Why AI ROI ultimately hinges on operational data infrastructure

For all the discussion around artificial intelligence, most organisations are facing a much simpler challenge in that they struggle to get consistent answers from their own data. 

This is not because they lack information. Finance, procurement, logistics and sustainability teams all collect vast amounts of it, often through sophisticated systems built to support specific operational needs. The problem is that, over time, different systems, processes and reporting requirements create different definitions for the same business concepts. An “order” may be defined one way in an ERP system, another in a warehouse platform and differently again in a customer-facing portal. Lead time may be measured from order confirmation by one team and from goods receipt by another, while sustainability data may be categorised by supplier group in one report and by transport route in another. 

Individually, these decisions make sense. Collectively, however, they create a challenge that many organisations only discover when they begin scaling AI initiatives and trying to connect information across multiple functions. 

The limitation AI cannot solve 

AI can analyse data, identify patterns and generate insights. What it cannot do is determine which of several conflicting business definitions is the correct one. If two datasets use different definitions of “late delivery”, an AI model may generate two answers that are both technically accurate and completely inconsistent with one another. Better prompts will not solve the issue, and more advanced models will not solve it either. 

The problem is not intelligence. It is interpretation. Many organisations underestimate how often AI projects depend on shared business meaning. When definitions vary across departments, every cross-functional use case becomes harder to scale, regardless of how capable the underlying technology may be. In those circumstances, AI is not amplifying clarity; it is amplifying ambiguity. 

Why pilots often succeed while programmes struggle 

This is one reason AI pilots frequently deliver promising results. Most pilots operate within a relatively contained environment, drawing from a single system, a limited dataset or a well-defined business process. Under those conditions, AI can produce meaningful outcomes quickly because the assumptions and definitions underpinning the data are largely consistent. 

The challenge emerges when organisations attempt to expand beyond that initial use case. A procurement model may need data from finance, while a sustainability initiative may require information from logistics and supplier management systems. Suddenly, teams are spending time reconciling metrics, debating definitions and validating extracts before any meaningful analysis can begin. What appears to be a technology problem is often a data alignment problem, and organisations can find themselves paying a hidden cost not in software licences or infrastructure, but in the hours spent reconciling information before decisions can be made. 

The groundwork that creates lasting value 

The most effective starting point is often far less ambitious than many organisations expect. Rather than launching a broad transformation programme, focus on a decision that is actively being considered today. For example, should demand be shifted to a lower-cost supplier? Should inventory levels be adjusted? Should a logistics route be changed? 

Before running scenarios, bring together the teams involved and agree on a small number of critical definitions. What constitutes an order? What qualifies as late? How is lead time measured? From there, establish a common identifier that links information across systems and document how key metrics are calculated. This work is not particularly glamorous, and it may take several workshops and require compromises between teams. Yet it often delivers greater long-term value than another technology deployment because it removes friction from every future initiative that follows. 

Better decisions, not just better reports 

When shared definitions exist, conversations change. Procurement can evaluate supplier costs alongside logistics performance, finance can access operational insights without waiting for separate reporting cycles, and sustainability teams can assess emissions implications before decisions are made rather than reporting on them afterwards. 

Disagreements do not disappear, nor should they. Healthy organisations debate decisions and challenge assumptions. The difference is that people are debating the implications of a decision rather than questioning whether the underlying numbers can be trusted. That shift alone can dramatically improve the speed and quality of decision-making. 

What successful AI programmes have in common 

The strongest AI programmes I have seen often look surprisingly modest on paper. They usually begin with a small set of agreed definitions, consistent identifiers across systems and a clear understanding of how important metrics are produced. These foundations rarely attract attention, but they make everything else easier because they remove uncertainty before it has a chance to spread. 

One useful test is to ask whether an AI project can explain, in simple terms, how key definitions were reconciled across the data sources it relies upon. If it cannot, that should raise concerns. The issue is not whether the model performed well in a demonstration. The issue is whether the organisation has reduced the ambiguity that will inevitably surface in future projects. Without that progress, the same reconciliation work simply reappears in different forms. 

The next phase of AI maturity 

Over the next year, the gap between organisations generating meaningful AI returns and those struggling to move beyond pilots is likely to become more apparent. The difference will not necessarily be the sophistication of the models they deploy. It will be the quality and consistency of the operational data beneath them. 

The organisations pulling ahead will spend less time debating what the data means and more time acting on what it tells them. Their questions will evolve from “Can AI summarise this?” to “What happens to cost, service levels and emissions if we act now?” Those are fundamentally different conversations, and they only become possible when leaders have confidence in the information supporting them. 

Building the operational foundations required to support those conversations is not the most visible part of an AI strategy. It is, however, one of the most important. Because in the end, sustainable AI returns depend less on the intelligence of the model and more on the reliability of the data it is asked to understand.