Rating Human-Machine Co-Production in Modern Performance Systems

A senior software engineer in Berlin submits four pull requests before noon. A financial analyst in New York produces a forty-page valuation memorandum in two hours. A customer operations manager in Toronto resolves eighty escalation tickets in a single shift. Historical performance rubrics would rate all three employees as extraordinary performers. Their output volume sits three standard deviations above historical team averages. Their turnaround times are unprecedented. Underneath these metrics lies a different reality. The engineer accepted ninety percent of the code suggested by GitHub Copilot without reviewing the memory allocation routines. The analyst passed raw prompts to Claude 3.5 Sonnet, missing two hallucinated EBITDA adjustments in the final appendix. The customer manager automated response chains using a custom agent that closed customer cases without verifying account verification tokens. Raw output volume has detached from human competence. When artificial intelligence generates draft zero in twelve seconds, volume metrics break. Evaluating employees based on finished artifacts served organizations well when human effort scaled linearly with output. That linear relationship no longer exists. Performance management must adapt to co-production. HR leaders, compensation committees, and line managers must address a fundamental question: what are you actually rating when the machine does the heavy lifting? ## The Shift from Production to Verification Traditional performance evaluation focuses on execution speed, output quantity, and basic accuracy. In knowledge work, execution speed historically correlated with mastery. A fast writer understood narrative structure. A fast coder understood algorithm efficiency. A fast paralegal understood case law taxonomy. Generative systems invert this dynamic. Execution speed is now a property of the software tier, available compute, and context window size. The human contribution has migrated to the ends of the production process: problem framing at the start, and verification at the end. Consider the breakdown of effort in a financial modeling task. Prior to generative tools, a senior analyst spent fifteen percent of their time framing the problem, sixty percent constructing formulas and populating spreadsheets, and twenty-five percent reviewing and refining scenario outcomes. Today, generative systems complete the construction phase in seconds. The analyst spends thirty percent of their time setting parameters and constraints, and seventy percent auditing the model for structural integrity, logical fallacies, and underlying assumptions. > Rating an employee on the volume of code, text, or analysis generated by a model creates a false signal of capability while actively disguising operational risk. Evaluating the analyst on how quickly the initial draft appeared measures machine efficiency, not human skill. Evaluating the analyst on whether they caught an inverted discount rate hidden within cell C42 measures human capability. Performance systems that fail to isolate the human contribution reward high-volume, low-scrutiny workers while penalizing diligent verifiers who take time to correct synthetic errors. ## Regulatory Boundaries: AEDTs, the EU AI Act, and Labor Rights Redesigning performance systems for co-produced work is not purely an internal operational choice. Regulatory bodies in North America and Europe have established strict legal boundaries around how technology monitors, measures, and rates employee performance. In New York City, Local Law 144 regulates Automated Employment Decision Tools (AEDTs). If an enterprise uses software to analyze prompt logs, keyboard metrics, or AI utilization rates to score, rank, or evaluate employees for promotions, salary adjustments, or terminations, that software may qualify as an AEDT. Qualifying tools require independent bias audits, public summaries of impact metrics, and explicit advance notice to workers. Across the European Union, the EU AI Act classifies artificial intelligence systems used in workplace management, performance evaluation, and task allocation as high-risk under Annex III. Organizations deploying performance analytics driven by machine learning face strict transparency mandates. Employers must complete fundamental rights impact assessments, maintain human oversight over automated scoring systems, and log system operations continuously. In Germany, worker representation laws introduce structural boundaries around performance tracking. Section 87, Paragraph 1, Number 6 of the Works Constitution Act (Betriebsverfassungsgesetz) grants the Works Council (Betriebsrat) mandatory co-determination rights regarding any technical device designed to monitor worker conduct or performance. Tracking individual prompt records, usage duration, or real-time acceptance rates of AI suggestions in tools like Microsoft Copilot requires a formal Works Agreement (Betriebsvereinbarung). Similar protections exist in France through the Commission Nationale de l'Informatique et des Libertes (CNIL) guidelines on workplace surveillance, and in Canada under Ontario's Electronic Monitoring Act. Organizations cannot simply deploy background telemetry to measure how employees interact with machine tools. Evaluation systems must rely on transparent, job-related criteria that measure human reasoning rather than raw interaction logs. ## Deconstructing the Four New Performance Vectors To rate human performance accurately during co-production, People Operations teams must discard generic quality scales. Evaluation frameworks must assess four specific human competencies. ### Vector 1: Intent Architecture and Context Curation Machines require context to generate relevant output. The quality of synthetic output depends directly on how well an employee defines the problem, structures constraints, and curates reference data. Intent architecture measures an employee's ability to translate complex business objectives into precise instructions for non-deterministic software. A weak performer issues broad prompts like "write a vendor contract for cloud hosting." A strong performer constructs detailed context windows, specifying jurisdiction, liability caps, indemnification exceptions, and service level thresholds. Evaluating this vector requires managers to assess input hygiene:

• Problem Decomposition: Breaking large strategic goals into discrete, logical processing steps.
• Boundary Setting: Explicitly defining what the machine must not do, including regulatory compliance rules and proprietary data boundaries.
• Context Selection: Feeding the system accurate, non-redundant, high-signal reference materials. ### Vector 2: Verification Velocity and Error Pruning Generative models produce authoritative-sounding statements regardless of accuracy. The primary human duty in co-production is quality assurance. Verification velocity measures how efficiently and thoroughly an employee audits machine output. High performance in this vector is not about speed alone. It is about error density detection per unit of time. Consider software engineering teams using GitHub Copilot or Amazon Q. Historical metrics tracked line count or pull request frequency. Modern metrics must evaluate review thoroughness:
• Edge Case Identification: Detecting boundary condition failures that pass initial automated unit tests.
• Logic Auditing: Verifying that synthetic code aligns with long-term architecture rather than introducing technical debt.
• Hallucination Pruning: Catching references to non-existent internal libraries, retired API endpoints, or invalid legal citations. An employee who rejects six incorrect machine suggestions and refactors the underlying logic is vastly more valuable than an employee who accepts ten code blocks without scrutiny. ### Vector 3: Domain Grounding and Strategic Alignment Generative tools operate on statistical probability across historic training data. They lack situational awareness regarding an organization's political realities, strategic priorities, and competitive positioning. Domain grounding measures an employee's capability to bridge the gap between generic output and institutional reality. A marketing team member might use an LLM to generate ten product launch angles. The machine cannot know that a major client is renegotiating terms, or that a competitor is launching a similar feature next week. High-performing employees take raw machine artifacts and adapt them to current business realities:
• Strategic Translation: Modifying output to align with unwritten organizational goals and executive preferences.
• Stakeholder Adaptation: Adjusting tone, nuance, and emphasis for specific external clients or internal leadership teams.
• Ethical Boundaries: Screening synthetic output for subtle bias, brand risk, or regulatory non-compliance. ### Vector 4: Workflow Design and System Orchestration Single-prompt tasks represent basic software literacy. High performers design repeatable, multi-step orchestration workflows that automate routine administrative cycles across departments. System orchestration evaluates an employee's ability to connect disparate software interfaces, databases, and generative models to optimize business operations:
• Tool Integration: Chaining specialized model outputs together to solve multi-stage operational problems.
• Workflow Standardization: Documentation of successful operational templates so peers can replicate efficiency gains.
• Process Maintenance: Updating context libraries and instruction prompts when underlying operational rules change. An operations specialist who builds a reusable, validated workflow for processing cross-border tax documentation performs at a higher level than an operator who manually enters prompts for each invoice. ## The Attribution Problem and Measuring Value Add Co-production creates an attribution problem for performance management. When a deliverable achieves an excellent outcome, how much credit belongs to the human, and how much belongs to the software? To solve this, organizations must calculate Human Value Add (HVA). HVA represents the qualitative difference between the baseline machine output and the final submitted work product. ``` Human Value Add = Final Artifact Quality - Baseline Machine Output Quality

2. Did the associate catch critical omissions that the software failed to flag?
3. Did the associate's modifications enhance the firm's negotiating position? This methodology separates software speed from human competence. It ensures that employees are compensated for their analytical judgment rather than their ability to click accept on automated suggestions. ## Managerial Workflows and Calibration Redesign Transitioning to co-production performance management requires retraining line managers. Most managers were promoted based on their mastery of traditional execution. They are rarely trained to evaluate human-machine interaction. Organizations must update managerial workflows across four distinct stages. | Stage | Traditional Evaluation Workflow | Co-Production Evaluation Workflow |
| :--- | :--- | :--- |
| Planning | Assign task and set delivery deadline. | Define machine tooling permitted, required context boundaries, and target quality benchmarks. |
| Execution | Monitor progress via status updates and draft reviews. | Spot-check context inputs, prompt strategy, and interim verification steps. |
| Review | Assess finished deliverable against static rubric for speed and format. | Calculate Human Value Add by comparing raw machine output against final submission. |
| Calibration | Rank employees by output volume, project count, and deadline adherence. | Evaluate error detection rates, domain judgment, and process orchestration achievements. | Calibration meetings must change. In typical tech and corporate services calibration sessions, managers bring list prices of completed projects. In co-production environments, calibration committees must examine verification samples and error-handling logs. During quarterly calibration at an enterprise software firm in California, managers formerly defended rating recommendations using commit counts and resolved tickets. Under the co-production rubric, engineering directors present code diffs showing where developers corrected synthetic logic flaws in automated code suggestions. A developer with forty high-scrutiny pull requests rates higher than a developer with one hundred low-scrutiny pull requests. This operational shift takes time. Managers spend approximately twenty-five percent more time during the initial two quarters auditing work processes rather than skimming final outputs. That investment is essential to prevent systemic quality degradation. ## Pay Equity, Compensation Bands, and Regulatory Compliance Co-production complicates compensation strategy. When software increases output volume across all performance tiers, traditional merit pay formulas break down. The EU Pay Transparency Directive (Directive 2023/970), which EU member states must transpose into national law by June 2026, requires employers to use gender-neutral job evaluation and classification systems. Objective, transparent criteria must govern pay progression and performance-related compensation. Variable pay tied to subjective managerial metrics will trigger legal challenge and mandatory pay assessments if gender pay gaps exceed five percent. If an organization awards performance bonuses based on raw output volume in an environment where AI tools are used unevenly, compensation equity is compromised. Male employees might adopt unvetted external AI tools at higher rates to inflate output metrics, while female employees follow internal risk guidelines and produce lower raw volume while maintaining higher verification standards. If performance ratings reward volume, the enterprise inflates its risk profile while violating gender pay equity rules. To ensure compliance with the Pay Transparency Directive and US equal pay frameworks (such as the California Equal Pay Act and New York Equal Pay Law), HR leaders must implement three safeguards: - Standardize AI Tool Access: Ensure all employees within a job band receive equal access to enterprise software models and compute infrastructure.
- Remove Volume-Weighted Scoring: Strip raw output volume, page counts, and simple task counts from variable compensation formulas.
- Objective Verification Metrics: Base performance bonuses on verifiable error-detection rates, domain adaptation metrics, and audited compliance standards. Skill premiums are shifting. Historical compensation models paid premiums for execution mechanics, such as writing syntax or manual financial modeling. Modern compensation models pay premiums for deep domain expertise and risk management. An entry-level analyst who relies entirely on synthetic outputs stays flat in salary progression. A domain specialist who understands underlying accounting rules, regulatory nuances, and machine failure modes commands a premium. Compensation bands must reflect this shift by expanding pay ranges for higher-level domain judgment while narrowing bands for generic administrative execution. ## Unresolved Tensions and the Three-Year Horizon As organizations refine performance management for co-produced work, three unresolved tensions will shape the next two to three years. ### The Skill Degradation Loop
When entry-level employees rely on generative tools to perform baseline work, they skip the basic operational execution that historically built domain expertise. A junior lawyer who never drafts basic non-disclosure agreements from scratch may fail to develop the mental models required to spot subtle errors in synthetic outputs five years later. Performance management systems do not currently account for cognitive offloading. Evaluating workers solely on current verification efficiency risks creating a workforce that loses foundational knowledge over time. HR leaders in North America and Europe must design performance expectations that include unassisted domain practice to maintain core competencies. ### Accountability and Liability for Synthetic Output
When a generative model hallucinations leads to a regulatory breach, legal penalty, or breach of contract, who bears internal performance consequences? If an employee publishes a financial disclosure containing an AI-generated calculation error, current performance systems hold the human employee entirely responsible. As synthetic systems become more autonomous and complex, isolating human fault from machine failure becomes difficult. Performance frameworks must define explicit responsibility thresholds. Did the worker follow established verification protocols? Was the failure a known edge case documented in system logs? ### Continuous Synthetic Calibration
Annual performance reviews are incompatible with fast-moving software releases. As foundational models improve, baseline machine output quality shifts every six months. A performance standard that made sense in January may become trivial by September. Leading organizations are shifting toward rolling six-month evaluation cycles grounded in continuous work sampling. Rather than assessing overall annual output, managers run periodic calibration exercises to measure an employee's performance relative to the capabilities of the software tools available at that exact moment. To align performance management with modern work, HR leaders must audit their current evaluation rubrics immediately. Remove volume-driven performance metrics from all job descriptions where workers use generative software. Define explicit verification protocols for every role co-producing work with machines. Train managers to evaluate input hygiene, error pruning, and domain judgment rather than speed. Organizations that continue to rate employees on execution speed and artifact volume will compensate machine outputs while leaving their systems exposed to operational risk.