AI Projects Are Failing | Here's What Everyone Gets Wrong
In this episode of An Hour of Innovation podcast, Vit Lyoshin speaks with Max Vermeir about why so many AI initiatives fail in real-world business environments, and why the problem is rarely the LLM models themselves.
Max explains that while large language models have advanced rapidly, most organizations struggle because their data is messy, unstructured, and not ready for AI to use effectively. From documents and spreadsheets to complex workflows, the real challenge lies in turning human-generated data into something machines can reliably interpret.
The conversation explores the difference between deterministic and probabilistic AI, highlighting why LLMs always produce an answer, even when it’s wrong, and what that means for trust and reliability in production systems. Max also shares practical examples where simpler solutions outperform complex AI models, emphasizing the importance of choosing the right approach instead of defaulting to AI for every problem.
Vit and Max discuss what product managers and engineers often underestimate when building AI-powered products, including hidden costs, edge cases, and the difficulty of scaling beyond a demo. Max outlines a practical framework for implementing AI features that actually work in production, focusing on data readiness, clear use cases, and measurable value.
The episode concludes with a look at the future of AI, including the rise of AI agents and how engineering roles are evolving toward orchestration rather than hands-on coding. Throughout the conversation, Max offers a grounded, no-hype perspective on AI, one that focuses on building systems that deliver real business impact.
Max Vermeir is the VP of AI Strategy at ABBYY, where he focuses on applying AI to real-world business problems, especially around document processing and automation. He works at the intersection of AI technology and enterprise systems, helping organizations turn unstructured data into usable insights. Max is known for his practical, no-hype perspective on AI, focusing on what actually works in production rather than demos. His experience gives him a unique view into why many AI projects fail and how to build systems that deliver real value.
Topics Discussed
- Why most AI projects fail in enterprise environments
- The real challenge of unstructured data for LLMs
- How document AI impacts business automation and workflows
- Deterministic vs probabilistic AI systems explained simply
- Why LLMs always return answers, even when incorrect
- Hidden costs of using AI models in production systems
- Why simple solutions often outperform complex AI models
- What product managers misunderstand about AI implementation
- Framework for building reliable AI products in production
- Future of AI agents and engineering productivity shifts
Timestamps
00:00 Introduction
01:35 AI in Business Operations
03:56 Understanding Document AI
06:57 Challenges in Document Processing
10:15 The Importance of Data Accuracy
11:59 Deterministic vs. Probabilistic AI
16:38 High ROI Use Cases for Document AI
18:50 Common Underestimations in AI Product Development
21:48 Engineering Challenges in AI Projects
25:26 Implementing AI Feature as a Product Manager
27:36 Frameworks for AI Implementation
32:00 The Impact of AI on Engineering Productivity
35:35 Skills for the Future: Adapting to AI
40:08 The Resurgence of Command Line Interfaces
41:55 The Rise of AI Agents
47:55 Innovation Q&A
Connect with Max
- Website: https://www.abbyy.com/
- LinkedIn: https://www.linkedin.com/in/maximevermeir/
This Episode Is Supported By
If you enjoy this podcast, you can support us by exploring the tools below.
These are affiliate links, meaning we earn a small commission at no extra cost to you.
- WhatConverts: Prove and Grow Marketing ROI.
- Webflow: Create custom, responsive websites without coding.
- Vista Social: Effortlessly manage high volumes of content across multiple social media channels.
- Optery: Remove your home address, phone, and other private info from Google, and 955+ sites.
For inquiries about sponsoring An Hour of Innovation, email iris@anhourofinnovation.com
Connect with Vit
- LinkedIn: https://www.linkedin.com/in/vit-lyoshin/
- Substuck: https://anhourofinnovation.substack.com/
- X: https://x.com/vitlyoshin
Episode References
Optical Character Recognition (OCR)
https://en.wikipedia.org/wiki/Optical_character_recognition
A technology used to convert different types of documents (scanned paper, PDFs) into machine-readable data.
Large Language Models (LLMs)
https://en.wikipedia.org/wiki/Large_language_model
AI models like GPT that generate and understand human language, widely used in modern AI applications.
Regular Expressions (Regex)
https://en.wikipedia.org/wiki/Regular_expression
A rule-based method for pattern matching in text, often simpler and more efficient than AI for structured tasks.
AI Agents
https://en.wikipedia.org/wiki/AI_agent
Autonomous systems that can perform tasks, automate workflows, and interact with tools or environments.
Cron Jobs
https://en.wikipedia.org/wiki/Cron
A time-based job scheduler in Unix-like systems used to automate repetitive tasks.
Deterministic vs Probabilistic Systems
https://www.rudderstack.com/blog/deterministic-vs-probabilistic/
Two approaches to computing - deterministic systems produce consistent outputs, while probabilistic systems (like AI) generate likely outcomes.
