A partner at a top consulting firm told me recently that her team shipped a client deliverable in four hours that would have taken two weeks a year ago. An agent pulled the data, built the model, drafted the deck. She reviewed it, changed six slides, and sent it. Then she said something that stuck with me: "I have no idea if I did good work. I don't know what I did."
That sentence is the real story of 2026, more than any benchmark or funding round. We've spent two years measuring what agents can produce. We have not figured out what a human is supposed to measure about themselves once the agent produces it.
The instinct is to keep the old scorecard and just raise the numbers on it. Slides per hour. Lines of code shipped. Client emails answered. If an agent gives you 10x output, surely the KPI is 10x output. This is the mistake nearly every organization I talk to is making right now, and it's why so much agentic AI adoption is quietly producing exhaustion instead of leverage. Forrester's read on the market this year calls it exactly that: companies chasing agentic AI, few actually catching it. The chasing usually looks like volume metrics stapled onto a workflow that fundamentally changed underneath them.
Here's the reframe that the good teams are landing on instead, and it's simpler than it sounds: when an agent handles first-pass execution, the human's job stops being production and becomes judgment. Not "did you make the deck" but "was the deck right." Not "how much did you generate" but "how well did you decide what to keep, what to cut, and what to push back on." One framework circulating among engineering leaders this year backs this up with a number worth sitting with — teams that shifted their KPIs from raw AI-touched output toward decision quality and rework rate saw cycle times drop by roughly 20 to 55 percent, while teams that kept celebrating volume alone saw rework creep past 15 percent, meaning the "10x" was partly an illusion that got quietly repaid in cleanup.
This is not a minor metrics tweak. It's a different theory of where value gets created. For a century of knowledge work, effort and output were roughly the same thing, so we measured output as a proxy for effort and called it a day. Agents break that proxy completely. Output is now cheap and abundant. What's scarce — what was always actually scarce, we just didn't have to notice — is the judgment that decides which output deserves to exist in the world. McKinsey's research on AI trust this year makes a version of this point about the agentic era specifically: as agents take on more autonomous execution, trust and value concentrate not in the volume of what agents do but in the quality of the human oversight wrapped around it.
Which brings me back to that partner's sentence. She was wrong that she didn't know what she did. She reviewed six slides out of what might have been forty, and in choosing which six, she was exercising exactly the skill that remains irreplaceable: knowing what "right" looks like for this client, this moment, this set of stakes. That's not a soft skill. That's the entire discipline of consulting, distilled down to its least automatable core, because it requires having sat in enough rooms, made enough mistakes, and absorbed enough context that you can recognize a wrong answer even when it's dressed up beautifully.
The practical shift this demands is uncomfortable for a lot of managers, because it means you can no longer proxy someone's contribution by watching how busy they look or how much they produce. You have to actually evaluate the six decisions instead of the forty slides. That's slower to assess, harder to standardize, and much more honest. It also means junior people, who used to build judgment slowly through years of doing the grunt work themselves, now need a deliberate substitute — structured review, exposure to real stakes, mentorship that specifically targets decision-making rather than task execution. Skip that, and you get a generation of professionals who can direct agents fluently but have never developed the judgment to know when the agent is wrong.
Ownership of outcomes was never really about who typed the words or built the model. It was always about who was willing to put their name behind the decision. Agents just made that the only part of the job left to measure.
