The hidden costs of automating work

This Carnegie Mellon University research freaked me out:
Automation makes humans 17.7% SLOWER.
The time "saved" by AI gets eaten by verification, debugging, and fixing what the AI produced.

Meanwhile, augmentation—where AI assists specific steps—makes humans 24.3% FASTER.

When humans use AI to AUGMENT their work (like using copilot suggestions while coding), the workflow stays 76.8% aligned to what they would normally do. But, when humans use AI to AUTOMATE their work, only 40.3% of the approach is the same.

Automating work fundamentally changes it.
Not always for the better.

When agents finish automated tasks, their success rates run 32.5-49.5% lower than humans. They're optimized to look productive—making progress rather than actually completing steps correctly.

Even the costs don't work.
Cost per successful task:

Pure human $29.30
Augmentation $23.29 (20% savings)
Automation $35.60-$63.68 (21-117% MORE expensive)

The verification tax kills the savings.

Now here's the tension:
Pure AI (without verification) IS genuinely faster—11.7 minutes vs 100.
And genuinely cheaper—$0.94 vs $24.79 per attempt.
But it only succeeds 47% of the time.
So automation only wins when you can skip verification.

The best things to automate meet four conditions:
1 High volume (thousands of tasks)
2 Automated evaluation (can objectively detect failures)
3 Low stakes (failures don't matter)
4 Independent retries (can do over again until it’s right)

That's maybe 5%-10% of knowledge work.

For everyone else, the moment you add human verification, you're slower and more expensive than just doing it yourself.

The problem isn't the technology. It's that AI learns from documented processes, but experts work differently than processes suggest. There's always a gap between "how we say we do it" and "how we actually do it." AI automation forces everyone back to the documented way. The slow, rigid, wrong way.

This is why your AI pilot felt clunky. Why adoption stalled. Why people quietly went back to the old tools.
You automated the process manual, not the actual work.

The path forward:
Watch your best people work. Really watch them. Record their screens. Have them narrate their work. Build workflows from observation, not documentation.
Identify which steps are readily programmable—repetitive, rule-based, high-volume.
Let AI augment those specific steps. Keep humans in the driver's seat.
Then—only then—consider full automation for steps where quality tolerance is high and cost pressure is extreme.

Most companies skip straight to Automate Everything. Then wonder why it doesn't stick.
The winners are augmenting first, learning what works, then automating strategically.

With 90%+ cost savings on the table, we can't ignore AI.
But we also can't ignore the 28% speed penalty and 21-117% cost increase from bad automation.

Choose augmentation.
Learn your workflows.
Then automate deliberately.


—
Source: How Do AI Agents Do Human Work? Wang et al, Nov 2026