n this episode, Noz Urbina interviews Ilya Venger, Data and AI Product Leader at Microsoft, to deliver a masterclass in practical AI implementation for business leaders. Ilya addresses the trillion-dollar question facing every executive: Should we build our own AI solution, buy off-the-shelf, or wait for the technology to mature? His answer: it depends on understanding your specific business problems, not chasing shiny technology. Key Takeaways The 80% Solution: Ilya reveals that AI systems work correctly about 80% of the time. Success isn’t about perfecting that last 20% through expensive fine-tuning – it’s about redesigning processes to work with AI’s probabilistic nature. As Noz puts it, “If you create a workflow with zero tolerance for error, you’ve designed a bad process.” The Fine-Tuning Trap: Ilya shares cautionary tales of companies spending millions to fine-tune models for specific problems (like the “six finger problem” in image generation), only to watch base models solve these issues within 18 months. His stark example: a model fine-tuned to be cheaper than GPT-4 became pointless when GPT-4’s price dropped tenfold. Data Reality Check: Both speakers agree that most organizations have “data heaps” – disconnected silos without understanding or metadata. Ilya’s metaphor: “You’ve got gold nuggets in a dark room. You need to turn on the lights first.” Organisations must understand their data landscape before implementing any AI solution. The Build vs. Buy Decision Framework: Build (Fine-tune): Only when you have extremely specific tasks with proprietary data (like recognizing manufacturing equipment or crop diseases) Buy: For most use cases, using off-the-shelf models with good system prompts and workflow design Wait: When your problem might be solved by next quarter’s model improvements What you’ll learn

• The build, buy, or wait decision framework – Clear criteria for when to fine-tune models (specific tasks with proprietary data), buy off-the-shelf solutions (most use cases), or wait for base models to improve
• Master the 80% solution – Why AI works correctly 80% of the time and three strategies to handle failures: improve the AI, modify your processes, or introduce human oversight
• Avoid the million-dollar fine-tuning trap – Real examples of why custom models become obsolete within 18 months and when fine-tuning makes sense
• Turn your “data heaps” into AI gold – How to assess and organize disconnected data silos before implementing AI, plus why most organizations fail at this critical first step
• Design systems, not magic genies – Why thoughtful system prompts and workflow design deliver 10x better ROI than chasing the latest AI model
• Handle AI’s “alien” failure modes – Understand how probabilistic systems fail differently than traditional software and build processes that expect interpretation errors
• Find your real competitive edge – Why your IP isn’t in having a custom model but in process design, context setting, and treating AI as “10,000 eager interns”
• Know when waiting beats racing – Recognise when today’s expensive problem (like the “six finger problem”) will be solved by next year’s base models

Noz Urbina interviews Rafaela Ellensburg, who has pioneered the content engineering discipline at Albert Heijn, one of the Netherlands' largest retailers. Rafaela discusses her journey from content specialist to content engineering leader, emphasising how structured content and metadata enable omnichannel measurement and personalisation at scale.

The conversation explores the evolution from content management to concept management, drawing parallels between content supply chains and traditional product supply chains.

Key topics include

• translating strategic business goals into measurable content metrics,
• implementing knowledge graphs and ontologies for cross-domain data connections, and
• preparing high-quality structured data to enhance AI reliability.

"You allow yourself as an organization to bring forward that message to whichever person it resonates with in the market, and you're able to do it on whichever channel that person is present. You get the relevance, and you get it at scale, at an omnichannel scale—making sure that the right message is sent to the right customer at the right moment and the right channel. That is the marketer's dream, right? That's what we all want." – Rafaela Ellensburg

"I like to compare content to products. People know products—they know shopping, they know logistics, they know that products are created somewhere and then have to be refined before they get to the stores. It's something that people can grasp, but we can do the same thing for content." – Rafaela Ellensburg

"We as humans actually have work to do to make our data of AI quality—more complete, richer, more consistent and truthful, so that whatever the AI does with that data, it becomes better. You do not get garbage in, garbage out, but you get value in, value out." – Rafaela Ellensburg

In this podcast, Noz and Amir Faizpour talk about how businesses can effectively implement AI beyond basic tools like ChatGPT.

Amir, who runs Aggregate Intellect, explains why companies might need more sophisticated AI solutions that integrate with their existing business systems rather than using standalone AI tools.

The conversation focuses on “AI agents” – systems that can independently use tools and execute complex tasks – and the importance of separating language models from actual business data to ensure accuracy and reliability.

A key takeaway is that instead of relying on one large AI model for everything, businesses might benefit more from using multiple specialised models for different tasks, much like how human workflows operate.

• Understanding why businesses might need solutions beyond standard AI tools like ChatGPT and Microsoft Copilot
• The importance of integrating AI with existing business systems rather than using siloed solutions
• How “agentic” AI systems work and their three key components: tool usage, memory, and planning capabilities
• The value of separating language models from actual business data and knowledge
• Why smaller, specialized AI models can sometimes be more effective than larger, general-purpose ones
• The role of human oversight and knowledge in AI systems
• How to evaluate and verify AI-generated content for accuracy and specificity.

“I have these nine commandments that I usually talk about when it comes to generative AI and one of them, which I think is the most important one, is the separating the knowledge, the data and the linguistic interface.”

“One of my biggest design principles is that the architecture of, or the anatomy, as you’re saying of the system, you’re building has to replicate the human workflow.” – Amir Feizpour