AI tools can now handle parts of early-career knowledge work, including document summaries, initial spreadsheet analysis, and standard memo drafts. That changes which parts of the work reveal an employee's judgment.
A field study described in Harvard Business Review (July 2026), "Research: Why Some Junior Employees Work Well with AI—and Others Don’t", reports that researchers from KPMG and the University of Texas at Austin evaluated 523 early-career professionals with tenure of 18 months or less as they completed domain-specific client tasks with enterprise AI agents.
Reported finding: In this study, critical thinking, domain knowledge, and AI literacy did not by themselves predict performance with AI. The researchers instead distinguished participants by how they directed, questioned, and revised the AI-assisted work.
Three reported workflow groups
The researchers compared participants' deliverables with an AI-only baseline using expert rubrics and grouped the observed workflows into three profiles. The descriptions below are my concise synthesis of the article's groups, not additional findings from the underlying study.
| Group used in the article | Share (%) | Skills reported | Workflow described | Reported result |
|---|---|---|---|---|
| AI Apprentices | 24.1% |
High (equivalent to Amplifiers) | Frequent micro-editing and cosmetic rewording, with attention moving to peripheral questions rather than task-level direction. | Below AI Baseline |
| AI Delegators | 25.8% |
Lowest across all metrics | Mostly accepted the initial draft, with little interrogation or hypothesis testing. | Equal to AI Baseline |
| AI Amplifiers | 50.1% |
High (strong domain + reasoning) | Actively directed the work, challenged assumptions, tested edge cases, and applied domain framing. | Significantly Outperformed |
Observed behaviors
1. AI Apprentices (24.1%)
Apprentices scored highly in critical thinking and domain knowledge, but their final performance was below the AI-only baseline.
When synthesizing financial filings and client exhibits, they spent many interactions on formatting, rephrasing, or marginal questions. The activity did not consistently improve the task-level analysis.
2. AI Delegators (25.8%)
Delegators had lower foundational-skill scores and relied heavily on the model's first draft.
They often pasted the brief, requested the deliverable, and accepted the result with little critique. In the study, their outputs were roughly equal to the AI-only baseline, so the workflow showed little added human judgment.
3. AI Amplifiers (50.1%)
Amplifiers combined stronger foundational skills with active direction and review of the AI's work.
In plain terms, their workflows included:
- Frame the problem: State domain constraints before asking for calculations.
- Challenge the draft: Ask why alternatives were rejected and identify missing evidence.
- Check intermediate work: Break a large task into steps and verify claims before writing the recommendation.
Possible implications for teams
My synthesis for team practice:
- Assess process as well as output: A final PDF or slide deck may hide how much judgment the employee applied. Reviews can include how candidates steer, verify, and test the model.
- Record decision rationale: Workflows can ask practitioners to log why an AI recommendation was accepted, modified, or rejected.
- Practice with domain-specific scenarios: Simulations can test whether practitioners identify unsupported claims and structural gaps in AI drafts.
Connection to agent harness design
The findings also suggest a design question for agent harnesses: whether the interface makes review and intervention visible.
A harness can expose the model's work and give the user explicit points to review, redirect, or reject it.
For tool routing, multi-turn loops, and execution sandboxes, observable controls can make human intervention part of the workflow. The study does not establish that an interface alone will move users from one profile to another.