The applause for AI pilots has stopped. Now boards want the money

For three years, an AI pilot was the easiest win in the building. Stand up a proof of concept, demo it to the executive team, collect the applause. That era is over. In 2026 the demos have stopped impressing anyone, and the question in the boardroom has hardened: where is the return?

The numbers behind the mood shift are stark. An MIT research group studied hundreds of enterprise generative AI deployments and found that around 95 percent of pilots delivered no measurable impact on the profit and loss statement. Only a small minority were extracting real value. Gartner, meanwhile, predicts that more than 40 percent of agentic AI projects will be cancelled by the end of 2027, pointing to escalating costs, unclear business value and weak risk controls. It also warns of widespread agent washing, where vendors rebrand ordinary chatbots as autonomous agents.

Why pilots stall

The failure pattern is remarkably consistent, and it is rarely about the models. Pilots succeed in controlled conditions, then hit the mess of real operations: documents that vary in structure, exceptions that multiply, people who behave unpredictably. Costs bite too. Token-based pricing makes agentic systems expensive in ways finance teams struggle to forecast. And governance questions that were waved away in the pilot phase, such as who is accountable when an autonomous system makes a bad call, become blockers the moment production is on the table.

What boards now demand

Directors have noticed. A recent Harris Poll survey of 600 CIOs across eight countries found that 98 percent report growing board pressure to show measurable return on AI investment, and 71 percent believe their AI budgets face cuts or freezes if targets are missed by mid-2026. Most expect their own pay to be tied to AI outcomes, and nearly three quarters regret a major AI vendor decision made in the past 18 months. Board-level commentary points the same way: directors increasingly want full visibility of the AI running across the enterprise, treating unknown and unmanaged AI as a fiduciary risk, and they want trust demonstrated through explainability and auditability rather than assurances.

The New Zealand picture

At home, adoption is broad but shallow. A Deloitte study for 2degrees published this year found 82 percent of New Zealand businesses now use some form of AI, though most early-stage users lean on features baked into existing software rather than anything deliberate. The upside for committed adopters looks real: the study estimated SMEs using AI earned around $400,000 more in FY25 than comparable non-adopters, and large adopters roughly $59 million more. AI already accounts for 29 percent of technology spending here and is expected to keep climbing. All of this against a backdrop of falling national productivity, which is precisely why the Government’s first AI strategy, released in July 2025, is unapologetically adoption-first: voluntary guidance rather than heavy regulation, aimed squarely at getting the private sector to invest with confidence.

The gap is the opportunity

Here is the encouraging part. The MIT research found the winners were not the organisations with the flashiest models. They bought specialist tools rather than building generic ones, embedded AI deep in real workflows, and chased unglamorous back-office value instead of shiny front-of-house demos. Purchased solutions succeeded at roughly twice the rate of internal builds. None of that requires frontier-lab budgets. It requires the discipline New Zealand tech leaders already apply to every other class of investment: a named owner, a real workflow, and a number the CFO recognises.

The pilots were the easy bit. The organisations that treat 2026 as the year of proof, not promise, will be the ones still holding budget in 2027.