There is a number that should change how any business leader thinks about their AI plans. According to a 2025 MIT study, 95% of enterprise AI pilots delivered zero measurable impact on profit and loss. Not low returns. Zero. The technology worked. The organisations were not ready to use it.
The gap between a successful AI pilot and a working AI deployment is almost never a technology problem. It is a data and governance problem. And both of those cost far more than most budgets account for.
Start with data. Most businesses assume that because they have data, it is usable. The reality, as shown in research from S&P Global, is that 42% of companies abandoned the majority of their AI projects in 2025 before reaching production, up from 17% the year before. The most common reason cited was not that the AI model was wrong. It was that the data feeding it was unreliable, poorly structured, or coming from systems that were never designed to talk to each other.
Data preparation alone can consume 60% to 80% of an AI project's timeline and between 25% and 40% of its total budget. That is before a single business result is produced. Most of that work is invisible in a business case: cleaning records that accumulated inconsistencies over years, standardising definitions that different departments apply differently, and deciding which historical data is trustworthy enough to act on.
Then there is the governance problem, which is separate and ongoing. Traditional oversight in most companies runs on quarterly cycles. A review committee meets, looks at what happened, and reports back. AI does not operate on that rhythm. A system embedded in a live business process, say, pricing decisions, customer outreach, or claims handling, can make thousands of decisions between any two meetings. By the time a problem surfaces in a report, significant damage may already have been done.
A 2025 survey of nearly 1,000 senior executives found that companies with real-time AI monitoring were 34% more likely to report revenue growth and 65% more likely to report cost savings than those without it. In the same survey, 99% of organisations reported some financial loss linked to AI risks, and 64% reported losses exceeding one million dollars. Governance is not a theoretical protection. It is a direct driver of financial outcomes.
The failure pattern almost always follows the same shape. A company deploys AI. The oversight model is designed for approvals before launch, not monitoring after. The AI begins operating in a live environment. Small errors compound. No one notices until the problem is large enough to cause visible damage. By then, the cost of correction is substantially higher than building oversight in from day one would have been.
One real example: in 2025, Deloitte Australia had to issue a partial refund to the Australian government after an AI-assisted report was found to contain fabricated references and a non-existent court judgement. The contract was worth roughly $440,000. The failure was not a bad AI model. It was a review process that did not catch what the model produced.
AI also exposes organisational problems that already existed. When a sales team wants to automate customer outreach but the legal team has privacy concerns, and the customer success team worries about relationship damage, those are not new conflicts. They existed before. Slower, manual processes just kept them hidden. AI forces organisations to resolve cross-departmental disagreements that were previously deferred indefinitely.
The companies that are seeing real returns from AI are not the ones with the most advanced models. They are the ones that invested in data quality, cross-functional oversight, and continuous monitoring before scaling. Those are unglamorous, time-consuming foundations. They are also what separates a working AI deployment from an expensive experiment.