AI Demand Sensing vs Classical Forecasting vs Planner Judgment

Benchmark verdict

Classical forecasting wins the baseline, AI wins selected signal-rich races, and planners win only when they add information

There is no universal champion across every SKU, location, channel, horizon, and lifecycle state. Classical forecasting remains the most credible default because it is inexpensive, reproducible, fast to retrain, and capable of expressing seasonality, trend, intermittency, and uncertainty without an enormous feature estate. AI demand sensing becomes compelling when demand is observed at fine frequency, relevant signals arrive early enough to change a decision, interactions are nonlinear, and the business can operate the data and model lifecycle. Planner judgment becomes valuable when a person holds material forward-looking information that neither method received and is required to state and later test that information.

That ordering matters. A team that jumps directly from a spreadsheet consensus number to a complex model cannot tell whether the new system beat simple history, an established statistical method, or merely a weak process. The benchmark should therefore run as a ladder. Begin with seasonal naive or another explicit no-change rule. Add the best classical model selected on rolling history. Add the AI challenger with only information available at forecast creation time. Add planner adjustments in a separate version. Every rung has a measurable incremental contribution, a cost, and a failure mode.

The most durable operating model is not a three-way compromise in which everyone touches every series. It is segmentation. Stable replenishment items may stay almost entirely automated with a classical method. Promotion-heavy, high-volume retail items may justify a machine-learning model using price, calendar, inventory position, and event features. New products may require analogues, scenarios, and accountable commercial input. Very intermittent service parts may need a probability-of-demand and inventory-policy treatment rather than a contest to predict each zero-heavy period. A supply-chain software platform should route the series to the right policy rather than force one fashionable method across the portfolio.

Evidence: m5-results, fva-2025, census-ecommerce-q2-2026

Three contestants

Compare operating systems, not marketing labels

Classical forecasting means a family of transparent time-series and regression approaches rather than one dated algorithm. Seasonal naive, moving averages, exponential smoothing, state-space models, autoregressive models, intermittent-demand methods, and dynamic regression can all be legitimate members. A production service may select or combine them by series. These methods can include causal variables and hierarchical reconciliation; the defining point is not that they are simplistic, but that their structure, training footprint, and behavior are usually easier to inspect than large machine-learning stacks.

AI demand sensing means a governed predictive pipeline that uses more granular, more recent, or more diverse information to update a near-term demand distribution. Candidate technologies include gradient-boosted trees, global neural forecasting models, representation learning, anomaly detection, feature stores, and probabilistic ensembles. Generative AI can help summarize event evidence or explain exceptions, but it should not be confused with the core numeric forecasting model. A chat interface that writes an S&OP narrative is not demand sensing unless a validated predictive layer produces the underlying quantities.

Planner judgment means an identified person or forum adjusts, selects, or contextualizes a system forecast. The judgment can be valuable: sales may know a customer contract will start, marketing may know an embargoed campaign, and product teams may know a launch date has moved. It can also import incentives and noise. A sales target, stretch plan, financial budget, unconstrained demand forecast, and supply-feasible plan answer different questions. Allowing the consensus meeting to blend them into one unexplained number destroys the ability to evaluate any method.

The benchmark must hold the forecasting task constant. Specify item, location, channel, unit, horizon, creation timestamp, required availability, demand definition, and treatment of lost sales and returns. Then compare what each approach knew at that timestamp. If the AI receives finalized promotion data that the classical model and planner did not, the contest measures information access as much as algorithm quality. That may still be a useful business finding, but it should be labeled accurately.

Evidence: m5-results, fva-2025, m5-robustness

Decision first

The forecast is useful only if it arrives before the decision and changes an action

Forecasting teams often compare models at a horizon that is convenient for data science rather than material for operations. A same-day signal cannot alter an overseas purchase order already frozen for twelve weeks. A monthly family forecast may be adequate for capacity but useless for allocating a promoted SKU across stores. Before benchmarking, map each decision: purchase release, production lock, labor scheduling, transport booking, store replenishment, digital availability, markdown, or customer allocation. Record its latency, frozen zone, service objective, and cost asymmetry.

This map separates sensing from reporting. Suppose an online item suddenly accelerates after an influencer mention. A model detects the acceleration within hours, but the replenishment network can only react after a nightly inventory feed and the next truck departure. The relevant forecast timestamp is the last point at which allocation can change, not the earliest point at which a dashboard can draw a steeper line. Measure decision latency end to end: source event, ingestion, feature calculation, inference, review, approval, system write-back, and execution.

The loss function must reflect the decision. Underforecasting a high-margin item during a limited promotion can cause lost sales and customer substitution; overforecasting a perishable product can cause waste and markdown. Aggregate error gives both units equal weight unless the evaluation explicitly introduces economics. That does not mean training directly on a complicated profit objective on day one. It means reporting error alongside stockout minutes, fill rate, waste, expediting, working capital, and a transparent value simulation using observed constraints.

ASCM's SCOR Digital Standard places planning in the broader work of determining requirements, gathering resource information, balancing requirements and resources, and identifying actions to correct gaps. That context prevents a forecast contest from claiming operational impact by itself. A more accurate unconstrained demand forecast only creates value when planning policies, inventory parameters, capacity, and execution can translate it into a better response.

• Name the decision and the last responsible forecast timestamp.
• Measure source-to-action latency, not only model inference time.
• Keep unconstrained demand, commercial target, and supply-feasible plan as separate versioned objects.
• Translate forecast changes through actual replenishment, capacity, allocation, and service policies.
• Reject improvements that arrive too late or cannot alter an operational action.

Evidence: ascm-scor-ds

Non-AI baseline

A tuned classical portfolio is a demanding opponent, not a straw man

The fair baseline is not last month's forecast copied forward, an untuned moving average, or a spreadsheet whose formulas nobody can trace. Start with simple naive rules because they reveal whether sophistication creates any lift. Add seasonal naive where recurrence is meaningful. Then evaluate a compact library of exponential-smoothing, state-space, autoregressive, dynamic-regression, and intermittent-demand candidates. Select by rolling-origin validation using the same horizon and data availability as production. Consider forecast combinations when no model dominates consistently.

Classical methods have practical advantages that rarely appear in AI demos. They can train across thousands of series on modest infrastructure, produce forecasts quickly during recovery, and expose components such as level, trend, seasonality, and residual variation. They tolerate a narrower data contract because they can operate on timestamps and demand history alone. That makes them a resilient fallback when a promotion feed, web event stream, or feature platform fails. Their assumptions can still be wrong, but the blast radius is usually easier to diagnose.

They also support causal features. A dynamic regression can incorporate price, promotions, holidays, weather, or availability while retaining interpretable coefficients and a time-series error structure. The important comparison is therefore not AI versus a history-only baseline in every case. It may be a boosted-tree global model against a well-specified regression with event features, or a neural probabilistic model against a reconciled state-space portfolio. Both must use the same cutoff-safe feature versions.

A classical portfolio needs governance too. Automatic model selection can churn when adjacent windows produce close scores. Structural breaks can make long histories harmful. Forecast intervals can be miscalibrated. Intermittent demand can make percentage metrics unstable. Use minimum history rules, stability preferences, residual diagnostics, capped transformations, fallback models, and explicit treatment of stockout-censored demand. Preserve the selected specification and training window so a planner can reproduce what changed between forecast cycles.

Evidence: hyndman-mase, m5-robustness, m5-uncertainty

Planner baseline

Planner judgment is an information channel that must be measured like one

Planner judgment is most defensible when it introduces dated, attributable information about the future. Examples include a confirmed customer opening, a signed tender, a changed launch, a retailer reset, a planned media burst, a delisting, or a known supply recovery that will release backorders. The adjustment should identify the event, affected series, start and end dates, expected direction, magnitude or scenario range, evidence owner, and expiration. This turns a number typed into a grid into a hypothesis that can be evaluated.

The 2025 International Journal of Forecasting study on forecast value added pooled 147,131 forecasts and actuals from ten organizations and twenty-two business units. It found that adjustments improved bias and accuracy for only just over half of SKUs, with variation across datasets, and that positive adjustments were more likely to worsen performance. That result should not be generalized into a rule that planners are ineffective everywhere. It demonstrates why organizations need their own reason-level evidence rather than a blanket assumption that consensus adds value.

Separate selection from adjustment. A planner might know which approved model applies during a temporary event, choose among scenarios, add an event quantity to a statistical baseline, or replace a forecast entirely. These interventions have different auditability. An additive event layer is often easier to test than an overwritten total because the business can compare whether the base, event, or interaction was wrong. Large changes should require more evidence, while trivial edits that cannot alter a decision should be eliminated to protect attention.

Forecast value added should be computed at each stage: naive, statistical baseline, AI challenger, planner adjustment, consensus, and final operational plan. Use the same actuals definition and frozen forecast snapshot. Report FVA by reason code, direction, planner, horizon, product class, and materiality, with sufficient sample sizes and uncertainty. The purpose is process learning, not a punitive league table. A team may discover that launch overrides help, routine upward nudges hurt, and supply-constrained actuals make some demand comparisons invalid.

Evidence: fva-2025, goodwin-judgment

AI architecture

Real demand sensing is a pipeline of time-safe evidence, not one magic model

A credible demand-sensing architecture begins with a forecast contract and temporal data model. Every observation needs an event time, an availability time, and a revision policy. Point-of-sale units, orders, searches, page views, price, promotions, inventory, returns, fulfillment promises, distribution constraints, calendar events, and external indicators arrive at different speeds and can be corrected later. The feature service must reproduce exactly what was knowable at each historical forecast cutoff; otherwise offline accuracy includes information from the future.

The model layer can be global, local, or hybrid. Global models learn across related series and can borrow strength for sparse items, while local models reduce cross-series interference and are easier to tailor. Tree ensembles are strong on structured interactions, neural models can learn representations and multiple horizons, and probabilistic heads produce quantiles or distributions. A practical ensemble may blend a classical base, an ML residual correction, and an explicit event component. Architecture should follow observed value, not a need to display every AI technique.

The serving layer needs more than predictions. It should return the model and feature versions, cutoff, training window, uncertainty, input freshness, segment, fallback status, and material drivers or counterfactual checks. If a required feed is stale, the service should fall back to a qualified baseline or abstain from an automated update. A planner should see which new information moved the forecast and whether that information is still active, rather than receiving an unexplained number with a confidence badge.

Generative AI can assist around the numeric pipeline by converting approved event documents into proposed structured features, summarizing exceptions, or drafting a review narrative. Those outputs require grounding and review because language models can misread dates, quantities, scopes, or negation. Keep extracted evidence linked to the source, restrict write access, and do not let a generated explanation become proof that the underlying numeric forecast is sound. The AI development task is a controlled product lifecycle, not a model endpoint.

Evidence: nist-ai-rmf, m5-results, m5-blend

Data readiness

Inventory availability, promotions, and actual demand definitions decide whether the benchmark is believable

Sales are not always demand. A store with zero on-hand inventory may record zero sales while customers would have purchased units. An order stream may include cancellations, substitutions, duplicates, fraud, forward buys, or allocations. Shipments reflect supply and fulfillment, not necessarily requested quantity. Before training, define the target by channel and decision. Preserve observed sales, orders, fulfilled units, lost-sales estimates, returns, and backorders as distinct measures so a data scientist cannot silently select the most convenient label.

Availability is one of the most dangerous omitted variables. If a promoted item stocks out, a model may learn that the promotion suppresses demand because recorded sales collapse after inventory reaches zero. Include on-hand, in-stock status, assortment, listing, promise, capacity, and suppression logic where available. Flag periods with censoring rather than inventing precise latent demand. For evaluation, distinguish forecast error on observable demand from the business simulation used to estimate stockout consequences.

Promotion data needs version history. The final promotion calendar often differs from what planners knew six weeks before launch. Training on final mechanics, realized discount, or post-event classification while evaluating an earlier forecast produces leakage. Store planned values as of each cutoff, including campaign status, price, display, media, channel, geography, funding, and cancellation. The same rule applies to weather forecasts, economic releases, marketplace ranking, and supplier capacity: use the vintage that was actually available, not the revised truth published later.

Create a series registry with stable identifiers and effective-dated relationships among product, location, customer, channel, hierarchy, pack, unit, and currency. Capture launches, delistings, substitutions, assortment changes, mergers, and product supersession. Data quality controls should measure missingness, duplication, late arrival, revisions, semantic drift, and impossible values by source. A data-management foundation is often the largest part of demand-sensing work because a sophisticated learner cannot repair an undefined target or a broken historical join.

• Target contract: demand measure, unit, grain, timezone, returns treatment, and censoring policy.
• Temporal contract: event time, availability time, revision timestamp, and forecast cutoff.
• Commercial contract: price, promotion, assortment, lifecycle, and customer-event versions.
• Availability contract: on-hand, in-stock, capacity, promise, and lost-sales flags.
• Identity contract: effective-dated product, location, customer, channel, and hierarchy keys.
• Quality contract: freshness, completeness, duplicates, corrections, and escalation owners.

Evidence: m5-results, m5-applicability

External signals

A timely signal is not automatically a predictive signal, and a predictive signal is not automatically causal

Demand-sensing programs accumulate signals quickly: web traffic, search trends, social mentions, weather, mobility, competitor price, logistics indices, macroeconomic releases, and news embeddings. Each seems intuitively relevant, but intuition is not out-of-sample evidence. A feature can correlate with demand because both respond to the same event, arrive after purchasing has already occurred, or merely encode the target through an operational feedback loop. Evaluate every signal against its realistic publication lag and revision behavior.

The New York Fed's Global Supply Chain Pressure Index is an instructive example. It integrates transportation cost data and manufacturing indicators from seven interconnected economies and is updated monthly. That makes it useful context for scenario planning or certain aggregate models. It does not mean the index predicts weekly demand for a retailer's individual SKU. Granularity, geography, lag, horizon, and mechanism must match. Treat such a series as a candidate feature and report its incremental contribution, not as a decorative claim of macro awareness.

Use ablation tests to remove each signal family and measure what changes. Repeat the test across rolling origins, event and non-event periods, high- and low-volume products, and several horizons. Check whether the feature improves point error but damages bias or interval calibration. Estimate the operational dependency: licensing, API reliability, ingestion delay, transformation, privacy, and incident response. A one-basis-point offline gain can be economically negative if the signal is expensive or fragile.

Causal interpretation requires stronger evidence than forecasting utility. A model may use search volume effectively without proving that searches cause sales. Avoid action recommendations such as raise price or increase promotion based solely on feature importance. For policy changes, controlled experiments, quasi-experimental analysis, or a structural causal design may be required. Demand sensing answers what is likely under the observed and planned conditions; it does not automatically answer what would happen if the business changed those conditions.

Evidence: nyfed-gscpi, nyfed-gscpi-method

Benchmark design

Rolling origins, frozen vintages, and untouched holdouts prevent an expensive false victory

Random train-test splits are invalid for most demand forecasting because they allow the model to learn from periods later than those it predicts. Use rolling-origin evaluation: train through a historical cutoff, predict the exact operational horizon, advance the cutoff, and repeat. Preserve gaps when feature construction or operational latency requires them. Hold out a final recent period for confirmation after model and policy choices are frozen. If the system is tuned repeatedly against that final period, it is no longer a holdout.

Separate related entities to test generalization. A new-SKU benchmark should hold out products or product families, not random days from products seen during training. A new-store benchmark should withhold locations. Promotion evaluation should ensure that copies of the same campaign do not span training and testing without an explicit reason. Deduplicate near-identical series and features. Otherwise a global model may appear intelligent because it memorizes the sibling of the item being scored.

The M5 competition supplied a valuable large-scale retail test, yet follow-on research found that top global bottom-up approaches lacked robustness across different time periods in those data. That is a warning against declaring a permanent winner from one window. Evaluate several demand regimes, seasonal peaks, supply disturbances, price environments, and lifecycle phases. Report distribution, not just the mean: median series lift, volume-weighted lift, tail degradation, and the share of series for which each approach wins materially.

Run the human stage prospectively when possible. Historical overrides may not preserve what the planner knew, why they changed the number, or which version was truly approved. In a shadow period, freeze the baseline, AI forecast, and planner-adjusted forecast before actuals arrive. Do not let later order updates overwrite them. Adjudicate data incidents separately from model errors. This design costs time, but it protects the company from scaling an improvement produced by leakage, revision, or selective memory.

• Use multiple rolling forecast origins and the actual business horizon.
• Reconstruct feature vintages as of each cutoff, including late and revised sources.
• Hold out products, locations, or event families when testing cold-start claims.
• Freeze one untouched confirmation period after model and threshold selection.
• Report performance distributions and failure segments, not one aggregate leaderboard number.
• Collect planner adjustments prospectively with reason codes and evidence.

Evidence: m5-robustness, m5-results, m5-blend

Metric stack

WAPE, bias, FVA, calibration, and service answer different questions

WAPE divides the sum of absolute errors by the sum of actual demand across the scored set. It is easy to communicate and weights high-volume errors heavily, which is useful for network and portfolio views. It can hide poor performance on low-volume strategic items and becomes undefined when the denominator is zero. Compute it on explicitly defined groups and periods, never by averaging a collection of precomputed WAPE percentages with different denominators. Show numerator and denominator alongside the percentage.

Signed bias reveals directional error. Define the sign convention in the dashboard because organizations use both actual minus forecast and forecast minus actual. Report total bias in units and normalized bias by segment. A model with acceptable WAPE and persistent positive forecast bias can still create excess inventory. A model with negative bias may repeatedly expose service. Aggregate bias can cancel, so include product, location, horizon, event, and lifecycle views plus the percentage of series materially biased.

Forecast value added compares each process stage with the prior stage or a naive reference. If planner adjustments reduce error relative to the AI forecast, the planner stage adds value for that sample. If the final consensus performs worse than the unadjusted statistical model, the meeting destroys value even if the final WAPE looks acceptable in isolation. FVA must state the error measure, aggregation, baseline, and materiality. Small changes within noise should not be celebrated as process improvement.

Probabilistic calibration asks whether outcomes fall below predicted quantiles at the promised rates and whether intervals are sharp enough to be useful. Service metrics ask whether inventory and execution decisions improved. Stockout rate alone is insufficient because a company can eliminate stockouts by holding excessive inventory. Use fill rate, on-shelf availability, lost-sales estimate, inventory days, waste, markdown, expediting, and working-capital measures. Simulate policy effects first, then verify through a controlled operational rollout where feasible.

Evidence: hyndman-mase, fva-2025, m5-uncertainty

Hierarchy and intermittency

A forecast can be accurate at item level and incoherent everywhere the business manages it

Retail and consumer-goods demand exists in hierarchies: units roll into product families, stores into regions, channels into markets, and days into weeks or months. Independent forecasts at every node will not necessarily add up. A store-item model may imply one regional total while the regional model implies another. Reconciliation methods can make forecasts coherent, but coherence alone does not make them accurate. Choose the hierarchy and reconciliation policy according to the decisions made at each level, then test both bottom and aggregate performance.

The M5 benchmark made hierarchy central by scoring 42,840 related series, including 30,490 at the lowest item-store level. Top solutions often used global bottom-up approaches, but the follow-on robustness work cautions that time-period choice affected rankings. For a company, the lesson is not to copy a competition architecture. It is to evaluate cross-series learning, aggregation, and reconciliation over several operational windows and to preserve a simple coherent fallback when the global model fails.

Intermittent demand needs separate treatment. When many periods are zero, MAPE can be undefined or misleading and WAPE can be dominated by a few nonzero events. The M5 uncertainty research describes overdispersion and sporadic demand, especially zeros, and argues for distributional treatment. For service parts or slow movers, predict probability of occurrence and size conditional on occurrence, or model lead-time demand for inventory decisions. A point estimate of zero every week can appear frequently correct while failing exactly when stock is needed.

Cold starts also break ordinary scorecards. A new item has no own history, and a new location may not share the same assortment or customer behavior as established sites. Use product attributes, analogues, hierarchy priors, launch curves, and scenarios, but test on historically held-out launches rather than reconstructing them with information learned later. Require planners to document analogue choice and event assumptions. AI may borrow strength better than a local model, while classical and human scenario methods can be safer when the product is genuinely outside previous experience.

Evidence: m5-results, m5-robustness, m5-uncertainty, m5-applicability

Failure-mode audit

AI fails through leakage and drift; classical models fail through stale structure; planners fail through incentives and anchors

AI demand sensing can fail spectacularly while offline metrics look excellent. Feature leakage may include final promotion outcomes, future inventory, revised weather, post-period price, fulfillment status, or target-derived aggregates. A model can learn that low sales follow low inventory without understanding that inventory constrained sales. Global models may transfer a pattern between unrelated products. Real-time features may go stale silently. Retraining can reinforce a bad feedback loop when forecasts influence availability and availability shapes the next labels.

Classical forecasts fail more visibly but not harmlessly. Fixed seasonality can persist after a channel shift, a long history can dilute a structural break, and automatic selection can prefer a model that won narrowly in one window. Intermittent-demand assumptions may not fit a changing customer base. A transparent model can still be confidently wrong if nobody monitors residuals, bias, or parameter changes. Simplicity reduces some risks; it does not eliminate the need for ownership and review.

Planner failures often reflect the organization. Sales teams may protect upside, finance may pull forecasts toward budget, operations may buffer against shortages, and product teams may defend a launch narrative. Anchoring on the system forecast can make the adjustment look independent when it is not. Group consensus can hide dissent and produce an untestable number. Frequent small edits may create activity without value, while large changes may escape scrutiny because seniority substitutes for evidence.

Hybrid failures are especially dangerous. A planner can accept an AI recommendation because it is polished, then the approval is written back as a label that makes the next model more certain. Alternatively, distrust can cause blanket overrides that remove real lift. A dashboard can show better WAPE while a high-margin tail deteriorates. Build a failure taxonomy and tag every incident: target error, data latency, leakage, model miss, distribution shift, hierarchy error, planner bias, policy mismatch, execution constraint, or measurement defect. Correct the system that produced the error, not merely the number.

• Leakage: the historical feature was not available at the forecast cutoff.
• Censoring: recorded sales fell because inventory or capacity prevented demand from being observed.
• Drift: product, customer, channel, event, or feature relationships changed after training.
• Feedback: forecasts changed availability or promotion, then those outcomes became future labels.
• Automation bias: a reviewer accepted a model result without evaluating material evidence.
• Incentive bias: an override moved demand toward a target, buffer, or negotiation position.
• Metric masking: aggregate gains concealed severe degradation in a consequential segment.

Evidence: nist-ai-rmf, nist-adversarial-ml, fva-2025, m5-robustness

Human oversight

Planners should govern exceptions and scenarios, not repaint every model output

Human oversight needs a precise job design. The planner should verify that material events are represented, challenge a forecast when new evidence exists, select a documented scenario under uncertainty, and escalate cases outside model coverage. The planner should not be required to click approval on thousands of routine series. Universal review creates automation theater: attention is spread too thin, acceptance becomes habitual, and accountability is claimed without meaningful inspection.

Route exceptions using both consequence and uncertainty. High-volume or high-margin items with unusual residuals deserve attention even when model confidence is high. Novel launches, large price changes, major campaigns, delistings, and contract events require evidence review. A stale input, disagreement between models, unusually wide interval, or out-of-distribution score should trigger investigation. Low-consequence stable items can flow under automated policy with sampled quality checks and a robust fallback.

Design the interface to reduce anchoring. For selected independent-review samples, show facts and event evidence before displaying the AI point forecast. Present the classical baseline, AI distribution, prior forecast, actual-to-date, and material drivers without visually declaring one answer correct. Require a reason and an affected time window for material overrides. Expire temporary adjustments automatically unless renewed with evidence. Allow planners to create alternative scenarios instead of overwriting the central estimate when uncertainty is genuinely unresolved.

Oversight must survive staffing and outages. Maintain the classical baseline and the ability to run a forecast cycle if the AI service or a key signal is unavailable. Train planners in error metrics, uncertainty, leakage, and model limitations, not only interface navigation. Review overrides and misses in a blameless monthly learning loop. NIST's AI RMF emphasizes documented human oversight, ongoing measurement, roles, and lifecycle risk management; those outcomes fit a demand-planning product even though the framework is voluntary and use-case agnostic.

• Auto-flow: stable, in-distribution series with fresh data, small consequence, and no material event.
• Planner review: event evidence, model disagreement, unusual bias, or high decision consequence.
• Specialist review: launch, structural break, exceptional customer commitment, or policy conflict.
• Scenario forum: unresolved uncertainty that should be expressed as alternatives rather than hidden in one number.
• Fallback mode: classical forecasts and predefined policies when features, models, or approvals are unavailable.

Evidence: nist-ai-rmf, fva-2025, goodwin-judgment

Security and governance

Prices, promotions, customer orders, and launch plans make the forecast platform a sensitive system

Demand-sensing inputs reveal commercial strategy. Future prices, campaign calendars, customer commitments, product launches, channel performance, inventory shortages, and supplier constraints can be competitively sensitive or contractually restricted. Customer- or location-level data may also create privacy obligations depending on the content and jurisdiction. Classify the data before architecture decisions. Define approved purposes, fields, regions, retention, sharing, training use, and deletion for every source and derived feature.

Use least-privilege access and separation of duties. A marketing user may propose an event but should not rewrite historical actuals. A data engineer may repair a pipeline but should not approve a forecast policy. A model service should read only the sources and features it needs, and downstream planning systems should accept writes through controlled identities. Encrypt data in transit and at rest, log read and write actions, test recovery, and verify that vendor contracts cover subprocessors, customer-data training, export, incident notice, and termination.

Machine-learning security extends beyond account compromise. NIST's 2025 adversarial-ML taxonomy covers poisoning, evasion, privacy, and misuse risks across AI lifecycles. In demand sensing, a compromised or manipulated signal could distort forecasts, while a poisoned event document could influence a generative extraction tool. Validate source identities, constrain accepted schemas and ranges, quarantine anomalous feeds, preserve raw records, and require review before untrusted text becomes a production event feature.

Govern the model as a named inventory item with owner, intended use, prohibited use, training sources, evaluation record, dependencies, deployment date, change history, monitoring, rollback, and retirement. Map NIST AI RMF's Govern, Map, Measure, and Manage functions into existing risk practice rather than treating them as a one-time checklist. The forecast owner remains accountable for the production policy even when a vendor supplies a model or managed service.

Evidence: nist-ai-rmf, nist-adversarial-ml, nist-csf-2

Cost benchmark

Total cost includes the feature estate, planner minutes, recovery design, and cost of being wrong

Classical forecasting usually has the lowest technical cost. A compact model library can cover a large portfolio with modest compute, and experienced analysts can explain or replace it. Its hidden costs are manual data preparation, coarse segmentation, delayed response to events, and the business impact of systematic misses. Measure those rather than assuming low software cost equals low planning cost.

AI demand sensing adds feature ingestion, stream or batch infrastructure, historical backfill, point-in-time storage, training compute, experiment tracking, model registry, real-time or scheduled serving, monitoring, security, on-call support, and specialist talent. External signals may require licenses. Global models can reduce marginal cost per series at scale, yet each new feature dependency raises operational complexity. Cost the champion and fallback together because resilience is part of production, not an optional add-on.

Planner judgment appears inexpensive when salaries and meeting time are already budgeted, but recurring review grows with SKU-location volume. Count preparation, meetings, exception research, follow-up, changes, and downstream reconciliation. Also count the opportunity cost: skilled planners inspecting harmless one-percent deviations are not resolving launches, constraints, or supplier risks. Automating routine flow can create value even when forecast accuracy changes little, provided quality does not decline.

Build a three-year scenario rather than a single ROI percentage. Separate fixed implementation, recurring platform, data licensing, compute, operations, model risk, planner labor, and change-management costs. Estimate benefit ranges from validated forecast lift passed through actual policies, with conservative attribution and sensitivity to inventory and service assumptions. Include shutdown and portability: can the company export features, forecasts, decisions, and evaluation history, and can it return to the classical baseline without interrupting replenishment?

Evidence: nist-ai-rmf, ascm-scor-ds

Implementation playbook

A twelve-step rollout moves from decision contract to controlled champion

A safe program begins with one decision, one coherent portfolio, and enough history to observe several relevant regimes. Do not start by connecting every enterprise feed or promising autonomous S&OP. Choose a segment where forecast timing matters, actuals can be defined, operations can act, and both success and failure are measurable. Preserve the incumbent process during shadow operation so the company can compare without disrupting service.

The steps below form a decision gate rather than a fixed calendar. A data-readiness failure should pause modeling. An AI challenger that does not beat the tuned classical baseline should not be disguised as a platform success; the company may still retain the improved data and classical service. A statistically promising model that cannot meet latency, reliability, security, or cost gates should remain a research artifact. Expansion is earned by observed production outcomes in the initial segment.

1. 1. Contract the decision

   Name the operational action, owner, grain, unit, horizon, cutoff, frozen zone, service policy, constraint, and economic asymmetry. Separate demand forecast from target and feasible plan.

2. 2. Freeze actuals and forecast versions

   Define observed demand, sales, orders, returns, cancellations, lost-sales treatment, and revision timing. Store every forecast stage with creation timestamp and prevent later overwrites.

3. 3. Map signals and availability

   Inventory each source, event time, availability time, revision behavior, ownership, sensitivity, cost, freshness expectation, and historical coverage. Remove features that cannot be reconstructed at past cutoffs.

4. 4. Build naive and classical baselines

   Implement seasonal-naive and a tuned compact classical portfolio with rolling-origin selection, hierarchy treatment, intervals, diagnostics, and fallback behavior.

5. 5. Define the benchmark scorecard

   Pre-register WAPE or chosen error metrics, sign convention for bias, quantile measures, FVA stages, service simulation, segments, latency, reliability, and cost gates.

6. 6. Train bounded AI challengers

   Start with a small number of architectures and signal families. Use point-in-time features, realistic horizons, reproducible pipelines, and a model card describing intended and prohibited use.

7. 7. Test robustness and ablation

   Evaluate multiple rolling windows, product and location holdouts, peak and ordinary periods, signal removal, stale-feed scenarios, and material failure segments.

8. 8. Design planner intervention

   Create exception rules, reason codes, evidence fields, materiality thresholds, scenario handling, expiration, and independent samples. Train planners on limitations and metrics.

9. 9. Complete security and governance gates

   Approve data purpose and access, vendor terms, logging, retention, recovery, adversarial-feed controls, model inventory, ownership, incident response, rollback, and retirement.

10. 10. Run a prospective shadow cycle

    Produce classical, AI, and planner versions on the real schedule without changing execution. Measure source-to-action latency, failures, overrides, and final outcomes.

11. 11. Pilot with controlled execution

    Apply the winning policy to a bounded group with a valid comparison, explicit inventory and replenishment parameters, daily incident review, and predefined stop conditions.

12. 12. Promote by segment and keep a challenger

    Deploy only where evidence passes. Retain classical fallback, monitor drift and FVA, refresh the holdout, and require a new gate before adding signals, autonomy, or portfolio scope.

Evidence: nist-ai-rmf, m5-robustness, fva-2025, nist-csf-2

Segmentation rules

Route each demand pattern to a policy instead of forcing every series into one leaderboard

High-volume, promotion-sensitive items with timely price, display, media, availability, and transaction data are the clearest AI candidates. Nonlinear interactions and cross-series learning can matter, and operational value can justify the feature estate. Still require a classical baseline and event ablation. If most lift disappears when realized promotion fields are removed, the offline result may not survive real cutoff timing.

Stable seasonal items often belong with classical automation. Exponential smoothing, state-space, or regression methods can be accurate, fast, transparent, and inexpensive. AI should be a challenger, not a mandatory destination. Planner intervention should be rare and event-based. This segment can become the fallback foundation that protects the wider planning cycle during a feature or model incident.

New products and discontinuities need scenarios and accountability. An AI model can use attributes and analogues, but uncertainty should remain visible. Planners should declare the analogue, commercial event, distribution build, cannibalization assumption, and confidence range. Evaluate launch forecasts on prior launches without retrospective information. After enough actuals accrue, transition the item into a regular policy rather than preserving special treatment indefinitely.

Slow and intermittent items should be tied to inventory decisions. A classical intermittent-demand method, a probabilistic AI model, or a hierarchical pool may each work, but point-error rankings can be misleading. Evaluate lead-time demand distribution, service, obsolescence, and holding cost. Strategic spare parts may require engineering criticality and supply-risk judgment even when the expected demand is near zero.

Large contractual or customer events need a structured overlay. A signed order or verified opening may be better represented as a separate event quantity than inferred from time series. The base model estimates ordinary demand; the event layer records committed and probabilistic increments; a planner owns evidence and expiry. This decomposition makes error analysis far more useful than asking a single model to learn rare negotiations.

• Stable seasonal: classical champion, AI challenger, exception-only human review.
• Signal-rich promoted: AI contender with ablation, latency, and leakage controls.
• New product: analogue and attribute model plus planner-owned scenarios and wide uncertainty.
• Intermittent or spare part: probabilistic lead-time demand and inventory-policy evaluation.
• Contractual event: explicit evidence-backed overlay separated from ordinary demand.
• Unreliable data: baseline or abstention until the target and feature history are repaired.

Evidence: m5-results, m5-uncertainty, fva-2025

Broader resilience

Use macro pressure as a scenario input without asking the forecast to replace supply-chain strategy

Demand sensing operates inside an uncertain supply network. Transport disruption, sourcing concentration, economic conditions, and policy changes can alter availability, lead times, prices, and customer behavior. The OECD's 2025 Supply Chain Resilience Review emphasizes agility, adaptability, and alignment and reports that simple relocalization can carry large economic costs without consistently improving resilience. Forecasting can help detect and quantify changes, but it cannot choose the sourcing, inventory, capacity, or commercial strategy by itself.

Keep demand and supply risk distinct even when they interact. A late inbound shipment can reduce recorded sales, making the demand model observe a false decline. A price increase caused by supply pressure can change demand, but the forecast should represent the price scenario rather than silently treating scarcity as preference. The S&OP or integrated business planning process should reconcile unconstrained demand, constrained supply, financial objectives, and risk scenarios as separate inputs with named owners.

External indices can trigger review or scenario ranges. For example, a material movement in a logistics or manufacturing pressure indicator may justify testing lead-time, price, or demand sensitivities. It does not justify adding an arbitrary uplift to every SKU. Specify the mechanism, affected segment, time lag, and decision. Monitor whether the scenario led to better preparation. The goal is an adaptable planning system, not a model that claims to internalize the entire world economy.

This is where a well-designed custom planning workflow creates leverage. The forecast service publishes distributions and material changes; the risk process contributes scenarios; inventory and capacity models translate them into actions; planners record decisions and evidence; execution feeds back observed results. Each object remains reviewable. Integration prevents forecast accuracy from becoming an isolated data-science metric while preserving the boundaries needed for accountability.

Evidence: oecd-resilience-2025, nyfed-gscpi, ascm-scor-ds

Final decision rule

Buy information advantage, preserve a transparent baseline, and make every override prove value

Choose AI demand sensing when the portfolio has a concrete short-horizon decision, granular and point-in-time data, plausible leading signals, enough scale to absorb the production cost, and a prospective benchmark showing material lift over tuned classical methods. Do not choose it because a vendor promises real-time intelligence or because a language model can narrate a chart. Require performance by segment, rolling-window robustness, ablation, uncertainty calibration, fallback behavior, source-to-action latency, and fully loaded cost.

Choose classical forecasting as the primary method when histories are stable, data beyond demand are limited, interpretability and resilience matter, or AI lift is small and fragile. Keep it everywhere as a benchmark and operational fallback. A simple method that is available, reproducible, and aligned with the decision can create more value than a sophisticated method that arrives late or depends on a broken signal.

Choose planner judgment when material information about the future is genuinely outside the models, the person can document evidence and scope, and the adjustment can be frozen and evaluated. Do not require human touches on every forecast. Route consequence and uncertainty to experts, expire event overlays, preserve independent samples, and review forecast value added by reason. The right human contribution is selective context and responsibility, not routine numerical decoration.

The winning system is therefore a controlled portfolio. It starts with an honest baseline, adds AI where information and nonlinearity create repeatable lift, and invokes judgment where the future is known through business evidence rather than historical patterns. It measures WAPE and bias but does not stop there; it follows forecasts into stockouts, inventory, service, latency, cost, and governance. That is how demand sensing becomes a supply-chain capability rather than an expensive synonym for forecasting.

Evidence: m5-results, m5-robustness, fva-2025, nist-ai-rmf, ascm-scor-ds