AI Doesn’t Fail Because of Bad Models. It Fails Because of This

This blog post is adapted from a keynote delivered by Mahdi Karabiben at the 2025 Data Innovation Summit Nordics.

‍

In the rush to embrace AI, most companies are focused on the shiny stuff: foundation models, copilots, real-time personalization. But beneath it all lies something less visible, more fragile, and absolutely essential: trust in the data.

And here’s the hard truth: without trust, none of it works.

You’ve likely seen it yourself. A machine learning model performs beautifully, until it doesn’t. A dashboard fails right before a leadership meeting. A key metric suddenly swings without explanation. Often, the root cause isn’t the algorithm or the analysis. It’s the data.

This isn’t about placing blame. It’s about understanding what happens when modern data environments become too complex to manage with old habits. And it’s about recognizing that data trust doesn’t scale on its own.

Why data trust breaks as you scale

A few years ago, a small team could manage a data pipeline end to end. But today? You’re dealing with cloud warehouses, reverse ETL, dbt models, dozens of dashboards, and multiple layers of business logic. It’s fast-moving and decentralized and while that unlocks speed, it also introduces risk.

Relying on spot checks or reactive alerts just doesn’t cut it anymore.

The consequences sneak up on you: minor bugs that cascade into outages. Hours lost debugging broken pipelines. A growing sense of hesitation among business stakeholders. Trust fades. And when it does, adoption slows, and the ROI of your entire data investment starts to wobble.

It’s not just a quality issue. It’s a confidence issue. A visibility issue. A systems issue.

Three ways to build observability that scales

So how do leading data teams build trust into the system and make it scale across use cases, personas, and platforms?

1. Start with strong foundations

Observability shouldn’t be bolted on. It needs to be woven into every layer of the stack from ingestion through transformation to final consumption.

That means using automation to detect anomalies and schema changes. It means treating metadata as an asset, not an afterthought. And it means integrating trust signals directly into the tools your team already uses.

If observability only lives in one tool or one team, it will always fall short. When it’s part of the system, trust becomes ambient and dependable.

2. Prioritize what really matters

You can’t monitor everything with the same intensity. That’s why the smartest teams prioritize observability around the data that drives real business outcomes.

Start with your top use cases, the dashboards that influence revenue, the models that power growth. Understand their dependencies. Put guardrails in place where they matter most.

Context is everything. A stale staging table is fine. A stale executive metric? That’s a crisis. Your observability approach should know the difference.

3. Make it work for everyone

Engineers. Analysts. Business stakeholders. They all need to trust the data, but in very different ways.

Engineers want control through code. Analysts need clarity in tools they actually use. Executives just want to know the numbers are right.

Scalable observability isn’t one-size-fits-all. It’s a shared language that adapts to different roles without adding friction.

Where this is all going

Data environments aren’t getting simpler. And AI is raising the stakes.

That’s why the future of observability lies in smart, autonomous systems that don’t just raise red flags, but help teams triage issues, understand impact, and act faster.

These systems won’t replace humans. They’ll make it possible for humans to keep up and stay ahead.

The quiet foundation every AI strategy needs

Great AI doesn’t start with bigger models. It starts with dependable data.

If you’re building toward an AI-powered future, now’s the time to ask:

• Does your observability strategy scale with your ambitions?
• Are you building trust across every layer and every team?
• Can your system spot risk before it turns into cost?

Because in this new era, it’s not just about faster decisions or smarter models. It’s about confidence. And confidence only comes from trust.

No trust, no AI.

And that’s a bet no business can afford to lose.