COMMERCE StoriesSubscribe
← Commerce courseCOMMERCE FROM ZERO · MODULE 20 · 15 min READ · FREE

AI across the chain

From the shopkeeper's memory to machines that plan, buy and move goods

Labels: [Sourced] checked against a source listed at the end. [General] standard knowledge, not tied to one source. [Our view] opinion, labelled as such.

1. The story: the shop that guessed a year ahead

For most of the 20th century, a clothing retailer had to decide what to make about a year before the customer walked in. Harvard's student write-up of Zara describes the traditional approach: forecast trends roughly 12 months out, then build the production plan around that guess. Zara did the opposite. Store managers sent back hard data (orders, sales trends) and soft data (customer reactions, "buzz" around a style), and head office reacted to trends as they appeared, running factories with spare capacity so they could speed up fast. [Sourced: HBS D3]

That is the whole story of AI in the supply chain in one idea: shrink the gap between what customers do and what the business does about it.

Zara did it with people, phone calls and fast factories. AI does it with data and software, across every step.

2. The one idea

Every step in a supply chain is a guess about the future. AI makes the guesses better, faster, and cheaper, and now it can also act on them.

Three kinds of AI show up in commerce. [General]

  • Predicting: how many will sell, which invoice will be paid late, when a truck will arrive.
  • Optimising: the best route, the best price, the best place to hold stock.
  • Acting (agents): software that does the task itself, such as reordering stock or booking a carrier, within limits you set.

3. The kitchen-table version

Think of a neighbourhood baker.

  • The old way: she bakes what she baked last Saturday.
  • The better way: she looks at the weather, the school calendar and a street festival, and bakes more.
  • The AI way: a helper does that homework every morning for every item, and tells her, or even places the flour order itself.

Same baker, same ovens. A better guess each morning.

4. Where AI sits, step by step

Our running example is a $12 water bottle. Follow it through the chain.

Step (module)The old guessWhat AI addsEvidence
Forecasting (4)Last year's sales plus a hunchModels that learn trends, holidays, promotions at SKU-store-week levelMcKinsey: AI forecasting can cut errors 20 to 50 percent, and reduce lost sales from unavailable products by up to 65 percent [Sourced]
Demand sensing (6)Wait for next month's planShort-term forecasts refreshed continuously from live orders, shipments and outside signalsInfor, SAP describe demand sensing as a short-term layer on top of forecasting; SAP's version covers roughly the next 4 to 8 weeks [Sourced]
Inventory (9)A fixed "keep 100" ruleStock levels that move with the forecastMcKinsey: autonomous planning at consumer goods firms reported up to 20 percent less inventory and up to 10 percent lower supply chain cost [Sourced]
Warehouse (10)Same pick path every daySlotting and routing tuned to today's orders[General]
Transport (11, 12)Fixed routesDynamic routing, delivery-time prediction[General]
Finance (13 to 16)Late invoices surprise youPredicting who pays late, matching invoices, spotting fraudcovered in Module 16
Merchandising and pricing (17)Cost plus a marginPrice tests, assortment by storecovered in Module 17
Buying itself (21)A human at a screenAI agents that shop for peoplebelow

Numbers marked "up to" are the best cases the authors report. Your result will be smaller. Treat them as a ceiling, not a promise.

5. Worked example: why a better forecast saves money

Suppose the bottle sells 100 a week on average, but your forecast is off by 30 units either way (a 30 percent error). To avoid running out, you keep a buffer of 30 extra bottles.

Now suppose a better model halves the error to 15 units. [Illustrative numbers, our own]

Old forecastBetter forecast
Average weekly sales100100
Typical miss3015
Buffer stock needed3015
Cost per bottle$5$5
Cash tied up in buffer$150$75

One product, one store, $75 freed. Multiply by thousands of SKUs and hundreds of stores and the value is real. The same halving also means fewer empty shelves. (This simple "buffer equals the miss" rule is a teaching shortcut. Real systems use statistics for the buffer. We cover it in Module 9.)

6. A short history of how forecasting got smarter

This is the history thread: each era shortened the time between customer action and business reaction.

EraWhat the retailer knewHow fast
Village shopThe shopkeeper's memoryDaily
Mail-order catalogs, 1872 onwardOrders that arrived by post. Montgomery Ward's first catalog in 1872 was a single sheet of 163 items [Sourced: Wards]Weeks
Barcode scanning, 1974Every sale, by item [Sourced, see Module 1]Daily
Big-box logistics, 1960s onwardWalmart opened its first store July 2, 1962 in Rogers, Arkansas [Sourced: Walmart], later famous for data-driven replenishment [General, to be expanded in Module 8]Days
Online retail, 1995Amazon opened as an online bookseller in July 1995 [Sourced: Amazon]. Every click is recorded, not only every saleMinutes
AI eraSignals from orders, weather, events, searchesHours

7. What changes next: from helpers to agents

Until now, AI mostly advised people. The new step is software that acts.

  • In 2025 and after, companies began publishing standards so AI agents can complete purchases for people. The Agentic Commerce Protocol is described by Stripe as an open standard created by Stripe, OpenAI and Meta for how AI agents interact with businesses to complete purchases for buyers. [Sourced: Stripe docs] Google published its own open standards: the Agent Payments Protocol (AP2), announced September 16, 2025, and the Universal Commerce Protocol (UCP), announced January 11, 2026. [Sourced: Google Cloud blog, Google Merchant docs, Google blog]
  • On the business side, "autonomous planning" aims to let software adjust forecasts and orders with people watching exceptions. [Sourced: McKinsey]

What we think, clearly labelled [Our view]:

  • The winners will not be the ones with the fanciest model. They will be the ones with clean product data (SKUs and GTINs, see Module 1), clean inventory counts, and clear rules for what an agent may do alone.
  • A note on words. You will hear "AI", "AGI" and "superintelligence". "AI" is the umbrella. "AGI" means AI that is broadly capable like a person, and nobody agrees on exactly where that line sits. "Superintelligence" is the classic term for AI far beyond the best human experts in nearly every field. Companies use it in different ways, and no one has shown it exists. We say "AI" in this course, and treat superintelligence as the direction of travel. [Sourced: Bostrom; Scientific American, see our research brief]
  • For commerce, the useful version is simple: software that handles routine decisions end to end, so people spend time on judgment, partnerships and exceptions. [Our view]

8. Limits and risks

  • Bad data in, bad guesses out. AI cannot fix a messy SKU list. (McKinsey notes that most organisations have enough data to start, but data quality still matters.) [Sourced, paraphrased]
  • Surprises. Models learn from the past. A shock with no precedent, like a port closure, still needs human judgment. McKinsey's own forecasting article stresses letting users adjust scenarios for events history cannot show. [Sourced, paraphrased]
  • Guardrails. An agent that can spend money needs spending limits and an approval step. [General]

8b. Watch-out boxes: where smart-looking AI goes wrong

Accuracy without bias. A forecast can score well on an average error number and still lean high or low every week. Pair an error measure with bias. [Sourced: Inventory Controller; see Module 4]

Slow sellers break the usual score. Percentage error blows up when sales are zero, and intermittent demand needs different methods and different error measures. [Sourced: Hyndman on Croston; see Module 4]

People overriding the machine. Forecast Value Added asks whether manual edits improve the forecast. A 2024 critique argues the metric cannot confirm that an override was right for the reason believed, so the debate is open. [Sourced: Foresight / Lokad critique; see Module 5]

One safety-stock number for everything. There is no universal formula; the right buffer depends on demand and lead-time variability and the service level you want, and works best with ABC/XYZ classes. [Sourced: IBM; MIT/APICS note; ASCM; see Module 9]

Hype check. Gartner predicts that by 2030 only 5% of organizations will make at least 10% of supply chain planning decisions autonomously, a useful counterweight to "up to" gains in vendor and consulting claims. [Sourced: Gartner, Sept 24 2026]

Words differ by context. "Demand sensing" is a vendor term for short-horizon forecasting with fresher inputs (Module 4, Section 8). "Superintelligence" is used three ways (see the note on words above). When a term is new to you, ask what it measures and who benefits from the label. [Our view]

9. Try it yourself (no code, 15 minutes)

  1. Pick something you buy every week (coffee, bread, diapers).
  2. Write what you think you will use next week and how wrong you were the last 4 weeks.
  3. Write one outside signal that would change your guess (a holiday, a sale, guests).
  4. Work out your own "buffer": how many extra do you keep to avoid running out?
  5. Ask yourself: what is one routine decision here a helper could make for you, and what limit would you set?

10. Self-check

  1. What does demand sensing add on top of forecasting? a) A one-year plan b) Short-term updates from fresh signals c) A new barcode. Answer: b.
  2. Which is an "agent"? a) A chart of sales b) Software that places an order within set limits c) A faster printer. Answer: b.
  3. A forecast miss halves, from 30 to 15 units, at $5 a bottle. How much cash is freed in this example? Answer: $75.
  4. Do the "up to 65 percent" results guarantee you the same? Answer: No. They are best cases.
  5. What did Zara do differently from the traditional 12-month forecast? Answer: reacted to trends as they appeared, using store data and fast, flexible factories.

Sources

CURIOUS? TEST THE CLUES

Curiosity check

Pick an answer and see why. No scores, no pressure. All shop examples are invented practice scenarios.

01 What does demand sensing add on top of forecasting?
02 Do the 'up to 65 percent' results guarantee you the same?