AI Demand Forecasting in ERP for Manufacturers | Likwid ERP
Discover essential data for AI demand forecasting, accuracy measurements, and tips to avoid inventory bloat. Enhance your efficiency with Likwid ERP!
<h1>Mastering AI Demand Forecasting in ERP: A Guide for Manufacturers</h1>
<h2>AI-powered demand forecasting in ERP: why leaders treat it as a closed-loop system (not a model)</h2>
<p>AI demand forecasting is moving from “nice analytics” to an operational requirement. In a September 2025 press release, <a href="https://www.gartner.com/en/newsroom/press-releases/2025-09-16-gartner-predicts-70-percent-of-large-orgs-will-adopt-ai-based-supply-chain-forecasting-to-predict-future-demand-by-2030" target="_blank" rel="noopener noreferrer">Gartner predicted 70% of large organizations will adopt AI-based supply chain forecasting by 2030</a>. At the same time, ERP roadmaps are shifting from embedded AI features toward agentic workflows that connect demand signals to supply decisions and customer promises.</p>
<p>For manufacturers and distributors, the practical takeaway is simple: forecasting only “works” when it’s embedded in the ERP execution loop. The leaders in 2025–2026 treat it as a closed system:</p>
<ul>
<li><strong>Data</strong> unified across ERP + CRM + supply chain</li>
<li><strong>Forecast</strong> generated with horizon- and SKU-class awareness</li>
<li><strong>Inventory policy</strong> translated into safety stock, reorder points, and order quantities</li>
<li><strong>Execution</strong> through purchasing, production planning, and allocation</li>
<li><strong>Feedback</strong> from what actually happened (including stockouts and overrides)</li>
</ul>
<p>This is the difference between improving a metric on paper and reducing working capital without sacrificing service levels.</p>
<h2>The minimum viable dataset: what you must unify across ERP, CRM, and supply chain</h2>
<p>You don’t need a “perfect data lake” to start, but you do need a coherent demand-and-supply truth inside your ERP. For most manufacturers and distributors, the minimum viable dataset falls into six buckets.</p>
<h3>1) Demand history (what customers wanted vs what you shipped)</h3>
<ul>
<li><strong>Orders</strong>: order date, requested ship date, promised date, cancellations, partials</li>
<li><strong>Shipments/invoices</strong>: what actually shipped and when (often differs from demand)</li>
<li><strong>Returns/credits</strong>: returns reason codes (quality vs commercial) to avoid “double counting” demand</li>
<li><strong>Customer and channel</strong>: customer segment, region, sales channel, contract vs spot</li>
</ul>
<h3>2) Inventory signals (so the model doesn’t learn the wrong lesson)</h3>
<ul>
<li>On-hand, available-to-promise, and backorders by SKU-location</li>
<li>Stockout flags (days out of stock), fill rate, and allocations/rationing events</li>
<li>Substitutions: what was sold instead when the desired item was unavailable</li>
</ul>
<h3>3) Supply and lead-time reality (to convert forecasts into purchase/production)</h3>
<ul>
<li>Supplier lead times (planned vs actual), variability, minimum order quantities (MOQs), pack sizes</li>
<li>Production constraints: capacity, yield/scrap, cycle times, changeovers, calendars</li>
<li>Inbound reliability: ASN/receipts, late deliveries, expedite flags</li>
</ul>
<h3>4) Pricing, promotions, and events (so AI doesn’t treat campaigns as “new normal”)</h3>
<ul>
<li>Price lists and effective dates, discounting and rebates</li>
<li>Promo calendar, bundles, end caps/placements (distribution), marketing pushes</li>
<li>One-time events: big projects, tenders, outages, recalls, competitor disruption</li>
</ul>
<h3>5) CRM pipeline and customer commitments (especially for B2B)</h3>
<ul>
<li>Opportunities and probability-weighted pipeline by product family</li>
<li>Contract volumes, blanket POs, forecasts from key accounts</li>
<li>Customer requested changes and churn/renewal signals</li>
</ul>
<h3>6) Master data discipline (the silent success factor)</h3>
<ul>
<li>SKU hierarchy (SKU → family → category), units of measure, substitutions</li>
<li>Location/warehouse hierarchy and service level targets by class</li>
<li>Lifecycle status: new item, active, superseded, end-of-life</li>
</ul>
<p>If your planning data is split between tools, the loop breaks. Likwid ERP’s positioning—procurement, production planning, inventory, and CRM in one platform—matters because it reduces the number of reconciliation steps where forecasting programs usually stall. If you’re exploring how AI connects to planning execution, see <a href="/ai-native-mrp">AI-Native MRP</a> and our broader perspective on <a href="/blog/transforming-erp-ai-business-solutions">AI in ERP</a>.</p>
<h2>How much history is “enough” (and what to do with new or intermittent items)</h2>
<p>There’s no universal minimum, but most teams get reliable lift when they can provide:</p>
<ul>
<li><strong>18–36 months</strong> of weekly history for seasonal SKUs (more if seasonality is strong)</li>
<li><strong>12–18 months</strong> for stable, high-run-rate items</li>
<li><strong>6–12 months</strong> can work if you have strong causal signals (price/promo/CRM) and clean stockout flags</li>
</ul>
<p>For <strong>new items</strong>, you typically need one (or more) of these playbooks:</p>
<ul>
<li><strong>Analog mapping</strong>: inherit patterns from “like” SKUs (family, price band, channel)</li>
<li><strong>Attribute-based models</strong>: use product attributes and customer segments to infer demand shape</li>
<li><strong>Launch governance</strong>: a structured manual forecast that is time-boxed, audited, and gradually replaced as real sales accumulate</li>
</ul>
<p>For <strong>intermittent demand</strong> (sporadic spare parts, project-based items), the goal isn’t to force a smooth forecast; it’s to combine an intermittent model approach with inventory policy tuned to service level and cost-to-stockout.</p>
<h2>Accuracy in 2025–2026: WAPE + bias, segmented by horizon and SKU class</h2>
<p>Many teams still default to MAPE, but leaders are moving toward weighted and scale-free metrics that behave better across long-tail catalogs. AWS explicitly calls out WAPE as a key evaluation metric for model selection and ensembling in its demand planning documentation: <a href="https://docs.aws.amazon.com/aws-supply-chain/latest/userguide/forecast-algorithims.html" target="_blank" rel="noopener noreferrer">AWS Supply Chain forecast algorithms and metrics</a>.</p>
<h3>MAPE vs WAPE vs MASE vs bias (plain-English differences)</h3>
<ul>
<li><strong>MAPE</strong>: average percentage error. Problem: low-volume items can explode the percentage and distort reality.</li>
<li><strong>WAPE (wMAPE)</strong>: total absolute error divided by total actual demand. Better for large SKU counts because it weights by volume.</li>
<li><strong>MASE</strong>: compares your model error to a naive baseline (e.g., last period). Useful to ensure you’re beating “do nothing.”</li>
<li><strong>Forecast bias</strong>: direction of error (systematic over-forecasting or under-forecasting). This is the KPI that predicts inventory bloat or chronic stockouts.</li>
</ul>
<p>For <strong>distributors</strong> with thousands of low-volume SKUs, WAPE plus bias is usually more actionable than MAPE. For <strong>manufacturers</strong> with capacity and BOM complexity, you still want WAPE/bias, but also a horizon-based view that reflects production lead times and frozen schedules.</p>
<h3>How leaders segment measurement (so the number means something)</h3>
<p>Instead of one blended score, segment accuracy along the dimensions that drive decisions:</p>
<ul>
<li><strong>By horizon</strong>: next week (allocation/replenishment), next month (purchasing/production), next quarter (capacity and S&OP)</li>
<li><strong>By SKU class</strong>: ABC (volume) and XYZ (variability), or margin/service criticality classes</li>
<li><strong>By level</strong>: SKU-location for execution, product family for S&OP, customer segment for commercial alignment</li>
</ul>
<p>This is also how modern suites are evolving. For example, SAP’s 2026 release highlights describe AI-generated explanations behind safety stock adjustments—tying recommendations to demand variability, lead-time fluctuations, and target service levels: <a href="https://news.sap.com/2026/01/sap-business-ai-release-highlights-q4-2025/" target="_blank" rel="noopener noreferrer">SAP Business AI release highlights Q4 2025</a>. Whether you use SAP, Likwid ERP, or another platform, the operational pattern is the same: accuracy must connect to policy.</p>
<h2>A practical playbook to avoid inventory bloat (without killing service levels)</h2>
<p>Inventory bloat is no longer purely accidental; many firms intentionally buffer risk. Netstock reports that <strong>30% of SMBs say more than 30% of their excess stock is strategic</strong> (up from 23% in 2024), which changes what “forecast success” means: you need governance around what risk you’re buying with working capital. (See the cited benchmark on Netstock’s 2025 report page.)</p>
<h3>Step 1: Correct for stockout-censored demand (lost sales)</h3>
<p>If you train on shipments alone, the model learns that demand drops when you stock out—exactly the opposite of reality. Fix this by:</p>
<ul>
<li>Capturing <strong>stockout days</strong> and suppressing those periods for training, or imputing demand using adjacent periods and customer signals</li>
<li>Using <strong>orders/backorders</strong> as the primary demand signal when available</li>
<li>Flagging <strong>allocation/rationing</strong> periods so the model doesn’t interpret constrained shipments as demand softness</li>
</ul>
<h3>Step 2: Control overrides with audit trails and thresholds</h3>
<p>Overrides are necessary (promos, tenders, known churn), but they are also a common path to bloat. Put structure around them:</p>
<ul>
<li>Role-based approval: who can override by SKU class and horizon</li>
<li>Reason codes: promo, one-time order, customer commitment, supply constraint, launch</li>
<li>Guardrails: maximum override % by class; alerts when bias drifts positive for multiple cycles</li>
<li>Post-mortems: track whether overrides improved WAPE <em>and</em> reduced bias</li>
</ul>
<h3>Step 3: Translate forecast uncertainty into safety stock and order policies</h3>
<p>Forecasts are distributions, not single numbers. To prevent over-ordering, convert uncertainty into explicit inventory decisions:</p>
<ul>
<li>Set <strong>service levels by class</strong> (not one-size-fits-all) based on margin, customer promise, and substitution options</li>
<li>Use <strong>lead time variability</strong> and demand variability together (both matter)</li>
<li>Apply <strong>MOQ/pack rounding</strong> intentionally: separate “policy stock” from “supplier constraint stock” so it’s visible</li>
</ul>
<p>Then connect those policies to execution via MRP/DRP. When this loop is embedded, you can keep AI from “chasing noise” while still responding to true demand shifts. For related execution best practices, see <a href="/blog/procure-to-pay-automation-manufacturing-best-practices">procure-to-pay automation in manufacturing</a>.</p>
<h3>Step 4: Detect promo and one-time order contamination</h3>
<ul>
<li>Tag promo weeks and project orders so the model can treat them as causal events, not baseline demand</li>
<li>Separate “base” vs “lift” (especially for distribution with frequent discounting)</li>
<li>Monitor bias after major price changes; price elasticity can look like a demand trend if not modeled</li>
</ul>
<h3>Step 5: Run a closed-loop KPI set (not just forecast accuracy)</h3>
<p>Leadership should track KPIs that reveal whether forecasting is improving the business, not just the score:</p>
<ul>
<li><strong>WAPE by horizon and class</strong> + <strong>bias</strong> (direction matters)</li>
<li><strong>Fill rate / OTIF</strong> and backorders (service reality)</li>
<li><strong>Inventory turns</strong>, days of supply, and excess/obsolete stock</li>
<li><strong>Expedite count</strong>, premium freight, and schedule instability</li>
<li><strong>Working capital</strong> tied in inventory vs service level achieved</li>
</ul>
<h2>Embedded ERP forecasting vs best-of-breed vs data science stack: what matters most</h2>
<p>The “best” option depends on complexity, team maturity, and time-to-value. But integration points are non-negotiable:</p>
<ul>
<li><strong>Bidirectional sync</strong> of items, locations, customers, and calendars</li>
<li><strong>Clean handoff</strong> from forecast to MRP/DRP, safety stock, reorder points, and PO/WO generation</li>
<li><strong>Override governance</strong> and auditability inside the execution system (not in a spreadsheet)</li>
<li><strong>Feedback loop</strong>: stockouts, late suppliers, substitutions, and actual lead times must flow back</li>
</ul>
<p>Enterprise vendors are publicly signaling this direction. Oracle, for example, has highlighted AI embedded across supply chain workflows in 2025: <a href="https://www.oracle.com/news/announcement/oracle-helps-customers-optimize-global-supply-chain-efficiency-2025-01-30/" target="_blank" rel="noopener noreferrer">Oracle helps customers optimize global supply chain efficiency</a>. Meanwhile, customer stories continue to market large gains—SAP’s FORVIA case study claims a 65% demand planning accuracy increase using advanced forecasting algorithms: <a href="https://www.sap.com/documents/2025/11/9ed21497-2c7f-0010-bca6-c68f7e60039b.html" target="_blank" rel="noopener noreferrer">FORVIA demand planning with SAP IBP</a>. The lesson isn’t that one vendor “wins”; it’s that the winners operationalize forecasting as a closed-loop process tied to inventory and execution.</p>
<h2>How Likwid ERP helps teams operationalize AI forecasting without inventory bloat</h2>
<p>Likwid ERP is built for manufacturing and distribution complexity: procurement, production planning, inventory, and CRM in a single AI-powered open-source platform. That matters because AI forecasting only delivers when your data, policies, execution, and feedback live in one system of action. If you’re ready to move from isolated forecasts to a closed-loop planning and execution workflow, explore <a href="/ai-native-mrp">AI-Native MRP</a>, browse more resources in our <a href="/blog">blog</a>, or start with <a href="/blog/welcome-to-likwid-erp">Welcome to Likwid ERP</a> to see how an integrated platform can reduce stockouts and prevent inventory bloat while protecting customer service.</p>