The Rework Tax

The Rework Tax

Status
Draft
Author
Created
Sep 17, 2026
Tags
Agentic AI
Accountability
Knowledge Work
Text
The real cost of AI-augmented work isn't measured in output. It's measured in what humans have to fix.
A colleague told me last month that her team's agent now drafts client decks "at 10x speed." She said it like a scoreboard number, the way you'd report a sprint velocity. I asked the obvious follow-up: how much of that time did she then spend fixing what the agent got subtly wrong? She paused. Nobody had measured that.
That pause is the real story of AI-augmented work in 2026, not the 10x.
The myth: output is the metric
The dominant way we talk about agentic AI in consulting and knowledge work is volume. GitClear's 2026 analysis found developers using AI tools now author four to ten times more code during peak usage weeks than non-users. Leadership decks cite numbers like this constantly, and they're not wrong, exactly. Agents genuinely do produce more artifacts per hour than a human alone ever could. But "more artifacts" quietly became a stand-in for "more value," and that substitution is where the trouble starts. A 10x increase in drafts, slides, or code commits tells you almost nothing about whether the work is right, whether it's the right work at all, or whether someone downstream is now paying for it in ways that don't show up on the same dashboard.
The reality: the cost hides downstream
There's a better number quietly emerging in how mature AI organizations track performance, and it's the inverse of output: the rework tax. It's the human time spent catching, correcting, or quietly redoing what an agent produced. Teams that actually measure it treat 30% of time saved as a rough ceiling — cross that, and the agent isn't augmenting the human, it's generating work for the human. I'd guess most consulting teams running agents today have no idea what their rework tax is, because nobody's incentivized to measure it. The 10x number goes in the client update. The three hours a senior associate spent re-checking the agent's assumptions before anyone could trust the deck does not.
This isn't a technology problem. It's an accounting problem. We built a scoreboard that counts what agents produce and stays silent on what humans have to catch. And the people best positioned to catch it — the ones with the pattern recognition to spot a plausible-sounding but wrong assumption, a citation that doesn't quite hold up, a recommendation that ignores the client's actual politics — are exactly the people whose contribution disappears from the metric entirely. Their value shows up as an absence: the errors that never made it to the client. Absences don't photograph well in a KPI review.
The 2026 enterprise AI conversation has started circling a version of this same tension from the governance side: firms are grappling with who is accountable for an agent's actions, and the consulting partners actually differentiating themselves aren't the ones with the flashiest automation — they're the ones doing the unglamorous work of post-deployment accountability, tracking whether an agent's output actually holds up against a business outcome over time, not just at the moment of generation. That's the same instinct, applied upstream. If you can't say who is accountable for an agent's mistake, you almost certainly aren't measuring how often it makes one.
What this asks of you if you're running or working inside one of these teams is a small discipline shift, but a real one. Stop reporting output multiples without a rework denominator next to them. When someone says "the agent drafted this in ten minutes," the next question is always "and how long did it take someone to make it trustworthy?" If you don't know, you don't actually know whether you've gained capacity or just relocated the work to a less visible stage of the process, done by someone whose time isn't being tracked as carefully as the agent's.
The consulting firms that will earn the premium in the next few years won't be the ones who can show the biggest velocity numbers. They'll be the ones who can show their rework tax going down over time — proof that the judgment layer, the human layer, is getting sharper at working with the machine rather than constantly cleaning up after it. Velocity is easy to claim. A shrinking rework tax is earned, one caught error at a time, and it's the only version of "10x" that actually means the work got better.