Key Takeaways

  • Molly Graham built her reputation on teaching fast-growing operators to give away their responsibilities like Legos, but she draws a hard line at delegating strategic judgment to machines.
  • Treating an LLM like an executive hire produces generic, low-grade outputs, because AI models lack taste and cannot define what quality means for your specific company.
  • Founders who ask AI to invent company strategy and paste the results into company channels are abdicating their core responsibility.
  • Lenny Rachitsky outlines a practical division of labor called the AI Sandwich (Human-in-the-Loop Workflow) to place human taste above and below automated execution.

The AI Sandwich (Human-in-the-Loop Workflow)

Graham spent a decade telling operators to hand off tasks as their startups expanded. When companies scale rapidly, hoarding responsibilities creates bottlenecks. Yet delegating to software is not the same as delegating to a trained colleague. Software does not possess taste, context, or moral accountability.

As Graham puts it: “There are some Legos that shouldn't be given away. There's just some work that shouldn't be outsourced.” She compares asking an LLM for strategic direction to asking a summer intern to run your product roadmap: “If you don't know the definition of good, how can you give it to your summer intern? If you don't understand what quality is, you can't outsource that yet.”

To prevent strategic drift and low-grade execution, Rachitsky recommends structuring work through three distinct layers:

  • Top Layer (Human Framing & Intent): The human defines the vision, strategic direction, problem framing, and criteria for what 'good' looks like. Do not ask AI to define the vision or strategy from scratch.
  • Middle Layer (AI Execution & Generation): AI agents and models handle the heavy lifting, code drafting, rapid prototyping, and data processing based on the human's constraints.
  • Bottom Layer (Human Review, Taste, & Accountability): The human conducts quality control, applies taste, checks for hallucinations or poor logic, and assumes full accountability before shipping the final product.

When This Works (and When It Doesn't)

This framework functions best when applied to high-stakes product design, engineering architecture, and executive communications where hallucinations or generic logic carry real business penalties. It forces you to supply the strategic bounds up front so the model produces useful assets rather than corporate boilerplate.

The framework fails if a founder uses it as an excuse to micromanage mechanical tasks or refuses to automate routine data pipelines. If a workflow has an objective, mathematically verifiable outcome (such as data formatting or syntax checking), you do not need human taste at the bottom layer. You only need an automated test suite.

What to Do With This

Take the next product brief or investor update you need to draft this week and run it through the sandwich structure:

1. Spend twenty minutes writing the top layer yourself. Bullet out the hard choices, the customer problem, your trade-offs, and your definition of an unacceptable outcome. Do not open an LLM prompt box until this framing exists on paper.

2. Paste your raw framing into the model and instruct it to draft the complete document, organize the data tables, and generate three counter-arguments against your thesis.

3. Pull the output back into your editor. Strip out the generic filler, verify every factual claim, sharpen the tone with your own voice, and sign your name to the final draft.