By Marko Matejcic, Product Leader, Adverity
The rush to invest in AI agents is palpable, but as the hype dust settles, the reality becomes much more sobering. While the industry is quick to spend on automation, there is a lingering disconnect between the ambition of these supposedly intelligent systems and the quality of the data supporting them.
AI can deliver some quick efficiency wins but it can’t go any further if there are cracks in your data foundations. Investing in AI without addressing data hygiene is not a shortcut; it is a way to scale existing errors faster than ever before.
The silver bullet is an illusion
The most common misconception in the C-suite is that AI adoption is a matter of bolting new tools onto an existing stack. We assume more tools equate to more efficiency, but in practice, many AI tools are effectively islands that don’t talk to each other. Instead of breaking down silos, we are inadvertently building new ones that are even harder to manage because the insights are too difficult to understand – unless you have machine learning scientists on your team.
When every tool in the stack introduces its own interpretation of your data, teams spend more time aligning disparate versions of the truth than actually acting on insights. This doesn’t democratise data; it increases dependency on specialists and slows down the very processes we intended to accelerate.
A step beyond generative: the move to agentic AI
We are moving beyond simple Generative AI toward a future of interconnected AI agents. However, the leap from a chatbot to an agentic system requires something traditional integrations cannot provide: Context.
Traditional APIs are a known quantity when it comes to moving data from one system to another. However, what they don’t do is preserve context. They respond to individual requests in isolation without remembering what came before. These capabilities work fine for fixed workflows and don’t need to be touched from an “if it ain’t broke, don’t fix it” perspective. However, this limitation creates roadblocks when you introduce natural language interfaces that need to preserve conversation history or agentic AI that needs to reason across multiple steps.
Model Context Protocols (MCPs) provide that continuity and shared understanding between user and machine. It allows AI systems to retain the context and parameters of every query so that follow-up questions can refer back to prior interactions instead of treating each like a blank slate. It reduces repetition and errors for the user, who also benefits from interactions that feel more natural. MCPs are what puts the intelligence in AI.
MCPs also serve as a sort of universal translator between AI systems that often have completely distinct logic from one another. It means a media buying agent can talk to an audience intelligence agent without getting their wires crossed, providing the connective tissue of agentic automation at scale.
These are not without their own particular brand of trade-offs, however. For advanced users of AI models, MCPs can consume significant portions of the platform’s context window (input tokens). As a result, some more sophisticated iterations still rely on direct API integrations in parallel, particularly where efficiency and tighter control over tokens are crucial. As we move towards production-scale agentic systems, these trade-offs are the tightrope every business must walk.
The composability mandate
It’s easy to see the appeal of end-to-end systems. Who doesn’t want the simplicity of a single piece of software handling a suite of functions? However, the reality is that such monolithic systems simply kick their complexity down the road to IT, data, and security teams who then have to try and keep them up-to-date to suit business needs. End-to-end systems are notoriously rigid and resistant to change; their code is a tangle of interdependencies that mean a tweak on one end can have unexpected knock-on effects on the other.
For marketers, this can be particularly frustrating. They’re the closest to changing customer behaviour, but they’re often dependent on systems that can’t keep up, and may not understand why beleaguered tech teams can’t green-light the changes they need. This is why I’m a massive advocate for composable technology, where modular components stick together Lego-style and can be easily swapped or tweaked without breaking the whole system.
The verdict? It’s about foundations over features
If you want to see genuine ROI from agentic AI, you have to start with the unglamorous fundamentals.
If the input is rubbish, AI will simply produce more rubbish, even faster. Moreover, agentic AI should give marketers greater access to insights and the autonomy to action them. And finally, everything – from MMM analysis to real-time campaign tweaks – starts with clean, unified, and democratised data.
Technology alone does not create value. Only a solid data foundation can turn AI from a costly experiment into a genuine growth engine.
That being said, no one can predict what the ideal AI stack will look like in a few years, or even by the end of this one. Building something that can evolve in the face of this unpredictability is what will future-proof an organisation.
Focus on building a stack that is adaptable rather than exhaustive. It’s far more sustainable to pursue a modular, composable setup where individual components can be updated or replaced without the whole stack toppling. Technology is moving especially fast right now, and you don’t want to be left with an outdated system because you put all your eggs in one basket.
