New research shows that only 38% of employers surveyed involve employees in discussions about AI. Only 40% keep checking what AI is doing to wellbeing, job satisfaction and skills development once systems are in place. These are organisations certified for excellence in people practices, so the numbers across the wider market are likely lower. “Organisations have moved fast on the technology and slowly on the conversation. Most employees have never been asked how AI is affecting their work, and that gap shows up later in ways that are much harder to fix,” says Sandra Botha, Lead HR Auditor at the Top Employers Institute.
She continues, “Nobody puts up their hand in a town hall and says they don’t trust the system. What we see instead is hesitation. People double-check the output or find ways to work around the system altogether, and the value the business expected from AI slowly erodes.”
From the boardroom, the doubt is easy to miss. UK Government research found that only 21% of workers believe AI improves their performance. And when that doubt has nowhere to go, it takes other routes. Some employees disengage from AI tools and revert to old ways of working. Others adopt AI agents on their own terms, outside official channels and beyond the oversight organisations rely on to manage financial, ethical and security risk.
Fairness feeds the problem. Where AI does lift performance, employees rarely see any of the benefit themselves, whether as shorter working hours, recognition or pay. Botha cautions HR leaders against assuming that the curiosity many employees brought to the first wave of AI tools will carry over into the next, particularly with early reports of AI-related burnout starting to surface.
Organisations that have laid the groundwork behave differently. Top Employers with an established ethical AI framework are more than twice as likely to evaluate how human and AI collaboration is working in practice, at 51% compared with 23% of their peers. Employee involvement in AI ethics discussions follows the same pattern, at 46% against 25%. Both practices show up more often in organisations with stronger engagement, retention and internal promotion outcomes.
“Adoption is not automatic. It has to be designed,” explained Botha, speaking at the Institute’s recent People Summit. “The organisations getting this right involve their employees early and keep the dialogue open. It is not a one-off communication at launch. It is an ongoing conversation about what AI is doing, where it is helping, and where there are concerns that need to be addressed.”
However, it’s not all doom and gloom; South African organisations are also leading in this space. Vodacom Group, for example, shows what that involvement can look like in practice. Rather than confining AI to a central team of specialists and asking everyone else to trust it, the company has trained employees across functions, including HR, finance, legal and compliance, to build and deploy their own AI solutions using no-code and low-code tools, within set guardrails and approval frameworks.
Employees get a direct hand in how AI enters their own work, and the results have followed: a community of 531 citizen developers has delivered 658 automated and data-driven processes and generated R18 million in quantified financial benefits. The model is now being extended into an Agentic AI Citizen Programme, and the company’s stated ambition is an organisation where every person has “the tools, the training, and the trust to shape how AI works in their world.”
The window for getting this right is narrowing. Salesforce research found that just 15% of HR teams had begun deploying agentic AI early in 2025; by the second half of the year, McKinsey put the share of organisations experimenting with it at around 40%. And unlike earlier tools, these systems do not sit alongside the work. They run entire workflows, which means every employee who quietly opts out now carries a far higher cost.
“The organisations that give employees room to raise their doubts early, while there is still time to respond, are the ones that will see AI deliver what the business case promised,” concludes Botha. “That conversation is much easier to have now than after the systems are already embedded.”
