Your AI Savings Are on Someone Else's Timesheet

Imagine an analyst using AI to prepare a report in ten minutes. The same draft used to take an hour.

Fifty minutes saved. An easy win for the AI rollout.

Then the report reaches a senior specialist. The specialist reopens the source documents, checks the calculations, questions two assumptions, and sends it back for clarification. A fifteen-minute review becomes forty-five minutes. The analyst spends another ten minutes answering the questions.

The draft got faster. The specialist's queue got longer. Both experiences are real.

That is the part of enterprise AI economics that deserves more attention: whose time a system saves, whose time it consumes, and which of those people the business can least afford to interrupt.

The arithmetic of a faster draft

Consider the arithmetic. This is an illustrative workflow with assumed times and fully loaded labor rates, not an EVE customer result. Drafting and follow-up cost $60 per hour; specialist review costs $120 per hour. Both workflows end with an accepted report of equivalent quality.

The company saves ten minutes of total labor. Its allocated labor cost rises by $20, before AI subscription or usage charges.

Payroll may be unchanged that month. The cost shows up in how the company uses the capacity it already pays for. Thirty additional minutes of specialist attention have to come from somewhere.

That could mean a customer waiting longer, a harder case staying unresolved, or an experienced employee spending less time teaching someone else.

What the research suggests

Research gives this concern a basis beyond hypothetical arithmetic. In January 2026, Workday published survey findings from 3,200 active AI users at organizations with at least $100 million in annual revenue. It reported that nearly 40% of AI time savings were consumed by rework, including checking and correcting outputs. The survey was fielded in November 2025. Workday research.

Glean's 2026 Work AI Index, based on a survey of 6,000 digital workers in the United States, United Kingdom, and Australia, reported 6.4 hours a week spent supplying context, supervising AI, fixing errors, and moving between tools. Its survey ran from December 2025 through January 2026. Glean research announcement.

These are vendor-sponsored surveys of reported experience. They do not establish the return on a particular deployment. They do give leaders a reason to measure the human work surrounding AI alongside the time employees say it saves.

Where the extra work lands

My concern is where that work lands.

The employee who knows the system's exceptions can become the natural destination for every uncertain output. A missing source, an unfamiliar calculation, or an ambiguous recommendation all arrive at the same desk with the same request: "Can you quickly check this?"

Each request seems reasonable. Together, they can quietly redesign that person's job.

The specialist becomes responsible for making everybody else's faster work usable. Their own output falls, even as colleagues report productivity gains. A manager looking only at individual completion counts could reward the people creating the review burden and question the performance of the person absorbing it.

How to evaluate an AI pilot

That possibility should change how an AI pilot is evaluated.

Follow a piece of work through to acceptance. Record the time spent preparing it, reviewing it, correcting it, and resolving follow-up questions. Include the people who receive the output. Apply the same measurement to comparable work completed without AI, including the review and errors that already existed.

Then ask four practical questions:

1. Did the cost per accepted outcome fall? Include labor across every participating role, tool charges, and an appropriate share of setup and maintenance.
2. Did quality hold up? Track whether the result met the same acceptance standard and whether it later had to be reopened.
3. Did the constrained team gain capacity? Check what happened to the specialist, approval, or operations queue that already limited throughput.
4. Where did the recovered time go? Identify the additional work completed or the capacity employees actually regained.

An AI workflow may still be worthwhile when it costs more. Better quality, faster customer response, or handling work that previously went undone can justify the expense. Make that benefit explicit in the business case.

Improve the handoff

Once the extra work is visible, improve the handoff.

Give the person receiving an AI-assisted report its source material, assumptions, unresolved questions, and the specific decision it is meant to support. A polished document with missing context makes the reviewer reconstruct the task before they can judge the answer.

For repeatable checks, use tested automation where it fits. Arithmetic, required fields, duplicates, and defined approval limits can often be checked before a specialist sees the work. Measure false alarms too: a control that sends every ordinary case to a human can create its own expensive queue.

For agents that change business systems, make the permitted scope and approval requirements explicit. Preserve a record of the proposed action and the control decision so employees can investigate an exception without rebuilding the entire sequence from memory.

This is where EVE's work connects to the economics. EVE CoreGuard focuses on authority checks before execution, and EVE Proof provides verifiable records of governance decisions. Those controls address a defined part of the workflow. Their effect on review effort and business outcomes should be measured in the deployment; a signed decision record does not establish that every underlying claim is correct.

The broader management responsibility remains with the business: design work that makes good use of both AI capability and human expertise.

Before the next AI success presentation, invite the person who receives the output. Ask what changed in their day.

They may confirm the savings. They may identify the next improvement. Or they may show you the hours missing from the slide.

Your AI productivity story is incomplete until the person checking its work gets a line in the business case.

Frequently asked questions

How can an AI workflow save time and still cost more?

Because the hours saved and the hours added are not billed at the same rate or drawn from the same person. In the illustrative workflow above, drafting drops from 60 minutes to 10, but specialist review rises from 15 minutes to 45 and follow-up adds 10. Total labor falls by ten minutes while allocated cost rises by $20, because the added time belongs to a more expensive reviewer. Any comparison that counts only minutes, and only the first role in the chain, will miss this.

What should an AI pilot actually measure?

Cost per accepted outcome, measured across every role that touches the work through to acceptance — preparation, review, correction, and follow-up — with the same measurement applied to comparable non-AI work, including the review and errors that already existed. Alongside that: whether quality held to the same acceptance standard, whether the already-constrained queue gained capacity, and where the recovered time actually went.

Why does it matter which person absorbs the review work?

Because the person who knows the system's exceptions is usually the one the business can least afford to interrupt, and they are the natural destination for every uncertain output. Their own throughput falls while colleagues report gains, so a manager looking only at individual completion counts can end up rewarding the people creating the review burden and questioning the person absorbing it.

Does this mean AI deployments are not worth it?

No. An AI workflow can be worth adopting even when it costs more — better quality, faster customer response, or handling work that previously went undone are all legitimate justifications. The argument here is narrower: make that benefit explicit rather than presenting a cost saving that the full workflow does not support.

What reduces the review burden rather than relocating it?

Two things. First, a better handoff: give the reviewer the source material, assumptions, unresolved questions, and the decision the output is meant to support, so they are not reconstructing the task before judging it. Second, tested automation for repeatable checks — arithmetic, required fields, duplicates, defined approval limits — applied before a specialist sees the work, with false alarms measured, since a control that escalates every ordinary case creates its own expensive queue.

Where do governance controls fit into this?

For agents that change business systems, an explicit permitted scope and approval requirement decides which actions clear automatically and which reach a person, and a preserved record of the proposed action and the control decision lets someone investigate an exception without rebuilding the sequence from memory. That addresses a defined part of the workflow. Its effect on review effort should still be measured in the deployment; a signed decision record does not establish that every underlying claim is correct.