Demand Sensing vs. Demand Forecasting: Why the Question Itself Is Getting Obsolete

Demand Sensing vs. Demand Forecasting

Last updated September 3, 2026 · 9-minute read

Picture a mid-market retailer that rips out a monthly forecasting cadence and goes all-in on a sensing layer promising near-real-time SKU-level accuracy. Six months later, the system is beautifully responsive — and confidently wrong about anything with a lead time longer than three weeks, because sensing was never built to see that far. That failure mode shows up often enough in practitioner write-ups and vendor post-mortems that it’s worth treating as a pattern, not a one-off — and it’s the mistake this article is written to help you avoid.

What each one actually does ESTABLISHED

Demand forecasting takes historical sales and statistical or ML models (ARIMA, gradient boosting, LSTM) and projects demand weeks to years out. It’s the backbone of budgeting, procurement contracts, and workforce planning — anything with a lead time long enough that you need to commit before you know the actual number.

Demand sensing consumes near-real-time signals — point-of-sale data, distributor sell-through, web traffic, local weather, promotional activity — to correct the near-term picture, typically inside a 0–14 day window. It doesn’t replace the long-range plan; it tells you how wrong that plan is turning out to be, this week, at the SKU-and-location level.

The number everyone quotes, and why you should squint at it PROBABLE

You’ll see “20–50% reduction in forecast error” attributed loosely to McKinsey across dozens of vendor blogs, usually stripped of the caveats about model type, data maturity, and category volatility that made the original figure meaningful. Kearney’s 2023 research, cited by supply-chain planning vendor Kinaxis, is more conservative and more specific: 5–20% forecast-accuracy improvement and 5–10% safety-stock reduction from demand sensing specifically — that’s the number I’d actually plan around for a first-year business case rather than the flashier headline figure.

The pattern worth watching for: the 20%+ gains tend to show up at companies that had genuinely messy baseline forecasting — spreadsheets, no probabilistic bands. If your forecasting is already decent, sensing buys you inventory precision, not a miracle.

What’s actually new in 2025–2026 PROBABLE

The meaningful shift isn’t sensing itself — that’s been production-grade for years. It’s that vendors are now shipping forecasting and sensing as one “agentic” loop instead of two tools a planner manually reconciles. Worth being careful here, because “agentic” in 2026 marketing copy is doing a lot of work for a term that’s hard to falsify — it can mean anything from a genuinely autonomous reallocation trigger to a chatbot bolted onto a dashboard.

What’s verifiable, from the vendors’ own announcements: SAP unveiled agentic supply-chain agents at Hannover Messe in April 2026 — including a Production Planning and Operations Agent that validates material, capacity, and scheduling constraints before a planner approves an order — and followed up by formally launching “Autonomous Supply Chain Management” at Sapphire 2026, bundling dozens of Joule-branded agents across planning, procurement, and logistics. Separately, Manhattan Associates introduced Sightline in May 2026, a decision-intelligence layer inside ActivePlanning that surfaces the reasoning behind each AI-generated forecast or inventory recommendation, aimed at building planner trust in the agent’s output rather than just its accuracy. Those are shipped, dated announcements — not roadmap slides.

The scale of investment is real too. Mordor Intelligence puts the inventory optimization software market at $1.31B in 2025, growing to $2.31B by 2031 — though that figure is worth reading as one analyst firm’s scope, not a market consensus: other research firms define “inventory optimization software” more broadly and land anywhere from roughly $2.5B to $8.7B for the same 2025 window. The wide spread reflects genuine disagreement about what counts (standalone SKU-level tools vs. broader supply-chain planning suites), not a typo — so treat any single market-size figure, including this one, as directional rather than precise.

BCG’s own research found agentic systems accounted for roughly 17% of total enterprise AI value in 2025, projected to reach 29% by 2028. That’s an enterprise-wide AI figure, not specific to demand planning — I’m citing it here as directional context for where investment is heading, not as proof that agentic demand planning specifically is delivering 17% of anything. Treat it the way I’m asking you to treat the market-size range above: real signal, soft edges.

Deloitte’s 2026 Global Retail Industry Outlook, based on a survey of 330 retail executives fielded October–November 2025, found 30% of retailers already use AI for supply chain visibility, with that figure expected to climb to 41% within the year, and 59% of executives expecting positive ROI within 12 months. That’s real adoption momentum from a named, dated report — but “expect positive ROI” is a survey response about sentiment, not a measured outcome, and it’s worth keeping those two things separate even though vendor decks often blur them.

Where the money actually gets tight ESTABLISHED

A Gartner survey published August 5, 2026 found 55% of chief supply chain officers are unclear on the ROI of their AI investments, even as AI now accounts for 67% of supply chain digital spending — drawn from two separate surveys, one of 394 supply chain professionals at $250M+ revenue organizations (Nov. 2025–Feb. 2026) and one of 135 senior supply chain leaders (Jan.–Apr. 2026). That gap between spend and clarity is the story underneath the accuracy statistics above. A model can cut forecast error and still fail to move the P&L if nobody’s changed the ordering policy, safety-stock formula, or approval workflow downstream of it.

Why fashion and grocery see different numbers

The category caveat above is real, so here’s the actual mechanism, not just the disclaimer. Short-lifecycle categories — fashion, seasonal goods, anything without three-plus years of clean comparable history — start from a weak forecasting baseline. There’s not enough historical pattern for a statistical model to lock onto, so real-time signal (sensing) does proportionally more of the work, which is why sensing’s reported lift looks bigger in these categories. Staple goods — grocery, commodity CPG, anything with stable weekly demand — already have a well-fit historical baseline. Sensing still helps at the margins (a heatwave spiking bottled water demand, a local event), but it’s correcting a forecast that was already close, so the percentage gain looks smaller even when the dollar impact is comparable at scale. If your category swings hard on trend or weather, weight your evaluation toward sensing. If it’s stable and high-volume, put the budget into forecasting model quality first.

What “clean downstream data” actually means

I said above that most companies aren’t ready for sensing because the data underneath it is messy. That’s a fair criticism to make specific, so here’s the practitioner checklist worth working through before recommending sensing to a client:

• SKU and location match cleanly across systems. If your POS system’s SKU for a product doesn’t map 1:1 to the same SKU in your ERP and your forecasting tool, sensing signal gets lost in translation before it ever reaches a decision.
• Sell-through data arrives on a predictable cadence. Daily is standard; weekly batch feeds defeat the purpose of a 0–14 day sensing window.
• Someone owns exception handling. A sensing signal that contradicts the baseline forecast needs a human or a clearly-scoped automated rule to act on it — not a dashboard nobody checks.
• Lead-time and supplier data is current. Sensing tells you demand shifted; it’s useless if your system doesn’t know your actual replenishment lead time to act on that shift.
• You can already answer “why did we stock out last month?” without a multi-day data-pull exercise. If that takes a week to answer today, sensing will just generate faster alerts you still can’t act on quickly.

If more than one or two of those aren’t true yet, fix them before buying a sensing layer — the tool will not fix them for you.

A rough decision frame

• Use forecasting alone if your lead times are long, demand is stable, and your main cost driver is procurement commitment, not shelf-level stockouts.
• Add sensing once the checklist above is mostly true and stockouts/markdowns are a measurable cost — not before.
• Hold off on full agentic orchestration until a human-run version of the sensing-to-action loop is already working. Automating a broken process just breaks it faster.

The forecast tells you what to expect. Sensing tells you what already changed. Most companies don’t need a smarter model — they need to actually look at both at the same time.