Why your AI agents keep failing – and it’s not AI
Most companies experimenting now with AI agents never manage to scale them. Fewer than one in ten do. The problem, almost always, is what’s underneath: the data.
The gap everyone is trying not to talk about
So, there’s a version of the AI story that sounds like this: companies deploy AI agents, the agents automate complex tasks, productivity soars, and everyone wins. That version exists. It’s just rarer than the headlines suggest.
In other words, according to a McKinsey study published in April 2026, roughly two thirds of enterprises worldwide have run experiments with AI agents. Fewer than ten percent have managed to scale them into something that delivers real, measurable value. And the failure isn’t usually the AI itself – it’s what the AI is running on.
“Eight in ten companies say fragmented, siloed data is what stops them from scaling AI agents.”
What good foundations actually look like
Actually, McKinsey’s research identifies four steps that separate organisations managing to scale agentic AI from those who get stuck in pilot purgatory. They’re worth understanding as a sequence – each one builds on the last.
- Find the right workflows to automate. Not everything benefits from an AI agent. The organisations getting results start by identifying a small number of end-to-end processes where autonomous decision-making could genuinely change outcomes – and map exactly what data those processes would need.
- Clean up the data architecture, layer by layer. This doesn’t mean rebuilding everything from scratch, it actually means modernising how data flows, connects, and becomes usable – progressively. Thus, data from different systems (CRM, supply chain, finance) needs to speak the same language.
- Move from cleanup sprints to continuous quality management. One of the most common failure modes is treating data quality as a periodic project. In an agentic environment, we should be able to monitor data quality in real time, with automated checks.
- Build governance for what agents are allowed to do. As agents gain autonomy, the rules governing their behaviour become the primary mechanism of control. Clear policies – defining what data an agent can access need to be automated and embedded. Human roles shift from doing the work to supervising and orchestrating agent-driven workflows.
Not to mention that the thread running through all four steps is the same: we need to treat data as infrastructure.
Where Zeren fits into this picture
In fact, this is the layer of work Zeren’s consultants operate in. We have consultants like Sânziana for whom the data architecture makes the AI models reliable.
Sânziana, one of Zeren’s senior data architects, is currently leading a Business Intelligence engagement for a major manufacturing company in the Nordics. Her framing of the problem captures it well:
“Without properly modelled data, AI cannot produce good results. We are among the fortunate ones for whom the rise of AI brings more work, not less.”
What Zeren Software does
Nevertheless, Zeren connects specialist data professionals – data architects, data engineers, BI consultants, AI engineers – with complex international projects. Our consultants work across industries building the data foundations that make AI actually usable in production environments.
The question worth asking now
So, if your organisation is planning an AI initiative, the most useful diagnostic relates to the data underneath the AI model. Is it connected or consistent? Is it governed? Do your agents have access to what they need, and only what they need? All these are questions to consider.
“In the agentic age, data foundations are becoming the primary source of competitive differentiation.”
Undoubtedly, for the companies already operating at scale, this prediction is already a sheet fact.
Curious about how your data infrastructure stacks up? Get in touch with Zeren.
