Skip to content
EN DE

From Chat to Delegation: The Autonomy Spectrum

L4 Lesson 1 of 5 — AI as Coworker
1
2
3
4
5
Level 4

AI as Coworker

Delegation, trust calibration, and compliance — the step from answers to autonomous work.

In L3, you learned to build persistent context — AI that knows you and your work. Now comes the question that follows: If AI knows your context, what can you hand over?

The transition from “AI answers my questions” to “AI does tasks for me” isn’t binary. It’s a spectrum — and navigating that spectrum consciously is the core skill of L4.

Agentic AI AI systems that can independently execute multi-step tasks — they plan steps, use tools, and make decisions without the human directing every move. operates on a spectrum from full human control to autonomous execution:

LevelYour RoleWhat AI DoesExample
L0: OperatorYou control everythingAI responds on demand”What’s the ROI of this campaign?”
L1: CollaboratorYou decide, AI suggestsAI makes proposals, you chooseCopilot suggests code, you accept
L2: ConsultantYou delegate subtasksAI works independently on subproblems”Research competitor X and summarize”
L3: ApproverYou approve resultsAI delivers complete proposalsAI drafts email, you review and send
L4: ObserverYou monitorAI acts autonomouslyAutomated reports, scheduled tasks

Most knowledge workers today operate between L0 and L2. The leap to L3 and L4 is happening right now — with tools like Claude Cowork and ChatGPT Agent Mode.

The autonomy level depends not on what the AI can do, but what you allow it to do. Using a highly capable model at L0 is just as valid as using a simpler one at L3 — if the task requires it.

Task Decomposition: The Art of Breaking Things Down

Section titled “Task Decomposition: The Art of Breaking Things Down”

Delegation works best when you break tasks into clear subtasks. That’s not an AI skill — it’s yours.

Example: Creating a Quarterly Presentation

Section titled “Example: Creating a Quarterly Presentation”

Weak: “Create the Q3 presentation.”

Strong — with decomposition:

StepTaskDelegable?Level
1Export quarterly data from the CRMYes (tool)L2
2Identify trends and anomaliesYes (with review)L3
3Define storyline and key messagesNo — that’s your strategyL0
4Create slidesYes (draft)L3
5Final review and adjustmentsNo — your name’s on itL0

The pattern: Delegate data work and drafts. Keep strategy and final accountability.

Research from Harvard Business School with 758 BCG consultants (Dell’Acqua et al., 2023) identified two fundamental modes of human-AI collaboration:

Like a centaur — half human, half horse — with a clear dividing line:

“This part is mine. That part is the AI’s.”

Works well for: Clearly separable subtasks, recurring processes, when traceability matters.

Example: You write the core arguments for a presentation, the AI formats the slides and creates data visualizations.

Human and AI work intertwined on the same task — like a conversation:

“Draft → Feedback → Revision → Feedback → Polish”

Works well for: Creative work, iterative processes, when the result emerges through dialogue.

Example: You write a paragraph, AI expands and revises, you adjust the tone, AI tightens — until the text is right.

SituationRecommended Mode
Clearly defined subtasksCentaur
Creative, exploratory workCyborg
High traceability requirementsCentaur
Tight deadline, result needed fastCyborg
You’re new to the topicCyborg (AI as sparring partner)
You’re the expertCentaur (AI as assistant)
  • Data work: Summarizing, formatting, analyzing, comparing
  • Drafts: First versions of emails, reports, presentations
  • Research: Gathering information, structuring, finding sources
  • Routine tasks: Recurring formats, templates, status reports
  • Strategic decisions: What should the company do next?
  • Relationship work: Difficult conversations, sensitive communication
  • Unclear tasks: When you don’t know what the result should be
  • High-consequence decisions: Contracts, legal documents, financial decisions — AI can assist, but you decide

Can I verify the result in under 2 minutes?

Yes → Delegation is worthwhile. No → Delegation doesn’t save time, because verification is as much work as doing it yourself.

The Harvard/BCG study with 758 consultants and GPT-4 delivered clear results:

  • +40% higher quality on tasks within AI capabilities
  • 25% faster on suitable tasks
  • But: On tasks outside AI strengths, quality dropped — because consultants trusted the AI output without checking

The researchers call this the “Jagged Frontier” — the boundary of what AI does well is irregular and hard to predict. Some tasks that seem complex, AI handles easily. Others that seem simple, it gets wrong.

The consequence: You need to experiment to learn where the frontier runs for your tasks.

Take a task you need to complete next week. Break it into 5–7 subtasks. For each: Delegable? If yes, at which autonomy level?

Complete the same task twice: once in Centaur mode (clear division), once in Cyborg mode (intertwined). Note: Which was faster? Which produced higher quality? Which felt more natural?

Give the AI 5 different tasks from your daily work. Rate each result: Usable without changes? Usable with edits? Unusable? After a week, you’ll have a personal map of AI capabilities for your work.

Delegation isn’t laziness — it’s a competency. The best knowledge workers won’t be those who do everything themselves, but those who know what to delegate, at which autonomy level, and how to verify results efficiently.

In the next lesson, you’ll learn about Claude Cowork — Anthropic’s desktop agent that makes delegation at L3 and L4 concrete.

Part of AI Learning — free courses from prompt to production. Jan on LinkedIn