An AI system recommends an action. A person reviews the reasoning, considers the context and decides whether to proceed.
That interaction captures one of the central questions facing organizations today: As AI moves from generating answers to taking action, how do we determine how much autonomy to give it?
The new Data and AI Impact Report: The New Economics of Trust provides useful evidence. Based on research involving nearly 2,700 decision-makers across 28 countries, it examines how data quality, model oversight, explainability and other trustworthy AI practices relate to business outcomes.
The report finds a striking relationship. Organizations with high trustworthiness scores are 15 times more likely to report strong or high AI returns than those in the lowest-scoring group.
Those trustworthiness leaders appear better positioned to determine when AI is ready for greater responsibility. They treat autonomy as something the system earns through performance, transparency and governance.
Human oversight is not the same as better judgment
One finding stands out to me: The report finds that overriding AI recommendations is common. The leading reasons are insufficient explanation and lack of context.
Human intervention remains essential, particularly when decisions carry meaningful consequences. But we should not assume that every override improves the outcome. Research in forecasting has shown that expert judgment can add context that a model lacks, but intervention does not improve every outcome. Organizations should understand when those interventions improve results and when they do not.
That suggests a useful distinction: The ability to override is an important safeguard. But organizations should also evaluate the outcomes of those interventions. Understanding when human judgment improves a result, and when it does not, provides important feedback. It may show that the system needs better data, more context or clearer reasoning. This feedback is essential to building systems that combine the strengths of people and AI.
Autonomy raises the standard
Trust declines as AI moves from generating answers to taking action, even as nearly nine in 10 respondents say agents already play some role in decision-making. At the same time, only 17.5% of organizations report operating data infrastructure at the highest maturity level.
The challenge is not that organizations are pursuing more autonomous AI. It's that many are doing so before the necessary data foundations are fully mature. Greater autonomy requires reliable data, traceable lineage and continuous validation so organizations can understand and evaluate AI-driven actions.
The goal should not be maximum autonomy or maximum human intervention. Autonomy is not an all-or-nothing decision but a continuum of responsibility. The most successful organizations treat autonomy as graduated responsibility, expanding authority only when systems consistently demonstrate reliability.
Then the question becomes: What level of autonomy has this system earned? Answering that question requires more than technology. The report’s leaders are distinguished by repeatable behaviors: how they govern, validate, explain and measure AI.
AI earns greater autonomy when its limitations are understood, its decisions can be examined and accountability for outcomes remains clear. Trust, in that sense, is not a feature of technology. It is the result of disciplined governance, measurement and oversight.
