Employees Work Faster with AI, But Companies Don't Necessarily Earn More

Employees Work Faster with AI, But Companies Don't Necessarily Earn More

By Hou Yanjun

When employees complete in half an hour what previously took two hours, where does that saved hour and a half actually go?

This accounting deserves deeper examination. Current discussions about AI efficiency often stop at "before vs. after" time comparisons. The demonstrations are compelling—a proposal that used to take half a day now gets generated in minutes. But executives invest in AI for more than just producing extra drafts.

Does the saved time enable serving more clients, reducing overtime, or addressing previously neglected issues? These all create value. The concern arises when time savings don't change business outcomes, yet get mechanically multiplied by hourly wages to claim project benefits.

If payroll costs remain unchanged, outsourcing fees stay level, and business volume doesn't increase—where exactly are the savings?

Saved work hours can estimate employees' freed-up capacity but shouldn't be directly counted as monetary savings. Only when this time replaces outsourcing or generates additional orders can financial benefits be calculated. If used for service improvements or reduced overtime, evaluation must reflect actual changes rather than counting everything as cash savings.

The same AI functionality creates vastly different value across companies and processes. Available business demand and downstream workflow readiness ultimately determine efficiency gains' real impact.

Consider recruitment as a hypothetical example: AI accelerates resume screening, enabling recruiters to produce candidate shortlists same-day. But if hiring managers remain unclear about requirements, struggle to schedule interviews, or keep changing position criteria, early shortlist delivery doesn't necessarily mean faster hiring.

Further increasing screening speed becomes meaningless when the real bottlenecks lie in undefined hiring standards, decision-making delays, and shifting requirements. While resume screening offers easily quantifiable metrics for demos, the actual hiring delays stem from unstable demand and protracted approvals.

Unless recruiters address these upstream issues, faster screening only means waiting earlier in the process. Changing requirements may force re-screening, negating initial time savings.

Therefore, companies selecting AI projects shouldn't focus solely on easily demonstrable quick wins. Pilot testing simple applications makes sense, but scaling requires evaluating end-to-end process improvements. Whether a previously month-long task now finishes earlier with maintained quality matters more than isolated step acceleration.

Another efficiency scenario demands careful accounting: faster generation but prolonged verification.

Producing reports in minutes beats staring at blank documents, but someone must still verify data accuracy, metric consistency, and conclusion validity—especially for deceptively smooth outputs hiding flawed details. Reviewers might need exhaustive source-data cross-checking.

Measuring only generation time while ignoring verification duration overstates efficiency gains. True before-after comparisons should encompass the entire workflow from data preparation to deliverable approval, including execution, review, revisions, and rework.

Saving one hour upfront but adding one hour in verification shouldn't count as pure gain. If verification requires higher-cost specialists, total expenses might actually increase.

The same logic applies to model costs. While pricing matters, cheaper models requiring frequent regeneration and manual correction may prove uneconomical. Companies bear total completion costs encompassing both tool expenses and human time.

A study of 5,172 customer service agents found AI tools increased productivity by approximately 15% measured by successfully resolved issues per hour. "Successful resolution" proves crucial—faster responses that don't actually solve problems lead to repeat contacts, negating service improvements.

This study offers valuable methodology: speed must be evaluated alongside completion quality. While it doesn't claim 15% profit growth, this metric better reflects real service impact than mere response counts.

Companies should evaluate internal projects similarly. Initial problem statements should determine success metrics—if shortening delivery cycles was the goal, don't settle for counting generated materials or tool usage frequency. Usage data indicates adoption but can't replace measuring actual cycle time, quality, and cost changes.

Evaluation must allow sufficient pilot observation periods. Data organization, training, and process adjustments all require investment. Absent immediate profit growth doesn't equal project failure, but leaders must define clear validation criteria, timelines, and scaling conditions.

For example: first confirm output usability, then compare total manual time requirements, and finally verify whether saved time generates additional business or reduces existing costs. Each step should inform subsequent decisions—neither declaring victory at deployment nor perpetually citing "exploration phase" to explain stagnation.

Some benefits won't immediately appear on income statements but remain valuable: previously neglected client follow-ups now handled, supervisors available to mentor juniors facing challenges, teams no longer relying on overtime to meet deadlines. Documenting these changes separately from financial gains enables better ROI assessment.

"Employees work faster" represents progress, but the accounting must continue until work gets properly completed, results become usable, and companies effectively utilize saved time.

If early candidate shortlists don't accelerate hiring, address decision-making bottlenecks. If faster report generation increases verification time, examine data and output quality. When business outcomes remain unchanged, identifying root causes proves more valuable than pushing AI to go even faster.

References

[1] Brynjolfsson, Li, Raymond, 2025: Generative AI at Work, The Quarterly Journal of Economics. Study covering 5,172 customer service agents at one enterprise, showing ~15% average increase in successfully resolved issues per hour, not corporate profit growth.