Every Manager Is Now an AI Manager

Ethan Mollick's Equation of Agentic Work — and what it looks like in practice

Based on "The Equation of Agentic Work" by Ethan Mollick, Wharton School — from his Substack One Useful Thing

This site was built in a single AI-assisted session to demonstrate the concepts it describes.

Scroll

Everything on this page comes from Ethan Mollick

Ethan Mollick is an Associate Professor at the Wharton School of the University of Pennsylvania, where he studies the effects of artificial intelligence on work, entrepreneurship, and education. He co-directs Wharton's Generative AI Labs and was named one of TIME Magazine's Most Influential People in Artificial Intelligence.

His Substack newsletter, One Useful Thing, has over 390,000 subscribers and has become one of the most widely read sources on the practical implications of AI for knowledge work. This entire site is built on ideas from a single article in that newsletter.

Everything that follows — the equation, the delegation framework, the three levers — is Mollick's thinking applied to a real-world system. We built this site to show what his ideas look like when you actually implement them.

Dozens of open tasks. Zero system.

Knowledge workers manage complex portfolios of concurrent work across email, documents, portals, and conversations. The traditional approach: memory, inbox scanning, and mental prioritization every morning.

Email
📄Documents
💻Portals
📞Phone
📅Calendar
💬Chat
📋Spreadsheets
📑Notes
🤔Memory

The old way

Without a system, everything lives in your head and your inbox. The most important thing competes for attention with whatever arrived most recently.

  • Priority is whatever feels urgent right now
  • Context lives in scattered email threads
  • Handoffs require verbal walkthroughs
  • Deadlines surface when they're already close
  • Nothing auto-escalates — you have to remember

When does delegation pay off?

Mollick defines three variables that determine whether delegating to an AI agent (or any agent) is worthwhile.

Variable 1 Human Baseline Time
vs.
Variable 2 Probability of Success
×
Variable 3 AI Process Time

You're trading "doing the whole task yourself" against "paying the overhead cost per attempt" potentially multiple times until you get something acceptable. The math favors delegation when Human Baseline Time is large and AI Process Time is small relative to Probability of Success.

Try the math yourself

7h
72%
Expected Attempts
1.4
Expected AI Time
1.4h
Time Saved
5.6h
Delegate — expected savings of 80%

Every field already has delegation paperwork

Effective AI use isn't about clever prompt engineering. It's about the same documentation and communication skills that every field has independently invented for human delegation.

Software Product Requirements Documents
Film Shot lists
Architecture Design intent documents
Military Five Paragraph Orders
Consulting Engagement scopes with deliverable specs

All of them answer the same six questions:

Delegation Question What It Accomplishes
What are we trying to accomplish, and why? Aligns the agent on purpose, not just mechanics
Where are the limits of delegated authority? Prevents the agent from making decisions outside their scope
What does "done" look like? Defines the acceptance criteria
What specific outputs do I need? Makes deliverables concrete and verifiable
What interim outputs do I need to follow your progress? Enables oversight without micromanagement
What should you check before telling me you're finished? Builds quality control into the process

When these are well-specified, both human and AI agents are far more likely to succeed — raising the Probability of Success and lowering AI Process Time.

How to shift the equation in your favor

Mollick identifies three ways to make delegation consistently worthwhile. All three are improved by subject matter expertise.

1

Better Instructions

Clear goals and context that the agent can execute on with a higher chance of succeeding the first time.

Raises Probability of Success
2

Better Evaluation

Faster, more accurate feedback loops. Fewer attempts needed to get the right output because you catch problems early.

Lowers Effective AI Process Time
3

Cheaper Evaluation

Faster to determine if the output is good or bad without deep review. Pattern recognition from domain expertise.

Lowers AI Process Time Per Attempt

What this looks like in practice

An insurance advisory practice managing complex client portfolios across multiple carriers. Dozens of open engagements at any time, each at a different stage, each with different deadlines and stakeholders.

✈️

The Account Manager (Pilot)

Owns the client relationship. Sets priority based on judgment, context, and experience. Has final authority over every decision. The system recommends — the pilot decides.

📡

The Flow Manager (Air Traffic Control)

Sees the full queue. Cross-checks human priorities against computed priorities. Surfaces conflicts and ensures nothing falls through the cracks. Manages the system, not the clients.

The Task — Routing Slip
Client Name Task Type Owner Status Priority (human-set) Next Action Task Number (search key) Recommended Priority Activation Step Action Briefing
The Work Unit — Flight Computer
Hard Bound Date Next Review Date Declared Priority Lifecycle Status Action Briefing Notes (running log) Effective Priority (computed) Activation Step (computed) Data Hygiene Flag (computed)
Email
📁
Documents
📊
Database

The Task is where team members work — a simple routing slip that says who, what, and where. It points to where the real work lives: email threads, documents, carrier portals.

The Work Unit runs independently in the background. It computes priority from dates alone, ignoring human judgment entirely. This is intentional — the point is to create an independent signal that can be compared against human assessment.

When human and system disagree, that's the signal

The system's highest-value moment is when the Task's human-set priority and the Work Unit's computed priority disagree.

System: Urgent
System: Not Urgent
Human: Urgent
Aligned
High urgency, everyone agrees. Take action now.
🟡
De-escalation Signal
Human feels urgent but dates say there's time. Is the human seeing something the system can't?
Human: Not Urgent
🔴
Escalation Signal
Deadline approaching but human hasn't prioritized it. Did they forget, or is there context the system doesn't know?
Aligned
Properly parked. Both agree it's not urgent right now.

Tasks auto-escalate as deadlines approach

The Work Unit computes priority from dates alone. As a hard deadline approaches, priority automatically escalates through four bands — without anyone touching it.

Must Do
≤ 10 days
Should Do
≤ 29 days
Could Do
≤ 44 days
Won't Do
> 44 days
45 days out
Current Priority: Won't Do

The pilot always has final authority. If the system says "Must Do" but you know the deadline moved, override it. If it says "Won't Do" but your gut says otherwise, investigate. The disagreement is the signal — not the answer.

"I don't know exactly what work looks like when everyone is a manager with an army of tireless agents. But I suspect the people who thrive will be the ones who know what good looks like — and can explain it clearly enough that even an AI can deliver it."

— Ethan Mollick, "The Equation of Agentic Work"

Management skills ARE AI skills

The skills historically dismissed as "soft" — communication, delegation, evaluation, judgment — are the hard skills of the agentic era.

The bottleneck is human clarity

The bottleneck is not AI capability. It's human clarity about what they want. An expert who can scope, specify, and evaluate will outperform a prompt engineer every time.

The system grows with you

Start simple — keep your tasks accurate. Add computation as patterns emerge. The system described here started as a spreadsheet and evolved into a priority engine. Yours will, too.

That's the theory. Here's what we actually built.

Signature Personal Insurance implemented these concepts using four tools: Airtable, Gmail, Google Drive, and Claude. The co-pilot structure page shows exactly how the pieces connect — real workflows at signaturepersonalinsurance.com.

✈️
The Co-Pilot Structure
Four tools, one system. How Airtable, Gmail, Google Drive, and Claude work together as a real-world delegation framework.
📊
Commission Reconciliation
How Claude processed thousands of lines of commission statements across a full year of reports. Coming soon.
Coming Soon