Enterprises are pouring billions into artificial intelligence, and many are not seeing the return they expected. The reason is rarely the model they chose.

The questions I keep hearing from senior leaders across Australia and New Zealand on AI have evolved rapidly. Twelve months ago, the conversation was about whether to invest in generative AI. Six months ago, it shifted to questioning the ROI of those early pilots. But today, as we witness the rise of agentic AI and what some call the SaaS-pocalypse, the question is more urgent: "How do we move beyond chatbots to AI that actually executes work?"

We are moving through a fundamental shift where software is no longer just a tool for humans to use, but a layer of reasoning that can call APIs, evaluate results, and iterate on tasks. Working for a building data infrastructure company for some of the world’s largest organisations, including 45% of the ASX’s top 20 companies and over 50% of the Fortune 500, I have a clear view on where these initiatives tend to break down. The problem is rarely the agentic AI process itself.

It sits in the data layer underneath, and specifically in whether an organisation can actually search across its data intelligently.

Everyone has been focused on the interface: the chatbot, the copilot, the agent. What tends to get overlooked is the engine that powers it. That engine is search, and for most large organisations right now, it is not able to search their full set of data and provide the context for the AI agents to do what the business is asking it to do.

The data problem hiding in plain sight

Here is a number worth sitting with: unstructured data, the emails, documents, images and SharePoint files spread across a typical organisation, is growing three times faster than structured data. Most AI systems are not searching for it properly.

This matters because when an AI agent produces an answer or takes an action for a customer, an employee, or a decision-maker, it is only as good as the data it can retrieve. If your search infrastructure cannot span your entire data estate, the agent is working with an incomplete picture, leading them to confidently deliver errors in potentially high-stakes situations.

Consider what this looks like for a retail customer who cannot find the product they want on a brand’s website. They do not ring to complain. They close the tab and find it somewhere else. Brand loyalty has become remarkably thin when the next option is seconds away. The same dynamic plays out across financial services, healthcare, and any other sector where customers have come to expect fast, accurate, and relevant responses.

Expectations have shifted, and they are not shifting back

Consumer expectations have moved substantially over the past two years, and many enterprises have been slow to register just how much. Millions of people now use AI assistants as part of their daily routine, and our recent survey showed that 62% of customers expect search to be as smart as ChatGPT. They have grown accustomed to receiving a single, coherent, contextually aware response to almost any question they ask. That experience has become the new floor.

When an enterprise search tool returns no results because a customer misspelled a product name, it is now a meaningful failure by the standard customers use to judge the experience.

In fact,72% say they have abandoned a favourite brand due to poor website search. Modern semantic search understands intent, handles typos and draws on behavioural context to surface the right result. The technology to do this exists and is widely deployed in leading platforms. The gap for most organisations is the data foundation required to support it.

Search has become a competitive lever in a way it simply was not five years ago. The organisations that invest in getting it right will find it increasingly easy to retain customers once they have won their attention. Those who leave it as an afterthought will keep spending on the AI layer above and struggle to understand why the experience is not landing.

Why the ROI keeps disappointing

The executives I speak with who are frustrated by their AI returns tend to share a common story: they added a chatbot to an existing workflow and saw modest productivity gains. Drafting emails. Tidying notes. But they failed to deliver transformational outcomes.

True transformation requires moving from simple data retrieval to also delivering relevance and context. This means delivering the right data, at the right time, in the right structure to allow an AI agent to reason effectively. Technology initiatives anchored with the right foundations can deliver genuine business outcomes; those chasing novelty do not.

The question worth raising in the boardroom

Boards are rightly focused on AI governance and strategy.  But as we move into the agentic AI era, there is another question worth adding to the agenda: “Can our AI tools find the data it needs, across every system it lives in, quickly enough to support the AI investments we are making?”

For most, the honest answer is that the search and data layer has not kept pace. Data is siloed across cloud environments and legacy systems, poorly indexed, and largely inaccessible to the AI agents that need it most.

The businesses that pull ahead in this race won't just be the ones that choose the fastest model. They will be the ones who master the ability to find and surface the full breadth of organisational knowledge at the speed customers expect. By layering intelligent search across your full company data, you move from AI that impresses in a demo to agentic systems that fundamentally transform how your business operates.

Jeremy Pell is Country Manager ANZ at Elastic, a search AI platform powering enterprise search, observability, and security for organisations globally.