An AI-supported employee-driven performance management system: a conceptual model

1. Introduction

In organizations with numerous measurable activities, information about performance often passes through managerial layers before reaching the person responsible for the task. This arrangement may delay correction and position employees as recipients of managerial interpretations rather than active users of performance information. The relevant question is not whether AI can produce more dashboards, but who receives its initial interpretation and what rights accompany it.

Research on AI-assisted human resource management identifies opportunities in assessment, analytics, personalized development, and decision support, while also documenting risks of opacity, bias, surveillance, and resistance (Malik et al., 2023; Meijerink & Bondarouk, 2023). An employee-facing system is not automatically empowering: its effects depend on design, governance, and actual opportunities to contest recommendations.

This paper proposes shifting the first analytical message toward the employee. The system does not replace a manager, assign blame, or make disciplinary determinations. It reports an observed deviation, offers provisional explanations, and suggests an appropriate next action. A human manager retains responsibility for consequential decisions and for situations requiring judgment, support, or resolution of disputed evidence.

2. Theoretical foundations and research gap

2.1. Performance management and feedback

Performance management extends beyond periodic appraisal to goal setting, monitoring, feedback, development, and employee reactions to evaluation (DeNisi & Murphy, 2017; Aguinis, 2019). Feedback does not improve performance automatically: its effect depends partly on where it directs attention and whether it enables useful task-level adjustment (Kluger & DeNisi, 1996). Self-regulation theory also emphasizes discrepancies between current and desired states as triggers for behavioral adaptation (Lord et al., 2010).

2.2. AI-assisted HRM and algorithmic management

The literature describes both the enabling and controlling consequences of algorithmic management. Algorithms may improve access to information and coordination while also redistributing authority and intensifying monitoring (Kellogg et al., 2020; Noponen et al., 2024). AI-assisted HRM research highlights strategic opportunities alongside fairness and implementation challenges (Malik et al., 2023; Pereira et al., 2023). The relationship between algorithmic management, autonomy, and performance is conditional rather than universally positive (Meijerink & Bondarouk, 2023; Lu, 2026).

Recent work has investigated configurations of algorithmic management and employees' experiences of AI-based appraisal (Chen et al., 2026; Pan et al., 2026). These contributions caution against treating technical availability as evidence of organizational benefit. In particular, employee-facing information should not be conflated with an employee's effective ability to act on that information.

2.3. Specific research gap

The general concepts of employee access to performance information, algorithmic feedback, autonomy, and human oversight are not new. The narrower design problem addressed here is how to combine contextual comparison, uncertainty-aware explanation, employee choice, and an explicit escalation route into one sequence of organizational decisions. The claimed contribution is an integrated design proposition, not the invention of its individual components or proof that this configuration outperforms alternatives.

Kellogg et al. (2020) foreground contested organizational control, while Meijerink and Bondarouk (2023) examine the duality of algorithmic management. Crucially, Bernstein and Li (2025) experimentally studied giving front-line employees direct access to performance data previously mediated by supervisors. Their findings demonstrate that employee-facing access is not, by itself, a new contribution. The present article therefore proposes a narrower, unvalidated design synthesis: context-sensitive comparisons, provisional AI explanations, employee discretion, reassessment, and specified human escalation. Whether this configuration provides incremental benefits remains an empirical question.

3. Research questions

RQ1. Under what conditions can AI shift initial performance diagnosis from the manager to the employee without creating more intensive surveillance?

RQ2. Which diagnostic functions can be supported algorithmically, and which judgments must remain with human managers?

RQ3. Which design and governance requirements are needed to protect autonomy, explainability, contestability, and human override?

4. Conceptual research approach

This paper is a conceptual design study rather than an empirical evaluation or a systematic literature review. Selected works on performance management, feedback, self-regulation, AI-assisted HRM, algorithmic management, and conceptual research methods inform the proposed relationships (Jaakkola, 2020; Snyder, 2019). No reproducible database search, exhaustive screening procedure, or PRISMA flow is claimed.

Three levels of claims are kept distinct: findings or theories reported in prior literature; the present article's design choices derived from comparison of those perspectives; and testable predictions that remain unverified. The synthesis is organized around five practical questions: who first receives the signal; how the expected result is defined; how interpretations may be contested; which action is available to the employee; and when managerial intervention becomes necessary.

5. Proposed employee-driven architecture

The architecture consists of seven linked stages: (1) task-relevant KPI and context data; (2) a contextual performance benchmark; (3) detection of a performance gap; (4) alternative hypotheses concerning the gap; (5) actionable employee-facing feedback; (6) employee choice and action; and (7) reassessment with human escalation when necessary. This is a proposed workflow, not an implemented or validated AI product.

Data and context → contextual benchmark → performance gap → provisional hypotheses → actionable feedback → employee action → reassessment / human escalation.

5.1. Contextual benchmarking

A uniform target or departmental average may be convenient for reporting but misleading for individual diagnosis. Employees can hold the same position while differing in tenure, portfolio composition, client mix, territory, or seasonality. The proposed benchmark therefore refers to an expected range under comparable working conditions, not an unadjusted organization-wide average.

Expected performance = f(role, tenure, portfolio, client mix, territory, seasonality, historical performance).

This expression is a conceptual list of potential predictors, not a fitted statistical equation. Implementation would require an explicit observation unit and prediction horizon, appropriate feature selection, held-out validation, uncertainty estimates, and fairness checks. Historical performance data can reproduce previous structural inequalities. A contextual benchmark must therefore not be assumed fair merely because it contains more variables.

5.2. Performance gaps and competing hypotheses

An observed gap should initiate inquiry rather than an automatic causal diagnosis. Weak sales of one product alongside normal results on other measures could reflect product knowledge, client demand, portfolio composition, incentives, or temporary operational constraints. A system should distinguish the observed fact from tentative interpretations, state the data used and uncertainty involved, and invite the employee to supply missing context.

5.3. Employee-first actionable feedback

The value of feedback lies in whether it supports a meaningful next step, not in the quantity of displayed metrics. Employee-facing messages should separate the result, the comparison basis, possible explanations, and a limited set of feasible actions. Employees should be able to select, modify, or reject a proposed action without automatic punitive consequences. This preserves the distinction between developmental assistance and algorithmic command.

5.4. Human escalation

Managerial involvement should not depend solely on how long a KPI remains below target. Immediate escalation may be necessary when underlying data are disputed, uncertainty is high, the employee contests the recommendation, or the decision has material legal, financial, safety, or employment consequences. Additional triggers may include repeated unsuccessful adjustments. Thresholds and response times should be set locally and validated before deployment, rather than invented as universal rules.

6. Illustrative retail-banking scenario

Consider a hypothetical employee in a retail-banking unit. This example is invented to demonstrate the sequence and is neither a field experiment nor a case study. It provides no evidence of effectiveness.

The system flags the negative deviation in credit cards but does not infer its cause. It could offer alternative explanations, request the employee's account of the context, and suggest product-specific training or a review of customer conversations. After a predefined period, the indicator would be reassessed. A disputed benchmark or failure to improve would prompt human review. The employee retains a meaningful right to challenge both the data and the proposed interpretation.

7. Testable propositions

P1. Under otherwise comparable conditions, developmental employee-first feedback will be associated with higher perceived autonomy than manager-first feedback only when employees can meaningfully select or reject suggested actions; the opposite effect remains possible.

P2. A properly validated contextual benchmark will reduce erroneous employee classification relative to a single unadjusted average, evaluated against independently defined reference labels and with group-level bias assessment.

P3. With equivalent access to underlying performance data, task-focused actionable feedback will reduce the time to initiation of an appropriate corrective action relative to KPI-only displays. Inappropriate actions and downstream outcomes must be measured separately.

P4. An accessible and effective contestation process with human review will be positively associated with perceived procedural fairness and feedback acceptance. The effect depends on actual review quality, not merely on the presence of a button.

P5. An employee-first architecture with explicit escalation criteria will reduce managers' time spent on initial diagnosis relative to a manager-first process without degrading decision quality, once the costs of false signals and escalation are included.

8. Governance, ethics, and boundaries

Explainability requires employees to understand the indicators and contextual assumptions behind a benchmark and the reasons for a suggested intervention. Contestability requires an accessible route to correct data, supply omitted context, and request human reconsideration. Neither principle is satisfied by a merely decorative interface.

A developmental design should minimize surveillance and collect information directly relevant to the task and the outcome. Accumulating the maximum possible behavioral data risks turning support into control. AI-generated developmental signals must not automatically become disciplinary evidence without appropriate human review. These concerns align with the literature on algorithmic control, autonomy, and perceived fairness (Kellogg et al., 2020; Noponen et al., 2024; Pan et al., 2026).

Human oversight must include genuine authority to reject algorithmic suggestions, not only responsibility for approving predetermined outcomes. The proposed system must also address data access, retention, privacy, bias audits, and the organizational incentives that may encourage misuse.

9. Theoretical and practical contribution

The theoretical contribution is a distinction among three forms of authority: access to performance data, authority to interpret that data, and authority to impose consequences. These are often conflated in discussions of management automation. The proposed architecture assigns the first opportunity for developmental interpretation to the employee while reserving consequential decisions and disputed cases for accountable human review.

Practically, the article offers design logic for organizations considering AI in performance management. Rather than asking only which managerial decisions can be automated, it asks which analyses can be placed directly in employees' hands while preserving meaningful interpretation, choice, and appeal. The resulting proposal remains a hypothesis requiring comparison with existing systems.

10. Limitations and future research

This study has no empirical dataset, implementation, validated predictive model, or exhaustive systematic review. It does not establish improvements in KPI outcomes, autonomy, trust, fairness, or managerial workload. The illustrative banking figures are synthetic. The extent to which the proposed combination is distinct from existing commercial and academic systems requires additional investigation.

A future field experiment could compare three groups: (a) manager-first feedback; (b) direct access to the same KPI data without AI recommendations; and (c) direct access to contextual AI recommendations with a defined human escalation path. Primary outcomes should be preregistered and include KPI changes, diagnostic error, time to appropriate action, managerial time, and decision quality. Secondary outcomes should include perceived autonomy, trust, fairness, appeal frequency, and adverse events.

The research protocol should specify randomization or allocation rules, sample size, exclusions, follow-up horizon, and statistical analysis. A comparison limited to manager-first versus employee-first would not isolate the incremental contribution of AI. Future studies should also examine benchmark bias, missing data, variation across roles and industries, and cross-national institutional differences.

11. Conclusion

AI-based performance analytics need not imply more intensive managerial control. An alternative is to shorten the informational distance between an observed outcome and the employee able to respond. Direct employee access to performance data has already been investigated (Bernstein & Li, 2025). The proposed model instead specifies how contextual comparison, provisional explanations, actionable feedback, employee choice, reassessment, and human escalation could be integrated. This architecture is a testable design proposal, not a proven improvement over existing practice.

12. Declarations

12.1. Conflict of interest

The author declares no conflict of interest.

12.2. Funding

This research received no external funding.

12.3. Data availability

No new empirical datasets were created within this conceptual article. All figures in Table 1 are hypothetical.

12.4. Use of generative AI

ChatGPT was used to assist with manuscript organization, language editing, drafting, and preliminary literature selection. The author retains responsibility for the accuracy of citations, interpretation of the literature, and all scientific claims.

12.5. Peer review

The manuscript underwent Double-Anonymous Peer Review by at least two independent subject-matter reviewers in accordance with the current editorial policy of IAE Review. The author is not affiliated with the journal's editorial office or publisher; the final publication decision was made by the editorial office.