Why Companies Struggle with AI: Between Hype and Reality

Summary

The podcast "KI-Podcast" discusses why companies struggle to successfully implement artificial intelligence despite high expectations. Hosts Gregor Schmalzried and Fritz Espenau speak with experts like Alexander Tam and Elisabeth Lorange about technical, organizational, and cultural barriers. The central problem lies not in the technology itself, but in lacking strategy, missing clear responsibilities, and employee-level resistance. The widely cited statistic of 95 percent failed AI projects is critically examined.

People

• Gregor Schmalzried
• Fritz Espenau
• Alexander Tam
• Elisabeth Lorange

Topics

• AI integration in companies
• Vertical vs. horizontal innovation
• Compliance and data protection as barriers
• Employee involvement and corporate culture
• Difference between AI-fication and AI transformation
• Autonomous vehicles and Waymo

Detailed Summary

The Starting Point: San Francisco vs. Germany

The episode begins with Fritz Espenau traveling to San Francisco for research purposes – the birthplace of the current AI revolution. He notices that practically every bus advertises AI tools, but not consumer products like ChatGPT or Claude, rather specialized solutions for AI startups. This demonstrates: providers expect their biggest profits not from individual users, but from corporate customers – and precisely those are struggling the most.

Vertical vs. Horizontal Innovation

A core concept in the discussion is the distinction between two approaches:

Vertical Innovation (top-down): Management decides centrally that AI should be implemented. This works well for highly specialized applications like radiology, where specialized systems are rolled out by experts – similar to traditional software implementation.

Horizontal Innovation (bottom-up): Employees experiment decentrally with available tools and discover applications themselves. The problem: not everyone has the time, energy, or motivation to experiment.

Alexander Tam, data science pioneer and founder of a consulting firm (founded 2012), describes his shift from strictly use-case-oriented work toward generalization through language models. While previously every problem was solved with specialized algorithms, modern LLMs like GPT-5 and Claude 4.5 enable flexible, reusable solutions – provided they are implemented correctly.

The Central Problem: Implementation

The main problem is not the technology, but concrete implementation. Many companies provide tools to employees without clear guidelines or training. If someone wakes up on Monday thinking "I should use more AI," they often don't know where specifically to start.

Organizational Barriers (according to Elisabeth Lorange from Deloitte):

• Procurement/Purchasing: By the time tools are purchased, they're already outdated
• Compliance and data protection: Strict regulations limit flexibility
• Works council: Codetermination rights delay decisions
• Missing responsibilities: Unclear who has decision-making authority (CFO, CTO, Chief AI Officer)
• Missing strategy and clear tools: Often there's no overarching AI strategy

The "95-Percent-Failure" Statistic

A frequently cited myth is deconstructed: the claim that 95 percent of all AI projects fail allegedly comes from MIT. In reality, it is based on:

• Only 52 interviews (small data, not big data)
• Very short success measurement: only 6 months to profitability
• Narrow definition of success that ignores many AI effects (difficult-to-measure quality improvements)
• Outdated data: Cloud Code didn't exist at that time

The actual effectiveness of AI projects is difficult to measure, as many benefits are subtle (better work, not just time savings).

Difference: Microsoft Copilot vs. Integrated Solutions

A concrete example is Microsoft Copilot. Microsoft CEO Satya Nadella is reportedly dissatisfied with it according to a December report from The Information. The Copilot often delivers unexpected results and cannot reliably solve simple tasks like "Find my first email to person X."

Ironically, Microsoft internally uses Claude Code itself, even though the company doesn't officially recommend this solution – simply because it works better.

Human Factors and Resistance

A major problem lies at the emotional and cultural level:

Fear of replacement: Studies from the Harvard Business Review show that employees who use AI are often disadvantaged by supervisors (fewer promotions, lower salaries).

Distrust of AI results: A listener named Marco reported that although his company recommended AI use, colleagues rejected results as soon as they realized AI was involved.

AI-Slop Problem: Employees use AI but don't revise the results. This leads to poor documents that other colleagues then have to rework. This creates systematic distrust.

Two extremes: Some people have exaggerated fears about AI (because they don't use it), others overestimate it (because they once had a great result). Both extremes block rational use.

How to Motivate Employees for Proper Use

Instead of prescribing AI to employees "from the top," companies should:

• Respect processes: Don't simply replace a legal work process with AI, but jointly develop a workflow that fits individual style
• Create experimental freedom: Time and space for mistakes
• Incentivize knowledge sharing: Show appreciation when employees share AI knowledge – don't punish with more work
• Corporate culture: A culture where innovation and mistakes are accepted
• Real training: Not just presenting tools, but accompanying the process

AI-fication vs. AI Transformation

An important conceptual distinction:

AI-fication: Digitally optimizing existing processes (e.g., AI summarizes meeting notes).

AI Transformation: Fundamentally reshaping processes so old steps become unnecessary (e.g., AI cross-references information, the meeting doesn't happen at all).

Most companies are still in the AI-fication phase. True transformation requires strategic thinking and courage for process change.

Four Dimensions of Successful AI Integration

Alexander Tam summarizes:

1. Strategy: Where does the company want to go? What problems should be solved? Clarify power relations.
2. Implementation: Which use cases are relevant? Product development mindset instead of purely functional organization.
3. Technology: Architecture, tools, data availability, integration.
4. Change management: Explanation, training, gamification elements, cultural accompaniment.

Critical: If even one dimension is neglected, the project fails.

San Francisco and the Waymo Experience

Fritz Espenau reports on rides in self-driving Waymo taxis. In San Francisco, these are completely normal (10–20% of street traffic). This raises an important question: is technological development splitting between the USA and Europe? Waymo is highly specialized for individual cities, Tesla's approach (cameras only, no LiDAR) could be more universal, but is not yet production-ready.

Practical AI Applications of the Hosts

Gregor: Used ChatGPT, Claude, and Google Gemini to find out why a hotel booking became cheaper. Only Gemini researched independently and discovered that a shoe trade fair was simultaneously happening in the city – explaining the higher prices in the first week.

Reader contribution (Kilian): Built a pipeline for political/social research: Deep Research in ChatGPT → Google Gemini → Notebook LM → generated 15-minute podcast to listen to. Highly efficient for complex topics, but impossible to measure in classical ROI terms.

Key Takeaways

• The biggest problem is not the technology itself, but implementation: Companies lack strategy, clear responsibility, and change management.

• The "95-percent-failure" statistic is misleading: It's based on 52 interviews with unrealistically short success measurement and ignores difficult-to-quantify improvements.

• Microsoft Copilot is a cautionary example: Despite big promises, it often delivers poor results in practice, while Microsoft internally uses Claude Code.

• Employees need trust, not prescriptions: AI only works if workers can experiment with it and mistakes are accepted – not if they're punished for using AI.

• Distinguishing between AI-fication and AI transformation is crucial: True value creation comes only through process redesign, not through mere digitization.

• AI integration requires four dimensions: Strategy, implementation, technology, and change management – omitting even one leads to failure.

• Regional differences are growing: San Francisco already has autonomous taxis in everyday operation, while Europe is moving significantly slower due to regulatory barriers.

• Success is difficult to measure: Many AI effects (better quality, easier work) cannot be translated into classical ROI metrics.

Metadata