Why AI Projects Fail Because of Data Architecture
This episode was recorded in German. YouTube may offer a dubbed audio track.
Show Notes
Many AI pilots start out promising and end up in a drawer after a few months. Gartner expects companies to abandon around 60 percent of AI projects that are not based on AI-ready data by the end of 2026. Matthias Dietrich and Philipp Krueger discuss what a data architecture has to deliver so that AI works in day-to-day operations, and what role the data lakehouse plays.
The path leads from the data warehouse, which delivers a report once a month, via the data lake to the data lakehouse, which makes data usable in ongoing operations. Using examples from retail and industry, they show how customer value and margins flow into search, CMS and the shopping cart. Talk to your data replaces the report that waits weeks in the sprint. Philipp describes a project in which the same product had a different contribution margin depending on the department. For an AI model, that remains a problem as long as finance, purchasing and sales have not agreed on one definition.
Finally, it gets practical. How does a company start when ERP, PIM, CRM, shop and POS each hold their own data? Which sources belong at the beginning, and how are product, customer and order defined consistently? And why does it pay to start with a concrete business question and a first department?
Alongside the episode, Springer Professional published the article "AI projects often fail because of data architecture" by Matthias Dietrich (in German). Using an omnichannel retailer as an example, he shows how a simple benchmark brought a stalled returns forecast into production.
Read the article at Springer Professional
What Works Tomorrow. The podcast for decision-makers in retail and industry. By Matthias Dietrich and Philipp Krueger, Co-CEOs of foobar Agency.
#AI #DataArchitecture #DataLakehouse #eCommerce #WhatWorksTomorrow
About the Podcast
What Works Tomorrow
What Works Tomorrow is the podcast for leaders in retail and industry. Philipp and Matthias talk with decision-makers about how companies use data, AI and commerce to reach their…
All EpisodesThe Hosts
Matthias Dietrich
Co-CEO
Philipp Krueger
Co-CEO
Subscribe
Chapters
- 00:00Preview
- 00:19Intro
- 00:40What it is about: making data usable
- 01:07The data warehouse as a warehouse on the edge of town
- 02:34Data lake and data swamp
- 03:23The data lakehouse
- 04:42Data in ongoing operations instead of a monthly report
- 05:37Customer value directly in the frontend
- 06:22Rigid ERP systems and fast steering
- 10:15AI and talk to your data
- 12:05Margen operationalisieren
- 16:02Omnichannel: Wer online optimiert, verfehlt den Laden
- 18:35Warum Datenprojekte als IT-Projekt scheitern
- 20:04Change und schrittweise Einführung
- 22:36Ein Produkt, mehrere Deckungsbeiträge
- 26:33Wie komme ich dahin?
- 28:24Architektur, Datenmodell und Integration
- 30:52Datenquellen von API bis AS400
- 32:59Nutzung und Ausleitung der Daten
- 35:21Datenplattform im B2B-Vertrieb
- 37:52Fazit
Related Articles
From Snowflake to AI: Predictions and LLMs on the Commerce Data Model
Three maturity stages — from deterministic calculation through probabilistic prediction to LLM-driven service co-pilots. What Snowflake Cortex delivers today and which prerequisites need to be in place first.
Data Foundation for B2B Industry: Forecasting, After-Sales, Service Data Flows
If your forecast accuracy hangs on the gut feeling of your head of sales, and your field-service technician collects a spreadsheet of wear-parts at the end of the day, you do not need the next platform — you need a foundation that brings ERP, service tickets and IoT signals onto a common state.
RFM and CLV: Who Your Most Valuable Customers Are — and Which Ones You’re Losing
CLV is not a reporting KPI. CLV is the answer to the question of which customers you should invest in today — and which ones you shouldn't anymore.
Snowflake in Commerce: From Data Silo to Shared Architecture
When marketing, sales and controlling each report a different revenue figure, it isn't a reporting problem. It's an architecture problem.