AI Showreel consulting-grade analysis, for everyone FR
← The index
Proof C Live confirmed

Otto

deep-learning demand forecasting triggering autonomous supplier orders

IndustryRetail & e-commerceLeverActivation / conversionFamilyPredictionImplementationHybridStagepurchase
7,5 milliards
Individual forecasts produced each month
"the AI makes 7.5 billion individual forecasts" S2

Otto predicts short-term demand with a deep-learning model that reaches 90% accuracy on 30-day sales, cuts stock by 20%, and automatically orders merchandise from third-party brands, producing 7.5 billion individual forecasts a month.

Objective

Predict what customers will order before they do, to order the merchandise ahead of time, cut overstock, and deliver faster from stock already in the right place.

The deployment

Otto is one of Germany's largest e-commerce retailers. A large share of its catalog comes from third-party brands, with lead times that made availability hard to hold. Otto put into production a forecasting engine (a deep-learning algorithm derived from particle-physics tools) that predicts short-term demand from hundreds of variables: sales history, site traffic and searches, weather, day of the week, holidays, competitive data. When the forecast is confident enough, the system orders the merchandise from the brands on its own, without human validation, to have it in stock before the customer order. On otto.de, the AI produces 7.5 billion individual forecasts a month to drive warehouse arrivals, sales volumes, and returns.

Results Proof C

7,5 milliards
Individual forecasts produced each month
"the AI makes 7.5 billion individual forecasts" S2
90%
Forecast accuracy on 30-day sales
"predict with 90% accuracy on any given day what items will be sold" S3
20%
Reduction in stock levels
"reduce inventory levels by 20%" S3

The Economist (major press) documents the Otto case and its figures by name, corroborated by Otto's official site (7.5 billion forecasts/month). The accuracy figure originally comes from a case study by the vendor Blue Yonder, hence no A level.

How it works

Documented architecture
commande autonome au-dela du seuilcas incertainsventes reelles reinjectees Ventes, trafic,recherches, meteo, joursferies, concurrence Moteur de prevision deeplearning Blue Yonder Systeme d'achat et dereapprovisionnement Otto Acheteur (cas sous seuilde confiance) Marques tierces /entrepot

The stack in detail

How it runs, concretely

For ops teams
CadenceContinuous forecasting at very large scale (7.5 billion forecasts a month), supplier orders triggered automatically when confidence is sufficient.
Operated byOtto's data science and purchasing team, with the vendor's forecasting engine at the core.
  1. 1
    Signal collection Otto systems / data team

    Sales history, site searches and traffic, weather, day of the week, holidays, and competitive data are aggregated.

  2. 2
    Demand forecasting AI model

    The deep-learning model predicts what will sell in the short term, with 90% accuracy on 30-day sales.

  3. 3
    Autonomous ordering AI model / purchasing system

    Above a confidence threshold, the system orders the merchandise from third-party brands on its own, with no human validation.

  4. 4
    Human control below threshold Otto buyer

    Uncertain cases stay with a buyer, who keeps control of items with unstable demand.

The signal that drives it

Predicted short-term demand per product. An order is only automated if the forecast is confident enough; below that threshold, it goes back through a human buyer.

How your customers perceive this type of use

Sourced studies

C'est la famille la moins acceptee : 68% des Americains jugent inacceptable un score financier personnel calcule par algorithme et 67% l'analyse video automatisee d'entretiens d'embauche (Pew Research, 2018). La demande d'explication et de recours est massive : 83% veulent savoir quelles donnees l'IA utilise et 91% veulent pouvoir corriger des donnees erronees (Consumer Reports, 2024). A l'echelle mondiale, seuls 46% se disent prets a faire confiance aux systemes d'IA et 70% jugent une regulation necessaire (KPMG / Universite de Melbourne, 2025).

68%
Americains qui jugent inacceptable un score de finances personnelles calcule par algorithme pour proposer des offres (2018)
67%
Americains qui jugent inacceptable l'analyse video assistee par ordinateur des entretiens d'embauche (2018)
58%
Americains qui pensent que les programmes informatiques refleteront toujours un certain biais humain (2018)

Acceptance conditions

  • Transparence sur les donnees utilisees : 83% des Americains la reclament (Consumer Reports 2024)
  • Droit de correction des donnees erronees : 91% le demandent (Consumer Reports 2024)
  • Explication de la logique de decision : 44% des consommateurs sont plus enclins a utiliser un agent IA si sa logique est clairement expliquee (Salesforce 2024)
  • L'acceptabilite depend du contexte de la decision : 50% des Americains jugent equitable un score de risque criminel pour la liberation conditionnelle, contre 32% pour un score financier applique aux consommateurs (Pew Research 2018)

Red lines

  • La decision opaque et sans recours sur l'emploi, le credit ou le logement : 45% tres mal a l'aise pour l'embauche, 39% pour le pret, 39% pour le logement (Consumer Reports 2024)
  • Le scoring des personnes a partir de donnees comportementales : 68% le jugent inacceptable pour les offres financieres (Pew Research 2018)

Sources: Pew Research Center 2018 · Consumer Reports 2024 · KPMG / Universite de Melbourne 2025 · Salesforce 2024

See full acceptance: by country, by use, by generation

How to replicate

Inference, not sourced

Data prerequisites

  • granular sales history
  • behavioral site signals (traffic, searches)
  • external data (weather, calendar)
  • supplier lead times and reliability

Org prerequisites

  • agreement to let an order go out without human validation above a threshold
  • supplier relationships compatible with advance orders

Possible stack

  • Blue Yonder
  • in-house forecasting (deep learning) + threshold rules
  • RELEX, o9
Team to operate2-3 data scientists/ML engineers + 1 data engineer + the buyers, who keep control of items below the confidence threshold

The plan, step by step

  1. Step 1
    Aggregate granular sales history, site signals (traffic, searches), and external data (weather, calendar), plus supplier lead times.Deliverable: Consolidated, documented feature dataset
  2. Step 2
    Train or configure the forecasting model on a category and backtest it against the buyers' current method.Deliverable: Accuracy measured vs the existing method, by forecast horizon
  3. Step 3
    Connect the forecasts to the purchasing process in recommendation mode: the buyer validates each proposed order.Deliverable: Order recommendations integrated into the buyer tool
  4. Step 4
    Define the confidence threshold with purchasing and management, then automate orders above the threshold on a limited scope.Deliverable: First autonomous orders in production
  5. Step 5
    Expand the automated scope and track stock, availability, and delivery times.Deliverable: Stock assessment and governance rules for autonomous orders

First step: Measure the accuracy of a short-term forecast on a category, define a confidence threshold above which the order can go out without validation, then open it up gradually.

Sources

  1. S1 How Germany's Otto uses artificial intelligence Established press economist.com · 2017-04-12 · accessed 2026-07-11 archive pending
  2. S2 Artificial intelligence at OTTO - already part of all business processes today Primary otto.de · 2023 · accessed 2026-07-11 archive pending
  3. S3 Autonomous Stock Replenishment at Online Retailer OTTO (Harvard TOM / HBS AI Institute) Secondary aiinstitute.hbs.edu · 2018 · accessed 2026-07-11 archive pending