Output Minus Rework: A Working Definition of AI Productivity

TL;DR

• AI productivity is useful output quality per dollar, net of rework, measured against a team's own baseline — not raw throughput. Volume is what AI inflates; net output is what pays.
• Raw throughput is the wrong unit because generative AI is designed to raise volume. More drafts, more replies, more pull requests — none of which is the same as more finished work.
• Rework is the correction that makes the number honest. One study found 41% of workers received "workslop" — plausible-looking but substandard AI output — at ~1h 56m to fix each instance (BetterUp Labs + Stanford, HBR, Sep 2025).
• The unit is output-per-dollar-net-of-rework — plainly, cost per good outcome: (outcome − rework) ÷ (AI cost + human cost to get it right), per team, vs. that team's own history.
• Only 28% of AI use cases fully meet their ROI expectations (Gartner, Apr 2026). The definition below is the difference between the 28% that can prove it and the rest still counting activity.

What is AI productivity? A working definition

AI productivity is the useful output quality a team produces per dollar of AI spend, net of the cost of redoing flawed AI work, measured against that team's own pre-AI baseline. It is not volume, not activity, and not speed. Return is the outcome that survives review; cost is the denominator; rework is the subtraction that turns an impressive-looking number into an honest one. That definition is deliberately narrow, and the narrowness is the point. Most "AI productivity" claims measure the one thing AI reliably increases — throughput — and skip the two things that decide whether the throughput was worth anything: the quality of what shipped, and the cost of fixing what didn't. This post gives the definition a shape you can quote, and contrasts it with the raw-volume version that AI quietly inflates. It is the conceptual bridge between why rework quietly erases AI ROI and our step-by-step framework for measuring AI ROI.

Why raw throughput is the wrong unit

Throughput is the wrong measure of AI productivity because generative AI is built to raise it. The technology's core function is producing more plausible output, faster. So when you measure output volume, you are measuring the tool doing exactly what it was designed to do — and learning nothing about whether the work improved. This is where "productivity theater" starts. The typical AI productivity story runs: adoption is up, drafts are up, tickets closed are up, therefore AI is working. Every one of those is a volume signal. Deloitte's State of AI in the Enterprise 2026 captures the split cleanly — 66% of leaders report productivity or efficiency gains, but only 20% see AI increasing revenue today, while 74% are still aspiring to it (survey of 3,235 leaders). The felt productivity is real. The output that reaches the P&L is a smaller, different number. Raw volume fails as a unit for three specific reasons:

• AI inflates the numerator artificially. More output is the default behavior of the tool, not evidence of return. Counting it rewards the inflation.
• It ignores quality. A doubling of drafts at half the accept rate is not a doubling of productivity. Volume treats a shipped brief and a rewritten one as identical.
• It ignores the redo cost. Every output that comes back for correction was counted once as "produced" and never subtracted when it was redone. The ledger only adds.

The correction: subtract the rework

Net productivity is what raw throughput becomes once you subtract the cost of the redo. AI rework is that cost: the human time and effort to correct, re-instruct, or re-do work an AI tool produced that looked finished but wasn't. It is the hidden denominator that turns apparent AI savings into a net-negative outcome — and it is the term almost every productivity claim omits. The clearest quantification of the symptom 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" — AI-generated output that looks polished but lacks the substance to advance the task — in the prior month. Each instance took an average of 1h 56m to resolve, which the researchers estimated at ~$186 per employee per month (self-reported, so directional rather than audited). Tellingly, the same study found only 28% of workers said AI improved the quality of their decisions. Volume up; quality flat. That gap is precisely what a net measure catches and a volume measure hides. Workslop is the borrowed name for the symptom (BetterUp/Stanford); AI rework is AIReturn's name for the measurable cost. You can have rework without anyone calling the artifact "slop" — an AI-drafted pull request that passes review and gets reverted a week later is rework all the same.

The perception gap: why volume feels like productivity

Net productivity matters because the people producing the volume genuinely feel productive — and the feeling is not the outcome. Rework is usually invisible to the person who created it: the effort moves downstream, to a colleague 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 gap between felt and actual speed. One caveat keeps this honest: METR later relabeled this specific result "historical" and redesigned its methodology in February 2026, so read 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 perceived and measured outcome. When perception says "faster" and measurement says otherwise, self-reported productivity is not a reliable input. A definition that leans on how productive a team feels will always overstate the return.

Output vs. rework: the same event, two ledgers

The contrast between raw throughput and net productivity is easiest to see as two ledgers reading the same week of work.

Net AI productivity, stated as a formula

Put the pieces together and the definition becomes a unit you can compute. Net AI productivity is output-per-dollar-net-of-rework — in plain language, cost per good outcome: the business output a team ships per dollar of AI spend, after subtracting the cost of redoing flawed AI work. Cost is the denominator; rework-adjusted output is the numerator. It is the same equation the pillar runs, scoped to the productivity question.

Net AI productivity = (outcome delivered − rework cost) ÷ (AI cost + human cost to get it right) — measured per team, continuously, against that team's own baseline. Three properties make this definition hold where raw throughput doesn't:

• The numerator is quality-gated, not gross. Only output that clears a quality bar counts — a merged change that survives ~14 days without a revert, a verified support resolution rather than a deflection, qualified or closed-won pipeline rather than a raw deal count, a spec-validated shipped decision — and reworked output is subtracted, not double-counted. Raw counts inflate under AI; the gate is what defends against that inflation.
• The denominator is real cost, granular. Not just tokens in aggregate — AI cost attributed by skill, by model, and by product, plus the human cost to get the output right. A precise denominator is what makes the ratio decision-grade rather than anecdotal.
• The comparison is the team's own history — never another team's. A support team's output isn't comparable to an engineering team's, so a cross-team leaderboard is noise. The only meaningful reading is this team, now, versus this team, before AI. That is why the honest measure is a baseline, not a benchmark.

Why the baseline must be per-team, never cross-team

The baseline is the reference point net productivity is measured against — and it has to be the team's own. Per-team baseline means each function is compared only to its own pre-AI level, because output and its friction are not comparable across functions. A resolved support ticket, a shipped design, and a merged pull request do not share units; averaging them into one "AI productivity" score produces a number that means nothing and hides the team quietly generating most of the rework. This is also why individual gains are a trap. Industry estimates put individual AI coding uplift at roughly 20–40% (Faros/DX, 2026, directional and vendor-adjacent). Even taken at face value, individual speed does not roll up to team output — not if the faster individual work generates rework that lands on the next person, or if the accepted-outcome rate falls. Net productivity is a team-level measure precisely because that is the level at which rework, cost, and outcome finally reconcile. Measured per person, the redo cost disappears into someone else's week; measured per team against its own baseline, it stays on the ledger. An honest limit belongs here, because this brand's credibility depends on it: the lag between AI-produced work and the rework it causes is not fully pinned in every function. In engineering it is concrete — code churn, reverts, reopened issues, and extra review rounds are already logged. For deals, support cases, and product initiatives, the proxy for rework and the separation of AI-caused from baseline rework is still being defined. The definition is strongest in engineering today and expands function by function as those proxies mature. Anyone claiming precise, universally validated rework attribution across every function in 2026 is overselling.

Why this definition is the one that gets funded

The reason to adopt this definition over the volume version is that the volume version is exactly what has stalled AI budgets. Only 28% of AI use cases fully meet their ROI expectations (Gartner, 782 I&O leaders, Apr 2026). The other roughly seven in ten are not failing because AI produced too little — most produced plenty. They are failing because "plenty" was measured as volume, and volume was never the return. A working definition of AI productivity has to survive contact with a skeptical CFO. "Output is up" does not; "output net of rework, per dollar, is up 11% versus this team's own baseline" does. The difference between those two sentences is the difference between activity and accountability — and it is the whole job of the framework CFOs use to measure AI ROI. To see the net number rendered as an actual deliverable, look at a sample per-team AI ROI report.

FAQ

How do you measure AI productivity? Measure useful output per dollar of AI spend, net of rework, against each team's own pre-AI baseline — not raw volume. Count only output that shipped without redo, subtract the cost of what came back for correction, and divide by AI cost plus the human cost to get it right. Volume and adoption are inputs; net output is the productivity number that maps to return. What is the definition of AI productivity? AI productivity is the useful output quality a team produces per dollar of AI spend, net of the cost of redoing flawed AI work, measured against that team's own baseline. In short: (outcome − rework) ÷ (AI cost + human cost to get it right), per team. It excludes hours saved, adoption, and raw throughput, all of which AI inflates without necessarily improving the work. Why isn't raw output a good measure of AI productivity? Because generative AI is built to raise output volume — more drafts, replies, and pull requests are the tool working as designed, not evidence of return. Raw output ignores quality and ignores the redo cost: an output counted once as "produced" is never subtracted when it's reworked. Deloitte found 66% of leaders report efficiency gains but only 20% see revenue impact (2026) — the gap volume hides. What's the difference between output and rework? Output is what a team produces; rework is the human cost of redoing the AI-assisted output that looked finished but wasn't. Net productivity is output minus rework. One study found 41% of workers received substandard "workslop," at ~1h 56m to fix each instance (BetterUp/Stanford, HBR, 2025). Counting the output while ignoring the rework is how high-volume AI use can still run net-negative. Should AI productivity be measured per person or per team? Per team, against that team's own baseline. Individual uplift (industry estimates put coding gains at roughly 20–40%, directional) doesn't roll up to team output if the faster work generates rework for the next person. Team level is where cost, rework, and outcome reconcile; per-person measurement lets the redo cost disappear into someone else's week, and cross-team comparison is invalid because functions don't share units.

Sources

• BetterUp Labs + Stanford Social Media Lab — "AI-Generated Workslop Is Destroying Productivity," Harvard Business Review, September 2025 (41%; 1h 56m; ~$186/employee/month; 28% decision-quality; n=1,150, self-reported). https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity
• METR — "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity," July 2025 (19% slower / 20% faster; result relabeled historical, method redesigned February 2026). https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
• Gartner — AI Projects in Infrastructure and Operations Stall Ahead of Meaningful ROI Returns (28% of use cases meet ROI, 782 I&O leaders), April 2026. https://www.gartner.com/en/newsroom/press-releases/2026-04-07-gartner-says-artificial-intelligence-projects-in-infrastructure-and-operations-stall-ahead-of-meaningful-roi-returns
• Deloitte — State of AI in the Enterprise 2026 (20% achieving revenue / 74% aspiring; 66% report productivity gains; n=3,235). https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/content/state-of-ai-in-the-enterprise.html
• Faros / DX — individual AI coding uplift ~20–40% (directional, vendor-adjacent; individual, not team-level), 2026. https://www.faros.ai/blog/lab-vs-reality-ai-productivity-study-findings

Related

• Start with the pillar: the framework CFOs use to measure AI ROI.
• Go deeper on the correction: why rework quietly erases AI ROI.
• The adjacent trap: why hours saved isn't AI ROI.
• The reference point: why the honest measure is a baseline, not a benchmark.
• Run it: our step-by-step framework for measuring AI ROI.
• See the output: a sample per-team AI ROI report.