The BCG study results:

• 758 consultants at Boston Consulting Group, randomized into AI and non-AI groups
• AI group: +12.2% tasks completed, 25.1% faster, 40% higher quality
• Biggest effect on below-average performers — AI compressed the skill distribution
• Tasks inside the AI frontier (where AI is capable) saw massive gains
• Tasks outside the frontier saw decreased performance when people over-relied on AI

Two integration models:

Centaur approach:

• Humans and AI handle separate tasks based on comparative advantage
• Human does the strategic thinking, AI handles the execution
• Requires clearly defined task boundaries
• Works well when you can cleanly separate "AI tasks" from "human tasks"

Cyborg approach:

• AI is woven into every step of the workflow
• Human and AI collaborate sentence by sentence, decision by decision
• More fluid and harder to systematize, but produces better outcomes
• The human is constantly steering, editing, and redirecting the AI

Key insight: AI changes which tasks are worth doing.

• Tasks that took hours (first-draft analysis, market research summaries, code boilerplate) now take minutes
• This doesn't just save time — it frees cognitive budget for higher-value work
• The competitive advantage shifts from "who can do the task" to "who knows which tasks to do"

The over-reliance risk:

• When AI encounters a task outside the frontier, people who rely on it too heavily produce worse output than if they'd worked alone
• The "falling asleep at the wheel" problem — human oversight degrades when AI is usually right
• Critical to maintain the ability to evaluate AI output independently