AIReturn

AI ROI — the fundamentals

AI Rework: The Hidden Cost That Erases Your AI ROI

AI rework — redoing AI output that wasn't right — silently erases AI ROI. See how to measure it and cut the cost, team by team.

Rodrigo Paredes BassiPublished 10 min read

TL;DR

  • AI rework is the redo cost of AI output that looks done but isn't — the human time to correct, re-instruct, or re-do it. Un-measured, it quietly cancels the productivity AI appears to deliver.
  • In one study, 41% of workers received "workslop" — plausible-looking but substandard AI output — costing ~1h 56m to resolve per instance and an estimated ~$186 per employee per month (BetterUp Labs + Stanford, HBR, Sep 2025).
  • Only 28% of workers in that same study said AI improved decision quality. Polished output does not mean better work.
  • FinOps sees the cost. Observability sees model quality. Adoption dashboards see usage. None of them see rework — the gap between "AI produced output" and "the work actually got done."
  • AIReturn makes rework measurable by tracking function-specific signals (code churn, reopened tickets, stage regression, requirements/spec volatility) against each team's own baseline — the missing variable in the AI ROI equation.

What is AI rework?

AI rework is the redo cost incurred when AI output looks finished but isn't: the human time and effort to correct, re-instruct, or re-do work an AI tool produced. It is the hidden denominator that can turn apparent AI savings into a net-negative outcome. When AI shifts effort downstream instead of removing it, rework is where that effort lands — and it rarely shows up on any dashboard a CFO already owns. The pattern is straightforward. A tool generates a draft, a summary, a block of code, or a customer reply in seconds. It looks done. Then someone downstream discovers it is missing context, subtly wrong, or generic enough to be useless — and has to fix it. The first-order gain was real and visible. The second-order cost was real and invisible. AI often shifts effort downstream — polished output that still has to be corrected, clarified, or redone. Un-measured, rework silently cancels the productivity gains AI appears to deliver.

AI rework vs. workslop: the symptom and the cost

These two terms are related but not the same, and the distinction matters for measurement.

  • Workslop (coined by BetterUp Labs + Stanford) is AI-generated output that looks polished but lacks the substance to advance the task, forcing the recipient to redo it. It names the symptom — the artifact that fails downstream.
  • AI rework is AIReturn's name for the measurable cost of that symptom: the effort spent redoing, re-instructing, or correcting the work. Workslop is what you receive; rework is what it costs you. You can have rework without formal "workslop" — an AI-generated pull request that passes review but gets reverted a week later is rework, whether or not anyone called it slop. The point of naming both is to move from a vibe ("the AI stuff needs a lot of cleanup") to a number.

The cost of fixing AI output, in dollars people actually recognize

The clearest evidence to date comes from BetterUp Labs and Stanford's Social Media Lab, published in Harvard Business Review in September 2025. In a survey of 1,150 US full-time employees, 41% reported receiving workslop in the prior month. Each instance took an average of 1h 56m to resolve. The researchers estimated the invisible tax at ~$186 per employee per month — which they projected to roughly $9M per year for an organization of 10,000 people, based on self-reported time. Two caveats keep this honest. The dollar figure is built from self-reported time, not instrumented telemetry — directionally strong, not audited. And the same study found only 28% of workers said AI improved the quality of their decisions. Output volume went up; whether the work got better is a separate question — and it is the one that decides ROI.

MetricValueWhat it measures
Workers who received workslop (prior month)41%Prevalence of substandard AI output landing on a colleague
Time to resolve one instance~1h 56mPer-incident rework cost
Estimated monthly cost per employee~$186Self-reported rework tax
Say AI improved decision quality28%The quality signal behind the volume
Source: BetterUp Labs + Stanford Social Media Lab, HBR, September 2025 (n=1,150 US full-time workers; dollar figures self-reported).

The perception gap: why teams don't feel the rework

Rework is easy to miss because it is often invisible to the person who created it. The effort moves to someone else, or to a later week, and the original author experiences only the speed-up. An early-2025 study by METR makes the gap concrete in engineering. In a randomized trial of 16 experienced open-source developers across 246 tasks in large, familiar repositories, developers using AI tools were measured 19% slower — while estimating they had been 20% faster (a 39-point perception gap). One important caveat: METR later relabeled this result "historical" and redesigned its methodology in February 2026, so treat it as a single early-2025 snapshot, not a timeless law that "AI makes developers slower." The durable insight is not the direction of the number — it is the size of the gap between felt speed and measured outcome. When perception says "20% faster" and measurement says otherwise, self-reported time-saved is not a reliable input to an ROI model. Rework hides in exactly that gap.

Why nobody's dashboard catches rework

The reason rework stays invisible is structural. Every category of AI tooling a CFO already pays for is designed to answer a different question:

  • FinOps and cost tools see the bill, not the worth. They allocate 100% of spend to the token. Knowing what AI cost is not knowing what it was worth. Cost is the denominator; it says nothing about whether the output survived contact with reality.
  • Observability and eval tools see model quality, not work quality. A 0.95 groundedness score does not tell your CFO whether the marketing brief shipped or got rewritten. Green evals are not the same as good work.
  • Adoption dashboards see usage, not impact. "Employees feel they saved three hours" is a survey, not a P&L line. High adoption of a high-rework use case can increase net cost while every activity metric climbs. Each tool is correct within its lane and blind to the next one. Rework lives in the seam between them — the distance between "AI produced output" and "the work actually got done." That seam is the 4th question of AI ROI that nobody answers continuously: did AI actually improve the work, by team, net of rework — and what do you fix?

How to measure AI rework, team by team

You cannot bill AI rework as a line item, but you can measure it through the signals it leaves in the systems teams already use. The method is the same across functions: pick signals native to how each team works, normalize them, and compare only against that team's own history — never across teams, because output and its friction are not comparable between functions. A support queue and an engineering backlog do not share units. Below are base-level rework signals by function. This is AIReturn's methodology: make rework observable where it actually accumulates.

FunctionRework signals (proxy for redo)Compared against
EngineeringCode churn, reverts, PR review rounds, reopened issuesThe team's own pre-AI baseline
SupportReopened tickets, reassignments, escalations, reply cyclesThe team's own pre-AI baseline
SalesStage regression, time-in-stage, extra touches per dealThe team's own pre-AI baseline
ProductRequirements/spec volatility (specs change, reopen, or get re-cut after work started), review loops; items moving backward through statuses as a hygiene sub-signalThe team's own pre-AI baseline
Signals are normalized to an index and read as a trend against each team's history — not as a cross-team scoreboard.
Read alongside the AI cost for that team, these signals plot each function on the Cost × Rework matrix — a 2×2 diagnostic mapping AI cost (high/low) against AI rework (high/low). The result is a per-team verdict: scale (low cost, low rework), keep (working, watch it), fix (high rework — the harness or workflow needs work before more spend), or cut (high cost, high rework, no output to show for it). That matrix is the CFO's decision artifact for funding AI team by team, rather than defending one org-wide "AI budget" nobody can substantiate. You can see a sample per-team AI ROI report for how the verdict renders.
An honest limit: the attribution window between AI-produced work and the rework it causes is not fully pinned, and isolating AI-caused rework from a team's baseline rework is the hardest open question in this work. In engineering it is concrete — churn, reverts, reopens, and review rounds are already logged. For deals, initiatives, and support cases, the proxies are still maturing. We would rather name that boundary than over-claim a precise dollar of rework across every function.

Rework is the missing variable in AI ROI

Put it together and the equation changes. Most AI ROI math looks at output per dollar and stops. The real unit is output-per-dollar-net-of-rework: the business output a team produces per dollar of AI spend, after subtracting the cost of redoing flawed AI work.

AI ROI = (business outcome delivered − rework cost) ÷ (AI cost + human cost to get it right), measured per team, continuously, across every AI vendor. Ignore the rework term and a high-adoption, high-rework team can look like a win while quietly running net-negative. That is not a hypothetical — it is the mechanism behind the wider measurement gap. CloudZero's June 2026 finance survey found 78% of finance executives cannot fully tie AI spend to business outcomes. Rework is a large part of what makes the tie so hard: the cost of the redo is real, distributed, and — until you measure it — unattributed. Getting it onto the ledger is what turns the framework CFOs use to measure AI ROI from an estimate into an accountable number, and it is why the real ROI unit is output minus rework.

Common ways rework stays hidden

  • Counting activity, not outcomes. More PRs, more replies, more drafts. Volume rises; you never see how much of it came back.
  • Trusting self-reported time-saved. The 39-point perception gap (METR, early 2025) shows why felt speed is not measured outcome.
  • Reading eval scores as work quality. A grounded, on-policy answer can still be the wrong answer for the task.
  • Averaging across teams. A blended "AI is working" number hides the one function silently generating most of the rework.
  • Measuring once. Rework moves as models, prompts, and workflows change. An annual snapshot misses it; a continuous read catches it.

FAQ

What is AI rework? AI rework is the redo cost of AI output that looks finished but isn't — the human time to correct, re-instruct, or re-do work an AI tool produced. It is the hidden denominator that can turn apparent AI savings into a net-negative outcome. Unlike model quality or usage, rework measures whether the work actually got done, which is what determines return. What is workslop, and how is it different from rework? Workslop (coined by BetterUp Labs + Stanford) is AI-generated output that looks polished but lacks the substance to advance the task, forcing the recipient to redo it. It names the symptom. AI rework is the measurable cost of that symptom — the effort spent fixing it. One study found 41% of workers received workslop, at roughly 1h 56m to resolve each instance (HBR, 2025). How much does fixing AI output actually cost? In BetterUp Labs and Stanford's 2025 study, resolving a single instance of substandard AI output took an average of ~1h 56m, which the researchers estimated at ~$186 per employee per month (self-reported). At 10,000 employees that projects to roughly $9M per year. The figure is directional, not audited — but it is large enough to erase the productivity gains AI appears to deliver. How do you measure AI rework per team? Track function-specific signals of redo — code churn, reverts, and PR review rounds in engineering; reopened tickets and escalations in support; stage regression and extra touches in sales; requirements/spec volatility and review loops in product. Normalize each signal and compare it only against that team's own history, never across teams, then read it beside AI cost on the Cost × Rework matrix. Why don't our existing AI tools show rework? Because each is built for a different question. FinOps tools see cost, observability and eval tools see model quality, and adoption dashboards see usage. Rework lives in the seam between them — the gap between "AI produced output" and "the work got done." Measuring it requires connecting delivery data across functions and comparing to a baseline, which single-lane dashboards don't do.

Sources

See the return on your own numbers.

Get a clear, per-team return-on-AI picture in weeks — output net of rework, reconciled to your invoice.