Harvard Just Proved It: AI Agents Cut 87% of Work Time and 94% of Costs — Here Is What It Means for the Map of AI

Research Analysis — A new study from Perplexity and Harvard Business School just quantified what we have been arguing with the Builder-PM framework: AI agents do not just assist workers. They replace the execution layer entirely.

The Numbers That Change Everything

The study (arxiv.org/abs/2606.07489) analyzed 84,000+ AI agent sessions across 18 domains, comparing Perplexity Compute — as explored in the economics of AI compute infrastructure — r (an autonomous agent) against traditional AI search (a conversational assistant). The findings:

The Operator-to-Supervisor Shift

The study’s most important finding is not the speed gain. It is the role transformation. From the paper:

“Users spend less time operating the workflow and more time specifying goals, supplying context, checking outputs, and asking for extensions.”

This is the Builder-PM transition, empirically validated. The human role shifts from executor (doing the work) to supervisor (directing the agent, verifying outputs, refining goals). The study found:

• 76% of agent tasks involve higher-order cognition (vs. 55% for traditional AI)
• 50% of agent tasks are at the “Create” level of Bloom’s Taxonomy (vs. 26%)
• 71% of agent work is abstract/non-routine (vs. 53%)

The agent handles the routine execution. The human handles the judgment. This is exactly what the Judgment Layer framework predicts: AI replaces execution, but judgment — the ability to decide what to build, what to prioritize, what to ship — becomes the scarce resource.

Scope Expansion: Workers Become Generalists

The second major finding: agents make workers broader, not just faster.

• 59% of agent queries cross occupational boundaries (vs. 50% with search)
• Agent tasks require 2.40 knowledge domains on average (vs. 1.74) — a 38% increase
• 51% of agent tasks require 3+ domains (vs. 17%)
• Work activities are 32-60% broader with agents

The implication for the Map of AI: as agents expand what individuals can do, the orchestration layer (Layer 8) becomes more valuable than ever. The person who can direct agents across multiple domains — the Builder-PM — replaces teams of specialists.

What This Means for Every Company

This study provides the empirical foundation for three structural shifts we have been tracking:

1. Builder-PM replaces traditional PM — If agents cut 87% of execution time, the person managing the workflow does not need to know how to execute. They need to know what to build and why. That is the Builder-PM.
2. Harness Theory applies to teams, not just companies — A single person with an agent harness can do the work of a 5-person team. The study shows 25x median speedup and 94% cost reduction. Companies that restructure around agent-augmented individuals will outperform those clinging to traditional team structures.
3. The Overhang is releasing — 87-96% cost reductions across 18 domains means the AI productivity overhang is not theoretical anymore. It is empirical. The companies that absorb this data and restructure will pull ahead. The ones that do not will face the same cost structure competing against companies with 94% lower costs.

The age of the AI assistant is over. The age of the AI agent is here. The data is in.

Read the frameworks:
Builder-PM — The New Role Replacing Product Management
The Map of AI — 9 Layers of the AI Economy
Harness Theory — How Non-AI Companies Win
Product Overhang Doctrine

Source: “How AI Agents Reshape Knowledge Work” — Perplexity + Harvard Business School (June 8, 2026). Full paper: arxiv.org/abs/2606.07489