Research Brief · AI Strategy

Why "What Are You Doing With The Extra Time?" Is the Wrong Question

Perception vs. reality, at a glance

The perception problem

Every organization that adopts AI well hits the same moment. A team that used to spend six hours building a competitive analysis now does it in ninety minutes. A manager notices, and asks — reasonably, sometimes even generously — "Great, so what are you doing with the rest of your day?"

The question feels neutral. It isn't. To an employee, it sounds like: your job just got smaller. The rational response is to hide the gain — pad timelines, quietly sit on finished work, or downplay how much AI contributed. Recent workforce surveys show this is already happening at scale: roughly a third to more than half of employees who use AI at work admit to concealing that use from managers, most commonly out of fear that visible efficiency will be read as reduced effort, reduced headcount need, or "cheating" rather than as skill.^[9,7,8]

This is a governance failure, not an employee failure. When leadership frames AI adoption as a time-recovery exercise — "how many hours did this save us?" — it invites employees to protect their hours by hiding the tool. When leadership frames it as a quality-of-output exercise, the incentive to hide disappears, because the metric that matters is no longer "hours worked" but "caliber of judgment applied."

The research base on this is now large enough, and consistent enough, to state the reframe with confidence.

The reality: hours stay flat, the composition of work changes

The strongest evidence for this comes from three separate, independently reviewed field experiments — two from Harvard Business School, one published in a top economics journal — plus corroborating survey data from McKinsey. None of them show workers doing less. All of them show workers doing different, harder, higher-judgment work in the same time.

1. The "Jagged Frontier" study (Harvard Business School / BCG, 758 consultants)

In the most cited field experiment on knowledge-work AI to date, Harvard Business School researchers partnered with Boston Consulting Group to give 758 consultants realistic, complex consulting tasks under three conditions: no AI, GPT-4 access, and GPT-4 with additional prompting guidance.^[1] Consultants using AI completed more tasks, finished faster, and — critically — produced higher-quality output on average 40% above the control group.^[1] The lowest performers gained the most, meaning AI narrowed skill gaps rather than just adding speed for people who were already fast.

The paper's key organizational insight is the "jagged frontier" itself: AI reliably improves performance on some tasks and degrades it on others, and the boundary is not obvious to workers in advance.^[1] That unevenness is exactly why judgment — knowing which tasks to hand to AI and how to verify its output — becomes the higher-value skill, replacing the lower-value skill of manually producing a first draft.

2. The "Cybernetic Teammate" study (Harvard Business School / Procter & Gamble, 776 professionals)

A follow-up pre-registered field experiment at P&G tested something more specific to the "extra time" question: what happens to collaboration, not just individual output. Professionals were randomly assigned to work with AI or without it, alone or in two-person teams.^[2]

The headline finding: a single person working with AI matched the performance of a two-person team working without it.^[2] That is not a story about someone finishing early and going home. It's a story about one person now carrying the combined analytical range — the "acumen" — that used to require pairing two people together. The extra time didn't evaporate; it was absorbed into a higher volume of expert-level reasoning per person.

3. "Generative AI at Work" (published in the Quarterly Journal of Economics, 5,172 customer-support agents)

This is the most rigorous large-N study of the "who benefits" question. Studying the staggered rollout of a generative AI assistant across more than 5,000 support agents at a Fortune 500 company, the researchers found a 14–15% average productivity gain — but with the gains concentrated almost entirely among newer and lower-skilled agents (up to 34% improvement), while veteran experts saw little to no speed gain.^[3]

The mechanism: AI disseminated the tacit, hard-won techniques of the company's best agents to everyone else, in effect giving novice employees access to expert-level acumen on demand.^[3] Customer sentiment and employee retention both improved alongside the productivity gain — a sign that the quality of the interaction rose, not just its throughput.^[3]

4. Corroborating evidence: quality gains generalize beyond consulting

A broader 2025–2026 field study of AI-assisted consulting, data-analyst, and management tasks found that AI usage raised measured output quality by roughly 18% (0.34 standard deviations) alongside an 81% increase in output per minute — and that total earnings per unit of time rose 146% once quality-based bonuses were factored in, because the work being produced cleared a materially higher bar.^[5] The pattern across every one of these studies is the same: time-on-task holds roughly constant; the difficulty and judgment-intensity of the task rises.

What this means for revenue-per-employee

This is where the "acumen-per-employee" framing connects directly to a metric finance teams already track. Revenue-per-employee (RPE) is a standard measure of labor productivity — annual revenue divided by headcount. Firms that redesign workflows around AI, rather than simply bolting a tool onto an unchanged process, are the ones that see it move. McKinsey's global 2025 State of AI survey found that among 25 organizational attributes tested, redesigning workflows around AI had the single largest measurable effect on EBIT impact — larger than which executive owns AI governance, larger than training programs alone.^[4] The same survey found that how a company manages the time freed up by AI — reassigning it to new, higher-value activities versus simply cutting headcount — was one of the specific attributes correlated with capturing bottom-line value.^[4]

Academic estimates of the underlying mechanism point the same direction: firm-level studies of AI investment find revenue-per-employee increases in the high single digits in the near term, growing to roughly 40% higher revenue within five years of sustained AI adoption, driven by upskilling existing staff rather than by adding more production labor.^[6]

The through-line: RPE rises not because employees work more hours, but because each hour carries more expert judgment. A 40-hour week that used to be 30 hours of production work and 10 hours of analysis becomes 10 hours of production and 30 hours of analysis, strategy, and decision-making — the parts of the job that actually move revenue.

Practical takeaway for leadership

Organizations that want the revenue upside — not just the cost-avoidance upside — need to change the question they ask employees. Not "what are you doing with the extra time," but "what harder problem are you now able to take on." The research above suggests three concrete moves:

1. Measure output quality and decision impact, not hours saved. Time-based metrics push employees toward concealment; quality- and outcome-based metrics don't.
2. Name the redeployment explicitly. McKinsey's data ties EBIT impact to workflow redesign specifically — not adoption alone. Tell people which higher-value tasks now belong to their role.
3. Reward disclosure, not just usage. The "shadow AI" data shows employees hide gains most when they fear the gains will be used against them. A visible, judgment-free channel for sharing what's working captures the exact "best practice diffusion" effect the Brynjolfsson study found — but only if people feel safe surfacing it.

References