better point forecast.
Foundation models become the cold-start default. Zero-shot forecasting on new items is moving from research to the standard first pass for products with thin history.Probabilistic output replaces the single number. Quantile forecasts let inventory policy be set from the distribution rather than from a point estimate plus a safety stock rule of thumb.Agents handle the exception queue. Rather than producing the forecast, AI agents for data analysis triage which forecasts moved, why, and which need a human. Multi-agent generative AI is being applied to exactly this coordination layer.
What will not change is the constraint. A forecast is only as good as the demand signal recorded, the drivers captured, and the willingness of the organization to act on the number without relitigating it monthly.
How Kanerika Builds Demand Forecasting That Reaches Production Most forecasting engagements fail after the model works, not before. Our approach front-loads the data and integration work that normally gets discovered late, when the budget is already spent.
The Engagement Sequence Decision mapping, weeks 1 to 2. We work with your planners to identify which orders and production runs the forecast will change, then fix the grain and horizon to match. This determines everything downstream.Data and signal audit, weeks 2 to 5. Demand history, hierarchies, price and promotion calendars and event data are assembled and reconciled, with stockout periods flagged so censored demand is not learned as real demand.Baseline and candidate models, weeks 5 to 10. We build the naive baseline first, then a tuned statistical model and a machine learning candidate, scored on identical folds with rolling-origin backtesting.Planning system integration, weeks 8 to 14. The forecast is written into the system the planner already uses, with model output and consensus number in separate fields and every override logged.Shadow run and handover, weeks 14 to 26. The model runs beside the existing process for a full quarter while we instrument drift, accuracy and override rate, and train your team to own all three.The Accelerators We Bring FLIP is Kanerika’s workflow automation platform. In forecasting work it carries the unglamorous load: recurring data pipelines, reconciliation steps and scheduled retraining runs that otherwise land on a data engineer every week.
Karl is our AI data analyst. It lets a planner ask why a forecast moved in plain language and get an answer traced back to the underlying transactions, which solves the explainability problem at the interface rather than inside the model.
Our data analytics practice handles the layer underneath, and our AI inventory management work covers what happens once the forecast is trusted enough to drive replenishment automatically.
Case Study
30% Faster Inventory Reconciliation With Karl
A UK building products manufacturer running Microsoft Navision used Karl to interpret coded item ledger entries. Reconciliation ran 30% faster, time-to-insight halved, and 10+ recurring variance patterns were detected automatically across 400+ reports.
Read the Case Study →
What We Have Learned Doing This Three patterns repeat across engagements, and they are worth stating plainly because they contradict how these projects are usually sold.
The data audit takes longer than the modeling. On a typical engagement, reconciling hierarchies and flagging censored demand consumes more calendar time than training every candidate model combined. Plans that budget the reverse start slipping in week three.A statistical baseline survives more often than clients expect. On stable, high-volume categories it frequently wins, and reporting that honestly is precisely what makes the machine learning result credible in the categories where it does win.Override rate predicts adoption better than accuracy. A model with good WMAPE and a high override rate has not been adopted, and the fix is almost always a missing driver rather than a better algorithm.
The handover matters as much as the build. We leave behind the backtest harness, the monitoring dashboards and the override log schema, so your team can rerun the comparison on a new category without calling us.
We also write down what we could not solve. If a category lacks the promotion history to support a driver-based model, that goes in the report as a data gap with a cost attached, rather than being hidden inside an averaged accuracy number.
The same applies to categories where the statistical baseline wins. We recommend leaving those on the baseline and spending the modeling budget where forecast value added is actually positive, which is usually the promotion-heavy and short-life categories rather than the steady movers.
For the FMCG and retail context specifically, our retail and FMCG practice covers the wider transformation, and AI assistance across business functions covers where the same data foundation gets reused elsewhere in the business.
Talk to Kanerika
Bring Us Your Hardest Forecasting Category
Pick the category your planners override most and we will scope a 90-day backtest against your current process, with a clear pass or fail on forecast value added before anything reaches production.
Book a Scoping Call →
Frequently Asked Questions
What is AI in demand forecasting? AI in demand forecasting is the use of machine learning to estimate future demand for a product, at a location, over a defined horizon. The model learns from sales history and from the external drivers that move demand: price, promotion, weather, holidays and events. The structural difference from classical forecasting is where the learning happens. A statistical method fits one model per time series, while a machine learning method fits one model across thousands of series at once, so behavior learned on established products transfers to new ones.
How is AI used in demand forecasting? A pooled model is trained across thousands of SKU-location series at once, instead of one hand-tuned model per series that nobody has time to tune. The drivers enter as engineered features: lagged demand, rolling means, price relative to a reference price, promotion type and depth, days to holiday, and a product attribute vector for cold-start analogues. Every feature has to be computable at forecast time from data that existed then, or the backtest looks excellent and production does not. Retraining runs weekly or daily, so the forecast reflects last week’s demand rather than last quarter’s, and the output is written into the planning system in a field kept separate from the consensus number.
Is AI good at forecasting? It depends on the pattern, and the honest answer is that a well-tuned statistical baseline frequently wins. Machine learning earns its place on promotion-heavy, short-life and driver-rich categories, while stable, high-volume series with long history are fast, transparent and genuinely hard to beat, and intermittent demand usually belongs to Croston’s method or TSB. McKinsey reported in 2022 that AI-driven forecasting can reduce errors by 20 to 50 percent, but those ranges come from organizations that finished an implementation and had the data to measure it, so treat them as an upper envelope rather than a plan. The only way to know which case you are in is to score every candidate against a naive baseline, last period’s demand or the same week last year, on the same folds, grain and horizon. A forecast nobody compared against a naive rule is an unmeasured forecast, whatever its architecture.
What is the best AI for demand forecasting? There is no single best tool, and the first decision is whether you buy a planning suite or build. On the buy side, SAP IBP fits SAP estates running S&OP on one stack, o9 fits enterprises wanting one planning data model across demand, supply and revenue, Blue Yonder fits retail and CPG with heavy promotional complexity, Kinaxis fits manufacturing that lives on scenario speed, RELEX fits grocery and store-level replenishment, and Anaplan fits finance-led consensus planning rather than SKU-level machine learning. On the build side, Azure Machine Learning and Vertex AI both offer managed forecasting, but neither supplies the planner-facing workflow, and someone still has to own retraining, monitoring and rollback. One note for older shortlists: AWS documentation states that Amazon Forecast is no longer available to new customers, so check current vendor documentation for every tool on your list.
How is AI different from traditional forecasting methods? An exponential smoothing model sees one column, past demand, so it cannot know that last March’s spike was a promotion or that a price cut is scheduled for next month. It reads the promotion as seasonality and repeats it a year later, and planners then correct that by hand, every cycle, on every affected item. A machine learning model pools thousands of series and takes price, promotion, weather, holidays and events as explicit features, which is the actual difference rather than raw sophistication. The trade is feature discipline: the model knows about a promotion only if you build the column, and only if that column is not quietly leaking information from the future. Neither approach removes the need for a naive baseline to measure against.
Which industries benefit most from AI demand forecasting? Six patterns cover most enterprise use, and they differ by driver set and by the cost of being wrong. Retail and e-commerce forecast at store and channel level to drive replenishment, allocation and markdown timing, with price, promotion depth, competitor activity and weather dominating the driver set. Consumer packaged goods and food turn forecast error into waste rather than inventory, because a week of over-forecast on fresh produce is discarded, not carried. Healthcare carries an extreme cost asymmetry, where an over-forecast is expiry and an under-forecast can be a clinical event, while manufacturing converts accuracy directly into raw material purchasing and production scheduling over lead time plus a review period. Financial services forecast cash and contact center volume, and energy and utilities run load forecasting, the oldest continuous application of statistical and machine learning forecasting anywhere in industry.
What are the key benefits of using AI in forecasting? Every figure worth quoting here carries a source. McKinsey (2022) reports that AI-driven forecasting can reduce errors by 20 to 50 percent and cut lost sales from product unavailability by up to 65 percent, and McKinsey’s work on distribution operations reports inventory reductions of 20 to 30 percent. IBM notes that retail executives identified demand forecasting, at 88%, as an area where they were already making strides with AI. Gartner frames the capacity benefit separately, as touchless forecasting that eliminates frequent manual inputs and regular human interventions, which moves planner time from producing numbers to judging them. One caveat belongs beside all of these: published ranges come from organizations that finished an implementation and had the data to measure it, so treat them as an upper envelope rather than a plan.
What are the challenges of AI in demand forecasting? Four obstacles account for most stalled programs, and none of them is a modeling problem. Censored demand is the quiet killer: every stockout produces a record of low sales the model reads as low demand, so the forecast learns to under-order exactly the items that already sold out. Promotion history is the second gap, because many organizations know a promotion ran but cannot reconstruct its depth, mechanic or store coverage, which removes the strongest single driver from the feature set. Integration is the third, since most enterprises run forecasting against an ERP never designed to receive an external forecast weekly, and duplicate item codes or an inconsistent unit of measure break a pooled model silently. The fourth is what happens after go-live: a planner who cannot see why the number moved will override it, models decay as patterns shift, and a program without a named owner reverts to spreadsheets within eighteen months.
Will demand planners be replaced by AI? No, and the change that is actually happening is a different one. What Gartner calls touchless forecasting means eliminating frequent manual inputs and routine interventions, not eliminating planners, so planner time moves from producing numbers to judging them and working exceptions. The practical mechanism is the override rate used as an automation gate: where planners override rarely and those overrides do not improve the number, release the forecast automatically, and where overrides are frequent and helpful, keep the human in the loop and find out what the model is missing. That rate is also the best adoption diagnostic available, because a model with good WMAPE and a 70% override rate has delivered nothing, and the fix is almost always a missing driver rather than a better algorithm. The consensus number is still negotiated between sales, supply and finance in the S&OP forum, and the model has no vote in that room.
What is the future of AI in demand forecasting? Three shifts are already visible in production work, and none of them is a better point forecast. Time-series foundation models such as TimesFM and Chronos are becoming the standard first pass for products with thin or no history, which is the cold-start case Gartner singles out as requiring AI. Quantile forecasts are replacing the single number, so inventory policy can be set from the distribution rather than from a point estimate plus a safety stock rule of thumb. Agents are moving into the exception queue, triaging which forecasts moved, why, and which need a human, rather than producing the forecast themselves. The direction of travel is what Gartner means by touchless forecasting, which is the removal of frequent manual inputs and routine interventions, not the removal of planners.
