Artificial intelligence usually enters an organisation with a familiar promise: work will become faster, more accurate and more efficient. In many cases that promise is kept. Reports are produced sooner, large volumes of information are processed in seconds, and routine cognitive tasks are handed to systems that do not tire or lose concentration.
The difficulty is that efficiency is visible, while the weakening of human judgement is not. Recent longitudinal research on accounting and auditing professionals has described this as the numb efficiency paradox (Yin et al., 2026). Employees may continue to meet their targets while becoming less sensitive to warning signs, ethical concerns and weak evidence. From the organisation’s perspective, the implementation looks successful because productivity holds steady. Beneath the surface, the professional scepticism needed to detect emerging risk may be eroding. Reliable performance requires more than an acceptable answer today. It requires people who will still recognise tomorrow when the system is wrong.
Reliance is not the same as trust
AI adoption is often framed as a question of whether employees trust the technology. I think that framing is too narrow. The more useful question is whether people rely on AI appropriately.
Appropriate reliance does not mean accepting AI output more often, and it does not mean challenging it as often as possible. Someone who rejects good advice without sound reasoning is no better calibrated than someone who accepts poor advice without question. Responsible reliance means knowing when to accept, verify, modify, reject or escalate a recommendation. That judgement depends on the accuracy of the system, the consequences of error, the person’s expertise, the evidence available, the time pressure involved and whether the organisation makes meaningful challenge possible.
For this reason, overreliance should not automatically be read as a personal weakness. It is often produced by the decision environment. An employee may appear overconfident because the system presents its conclusions with unwarranted certainty. Another may fail to challenge because the organisation expects fast decisions, offers no clear escalation route, or makes disagreement hard to defend. A third may remain formally accountable for a decision while having little practical influence over its outcome.
Cognitive integration is a continuing process
The healthier alternative is not to keep AI at arm’s length. It is to integrate it without surrendering the cognitive processes that sound judgement depends on. Effective human and AI interaction appears to follow a repeating cycle. The person monitors their own reasoning, evaluates the AI as an external source of information, and decides how much cognitive responsibility to allocate to the system given the task, the evidence and the likely consequences. They then examine the outcome and adjust their approach next time.
This is considerably more demanding than writing a good prompt. A person may be technically skilled at using AI and still be behaviourally unprepared to work with it responsibly. They may produce impressive results without being able to explain the reasoning, identify weak evidence or do the task when the system is unavailable. The capability that matters is therefore not AI literacy alone. It is the capacity to use AI while retaining ownership of consequential judgement.
Productivity can mask behavioural deterioration
This creates a governance problem. Most organisations monitor what is easiest to measure: speed, output, usage, cost reduction and task completion. These indicators show whether AI is being adopted and whether it delivers operational benefit. They cannot show whether human judgement remains healthy.
A team can become faster while becoming less questioning. Employees can keep delivering while slowly losing their sensitivity to anomalies. The organisation may celebrate adoption while its ability to detect an unusual or ethically troubling situation quietly weakens. Stable output can therefore reflect two very different conditions. In resilient efficiency, people use AI productively while keeping their scepticism, ethical vigilance and ability to intervene. In numb efficiency, output continues but people grow detached from the warning signals that reliable judgement relies on. On a conventional performance dashboard, the two can look almost identical.
Accountability alone is not protection
A common response is to insist that a human remains accountable for the final decision. That sounds reassuring, but accountability without capability and authority can become symbolic. Nobody can provide meaningful oversight without the expertise to evaluate the output, the time to investigate uncertainty, the information to challenge it and the authority to change the resulting action. Accountability pressure may sharpen attention, but it may equally produce anxiety, defensive documentation and routine approval when the person cannot genuinely influence the outcome.
The governance question is therefore not only who signs off. It is whether that person can independently evaluate the recommendation, see what the AI has done, take enough time to examine uncertainty, pause or escalate the process, and change the outcome through disagreement, and whether the organisation will support a justified challenge. Where the answer to these questions is no, the human carries responsibility without exercising control.
The hidden cost of monitoring
There is a further complication. Employees who take oversight seriously may stay cognitively attached to AI-delegated work long after the task has formally been handed over. The system does the work, yet the employee keeps wondering whether it misread the instruction, missed important context or produced an error that will surface later. AI can reduce execution effort while increasing monitoring burden.
This is especially likely when people remain accountable for opaque processes they cannot observe continuously. The organisation gains the benefit of automation while the individual absorbs the uncertainty. Responsible adoption should therefore ask not only whether employees monitor the system, but whether that monitoring is proportionate, sustainable and supported by appropriate controls.
What organisations should measure
To judge whether AI adoption is behaviourally stable, organisations need to look beyond usage and productivity. They should examine whether employees can tell reliable advice from unreliable advice, retain professional scepticism, detect anomalies and explain why they accepted or rejected a recommendation. They should also check whether escalation routes are visible, whether interventions are recorded and whether challenging the system actually changes decisions.
The most informative indicators are likely to include accuracy when AI advice is correct and when it is wrong, the quality of verification, the ability to spot weak or missing evidence, appropriate pausing and escalation, and effective correction rather than mere challenge frequency. Performance when AI support is unavailable matters as well, as do ethical vigilance, professional scepticism and recovery from the cognitive burden of monitoring delegated work. Together these recognise that behavioural readiness is not the same as willingness to use AI.
The real test of responsible adoption
The central question is not whether AI makes people faster. In many settings it already does. The more important question is what happens to human judgement while that efficiency is being created.
A responsible organisation should be able to improve performance without weakening the human capabilities that keep its decisions safe, defensible and adaptable. Its people should be able to use AI productively, recognise uncertainty, challenge poor recommendations and remain accountable without carrying an impossible monitoring burden. AI adoption becomes behaviourally stable when efficiency and judgement improve together. Anything less may look like progress on the dashboard while quietly increasing risk underneath it.
Reference
Yin, S., Hu, W., Tian, Z., & Zhang, Y. (2026). The numb efficiency paradox: AI work pressure, affective numbing, and professional judgment in organizationally digitally mediated work. Frontiers in Psychology, 17, Article 1900940. https://doi.org/10.3389/fpsyg.2026.1900940