In two-thirds of the organisations surveyed, AI can update core business records without human approval. If it makes a mistake, who answers for it: the employee expected to validate the result, the manager overseeing the process or the executive who approved the system?
The survey does not ask. It finds that 66 per cent of leaders say their organisations permit AI to read and write to core systems of record without human approval, while only 33 per cent report fully implemented policies specifying which decisions AI cannot make autonomously. These are separate measures, not proof that the same organisations lack safeguards.
The figures come from Kyndryl’s 2026 People Readiness Report, a commercially commissioned survey conducted by Edelman Intelligence among 1,100 business and technology leaders in eight countries between 24 March and 30 April 2026. Kyndryl provides enterprise technology and AI consulting services. Half the respondents were C-suite executives; India, Japan and the US contributed 200 each. The results describe leaders’ assessments, not independently verified organisational practices. The published methodology does not explain respondent recruitment or country weighting.
Authority ahead of confidence
Eighty-one per cent expect autonomous AI agents to make decisions with material business impact within 12 months. Yet only 25 per cent completely trust AI systems that decide or act without human oversight; for agents interacting with customers without employee supervision, the figure is 24 per cent.
These answers do not establish that autonomous decisions are unsafe. They do reveal a tension between the authority organisations are granting AI and the confidence leaders express in its unsupervised actions. When a system changes a record or affects a customer, someone must be able to recognise an error and intervene.
Rules lag behind
Only 23 per cent of leaders consider their governance and compliance capabilities ready to support AI. Forty-one per cent are completely confident in their organisations’ AI safeguards. Beyond the 33 per cent with fully implemented decision-limit policies, just 27 per cent report a central registry and monitoring system covering all AI systems in their environment.
A policy defines where automated authority ends; a registry identifies the systems operating; monitoring helps track their actions. None automatically settles who is responsible for a harmful decision. Nor do the survey’s separate questions establish how many organisations allow AI to act without safeguards.
The readiness expectations point in the same direction. Fifty per cent anticipate complete readiness in technology infrastructure by the end of 2026, compared with 41 per cent for leadership decision-making and accountability structures and 35 per cent for governance, risk and compliance.

Responsibility without preparation
Forty-seven per cent of leaders expect human accountability to remain but shift from execution towards oversight and validation. Another 32 per cent expect it to increase as employees validate AI-driven decisions; 13 per cent expect it to decrease. Thus 79 per cent foresee accountability remaining with people or growing, even as machines undertake more decisions.
But oversight is not control. An employee asked to validate a system’s decision needs enough information to judge it, authority to challenge it and a clear route to intervene. The survey does not establish whether those conditions exist.
Preparation is uneven. Only 29 per cent report fully implemented development of AI-related skills, including oversight and output evaluation. Thirty-three per cent have fully implemented training in human-AI collaboration, and 22 per cent have redesigned workflows to make AI a collaborative partner.
Those below the C-suite report greater obstacles: 55 per cent identify skills or talent gaps as a major challenge, against 43 per cent of executives. On redesigning roles and workflows for human-AI collaboration, the figures are 48 and 42 per cent respectively.

Who does the overseeing?
The division of work is becoming less clear even in accountability-related activities. Asked where organisational value will primarily come from over the next three years, only 41 per cent expect people to remain the primary source in oversight and accountability for outcomes, compared with 51 per cent in complex decision-making. Thirty per cent expect people and AI to contribute equally to oversight and accountability; 28 per cent expect AI to contribute more.
This measures expectations about value, not who bears legal or managerial responsibility. AI may increasingly perform checks while a person remains answerable for their consequences.
Four in five leaders expect human accountability to remain or increase. Yet fewer than half expect their leadership decision-making and accountability structures to be completely ready this year. Organisations are expanding AI’s authority faster than they expect to be ready to govern it.


