Your AI agent is only as good as the data underneath it.

There is a version of the agentic AI story that the technology industry has been telling for the past two years. It goes roughly like this: powerful AI models now exist, no-code tools make them deployable by anyone, and organisations that move quickly will capture an enormous productivity advantage. Deploy agents. Automate workflows. Win.

The version the data is now telling is more complicated. Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027 — citing high implementation costs, unclear ROI, and inadequate risk management. And that failure rate is not uniformly distributed. It falls heaviest on organisations that moved fast, skipped foundations, and discovered too late that an AI agent running on bad data is not a productivity tool. It is a liability.

01The infrastructure problem nobody wants to talk about

The most consistent finding across independent analyses of agentic AI failures is one that the industry would prefer to minimise, because it is unglamorous and slow to fix. Research finds that the majority of companies lack the foundational infrastructure to build effective agentic AI. Not because they lack access to capable models — they don't. Not because the tools aren't good enough — they largely are. But because the data those agents need to operate on is siloed, unstructured, inaccessible, or simply not trusted.

The solution, Stone argues, is to pair AI agent initiatives with genuine investment in core infrastructure: semantic search capabilities, clean structured data catalogs, proper API integration, and hybrid search functionality. These are not exciting announcements. They do not generate press releases. But they are the reason an agent can actually do what you ask of it.

The car analogy from Lucidworks CEO Mike Sinoway captures it succinctly: if agentic AI is a car and generative AI is the engine, then data is the fuel. Too many organisations, his research finds, are trying to build Formula One racers around go-kart engines — and they might not even have enough gas to fill the tank.

02What Microsoft's AI Red Team found

Microsoft's AI Red Team (AIRT) published a comprehensive taxonomy of failure modes in agentic AI systems — the product of systematic interviews with external practitioners and internal analysis across Microsoft Research, Azure, Security, and the Office of Responsible AI. Their findings are worth studying carefully because they go deeper than "bad data produces bad outputs." They identify specific, catalogued failure modes that emerge when agents are given real autonomy without proper foundations.

The taxonomy distinguishes between novel failure modes unique to agentic systems and existing AI failure modes whose impact is dramatically amplified in an agentic context. Several are directly caused or exacerbated by poor data infrastructure:

The AIRT whitepaper is direct about this: the potential impact of existing AI failure modes is "greatly increased when contextualised in an agentic AI system." A hallucination in a chatbot is an embarrassment. A hallucination in an agent that is autonomously managing approvals, routing documents, or generating invoices is a material operational failure.

03The context problem — why generic AI isn't enough

There is a version of AI adoption that is seductive precisely because it is easy: connect a generic large language model to your business, hand it a few tools, call it an agent. The model is capable. The interface is polished. The demo looks impressive.

But a generic model does not know your business. It does not understand the specific meaning of a term in your contracts, the particular workflow your compliance team follows, the implicit rules encoded in two decades of operational decisions. It knows what the internet knows — which is to say, it knows almost nothing about the specific competitive intelligence locked inside your organisation.

"An agent running on generic knowledge is not your agent. It is a general-purpose tool wearing your logo."

This is the fundamental limitation of deploying AI agents without first building a knowledge foundation grounded in your own verified data. The agent will produce outputs — it will always produce outputs — but those outputs will reflect the model's general training, not your organisation's specific context. In domains where context is everything — insurance, healthcare, legal, finance — that distinction is the difference between a useful tool and a dangerous one.

04The foundation is not a feature — it is a prerequisite

TrustedHub has spent 25 years building data infrastructure. We did not start with AI agents. We started with documents — the physical records that organisations hold but cannot use, the institutional knowledge locked in paper that no model can access because it has never been structured, enriched, or made readable by a machine.

The Trusted MultiVerse was not designed around AI as an afterthought. It was designed from the ground up around the principle that every organisation is sitting on a treasure trove of information — in their archives, their files, their legacy records — that is entirely invisible to AI until it has been properly processed. DataVerse converts that information into high-dimensional vector embeddings. KnowledgeVerse provides a governed, queryable layer that grounds agent responses in verified organisational knowledge. ActionVerse then builds agentic workflows on top of that foundation.

The sequence matters enormously. It is not possible to run effective agentic workflows on models that do not understand the context of the business. An agent that cannot answer "what does our underwriting policy say about this specific case?" because that policy was never digitised, never structured, never made queryable — that agent is not enhancing your business's competitiveness. It is performing theatre. Every other organisation deploying the same generic model has the same capability. The differentiator is never the model. It is always the data underneath it.

05What this means for organisations building with AI now

The organisations that will look back on 2026 and 2027 as the years they built a durable AI advantage are not the ones that deployed the most agents the fastest. They are the ones that asked a harder question first: what does our AI actually know about us?

If the answer is "whatever the base model was trained on," then no amount of agentic capability will change the fundamental output quality. The agents will be fast. They will be confident. They will produce outputs at scale. And a meaningful fraction of those outputs will be wrong in ways that are expensive to catch and correct.

If the answer is "our agents are grounded in our own verified, structured, governed data — digitised, enriched, and made queryable through a certified process" — then the AI is not just a tool. It is an institutional intelligence that compounds in value the more your organisation adds to it. Every new document processed, every new record digitised, every new workflow built on top of that foundation makes every agent smarter, more contextual, and more reliable.

That is the difference between AI that depletes your organisation — through failed projects, wasted costs, and eroded trust — and AI that builds it. The foundation is not a feature you add later. It is the prerequisite for everything that comes after.