Does human oversight improve perceived AI-audit outcomes?

Who answered the survey and what they were asked

The researchers collected 150 online responses over one month from ESG auditors, compliance officers, sustainability managers and AI-governance analysts in private, public and non-governmental organisations. Recruitment used professional networks, audit forums and LinkedIn groups focused on ESG. Participants were chosen purposively and then through snowball referrals because experience with AI-driven ESG auditing is specialised.

That method can reach knowledgeable respondents, but it does not produce a representative sample. People active in professional online groups or referred by peers may be more engaged with AI and governance than the wider profession. The paper's authors are based in Greece, but respondent geography is not reported clearly enough to treat the sample as nationally representative or globally balanced.

Respondents used five-point agreement scales to rate AI automation, AI decision-making, human oversight, audit efficiency, accountability and transparency. The analysis used partial least-squares structural-equation modelling. The achieved sample exceeded a simple ten-times rule but was slightly below the paper's 155-to-160 target from an inverse-square-root calculation for detecting a small effect around 0.20.[1]

What the structural model found

Reported AI automation was positively associated with perceived efficiency, with a standardised path coefficient of 0.351, accountability at 0.288 and transparency at 0.341; all had p-values below 0.001. AI decision-making had smaller positive direct associations with efficiency at 0.203, accountability at 0.248 and transparency at 0.297. The first two had p-values of 0.001 and the third below 0.001.

Both AI variables were also associated with human oversight: 0.494 for automation and 0.440 for decision-making. Oversight in turn was associated with efficiency at 0.426, accountability at 0.436 and transparency at 0.397, all with p-values below 0.001. Significant indirect paths remained alongside direct paths, so the authors described partial mediation.

The indirect association between automation and efficiency through oversight was 0.210; corresponding figures were 0.216 for accountability and 0.196 for transparency. For AI decision-making, they were 0.187, 0.192 and 0.175. These coefficients describe questionnaire constructs in this sample. They are not percentages of improvement and should not become claims that oversight reduced errors or raised productivity by a particular amount.[1]

Why perception and performance must stay separate

A respondent enthusiastic about an organisation's AI programme may rate automation, oversight and outcomes positively at the same time. Organisational culture, senior sponsorship or personal attitude could raise every score. The authors ran a Harman single-factor check for common-method bias, but a statistical diagnostic cannot remove the basic same-source, same-time design problem.

Efficiency is especially easy to overread. The survey captured agreement with statements about process, not timestamps, staff hours, cost per engagement or time to assurance. Accountability and transparency were also reported constructs, not independently reviewed evidence of traceable decisions, successful challenges, disclosed limitations or comprehensible models.

The paper distinguishes process transparency from technical explainability. A well-documented workflow may still rely on an opaque model; an interpretable model can sit inside a process with poor responsibility and escalation. Organisations need both operational records and appropriate model evidence rather than using a favourable survey score as proof of either.[1]

What firms, auditors and regulators can take from it

The defensible implication is to specify oversight before automating an ESG-audit task. A firm can name the person accountable for each model-assisted judgment, define which evidence must be checked, preserve source-to-output traceability and document when a human changed or rejected a recommendation. High-risk assertions, such as emissions baselines or materiality decisions, need stronger review than routine document sorting.

Audit committees should ask for operational measures alongside staff sentiment: elapsed time, rework, exception rates, confirmed misstatements, false alarms, unresolved data gaps and the proportion of model outputs receiving substantive review. Regulators and standard-setters need evidence that oversight is capable of changing a decision, not merely that a human clicked approve.

For workers, meaningful oversight requires time, access to underlying evidence and authority to challenge the system. If workload targets make review nominal, a human-in-the-loop label adds little protection. Training should cover both ESG subject matter and the limits of the AI system, while escalation routes should protect staff who identify unreliable evidence or misleading claims.[1]

Evidence limits, ethics and disclosures

The cross-sectional design prevents causal inference and cannot establish the sequence implied by the mediation model. Purposive and snowball recruitment means coefficients describe this sample and its networks rather than population parameters. The marginally smaller-than-target sample, self-report measurement and unclear geographic distribution add uncertainty. The model-fit SRMR was 0.063, while NFI of 0.795 was below commonly preferred levels.

The authors acknowledged that measuring predictors and outcomes at one time may inflate relationships. They recommended longitudinal work, sector comparisons and explainable-AI variables. Data are available from the corresponding author on request because of confidentiality commitments rather than through an open repository.

The study received no external funding, and the authors declared no conflicts. The university ethics committee waived formal review under its rules because the project had no funding, collected no identifying or sensitive personal information and involved no vulnerable populations. The authors reported obtaining informed consent.[1]

What would change the assessment

Confidence would rise with a preregistered study following audit teams over time and combining surveys with system logs, reviewed files and independently adjudicated outcomes. A stronger comparison could examine similar engagements before and after a defined oversight intervention, or compare teams assigned different protocols while controlling for task difficulty and organisational support.

Useful denominators include total engagements, model-assisted decisions, human overrides, confirmed errors, unresolved exceptions and staff hours. Reporting should separate organisations, sectors and jurisdictions and test whether relationships hold outside networks already interested in AI. Evidence would weaken the interpretation if objective outcomes fail to improve, if oversight becomes ceremonial, or if automation shifts work into hidden review and correction.[1]