The Human Sandwich
Set the direction, let AI execute, then refine the result.
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 4
- Confidence
- 96%
The Human Sandwich places AI between two deliberate human interventions. A person begins by defining the creative intent, desired outcome, and relevant constraints. The AI then performs the labor-intensive middle work, such as finding reviews, drafting an article, comparing products, or generating alternatives. Finally, the person inspects the result and adjusts it using preferences or context the model missed, such as budget, urgency, tone, or quality requirements. The mechanism preserves human agency while shifting research and production effort to the model. It works as an iterative loop when necessary: each review supplies better constraints for another AI pass until the output is accurate, useful, and personally appropriate.
Origin
Ben Tossell described the model on Marketing Against The Grain while explaining how people could remain in control as AI performs more of their work.
Core principles
- 01Humans supply intent and judgment.
- 02AI handles research and production work.
- 03Human review protects quality and personal fit.
- 04AI output is a draft or recommendation, not the final authority.
How to run it
- 1
Set the creative direction
State what you want to achieve and provide the initial idea, question, or task. Include the constraints that materially affect a good result.
Pro tip Specify the audience, desired format, budget, timing, and quality threshold up front.
Watch out A vague starting point forces the AI to fill important gaps with assumptions.
- 2
Delegate the middle work
Ask the AI to research, compare, draft, or generate the requested result. Let it handle the time-consuming production stage while keeping the conversation available for follow-up instructions.
Pro tip Break complex work into visible stages so errors can be caught before they compound.
Watch out Do not treat confident output as verified evidence.
- 3
Apply human judgment
Review the result for correctness, fit, taste, and practical constraints. Identify anything that conflicts with your actual preferences or circumstances.
Pro tip Explain why a recommendation is wrong so the next revision incorporates a stronger decision rule.
Watch out Skipping review turns assisted work into unaccountable automation.
- 4
Refine and approve
Give the AI targeted corrections and request a revised result. Repeat until the output meets your standard, then make the final decision yourself.
Pro tip Change one major constraint at a time when you need to understand its effect.
Watch out Endless refinement can cost more time than accepting a sufficiently good result.
In the wild
A podcaster asks AI to compare microphones used by similar shows. The AI gathers options and reviews, but the podcaster rejects the cheapest recommendation, explains that studio quality matters more than price, and adds a requirement for same-day delivery. The AI revises the shortlist around those human constraints.
→ The podcaster saves research time while retaining control over cost, quality, and urgency.
A marketer supplies the thesis, audience, and desired tone for an article. AI researches supporting material and writes a draft. The marketer then corrects unsupported claims, removes generic passages, and reshapes the conclusion around firsthand experience.
→ AI accelerates production while the human preserves credibility and distinctiveness.
Common mistakes
Removing the final human review
The model depends on human judgment to catch incorrect recommendations, missing context, and unwanted trade-offs. Accepting its first output defeats the framework's control mechanism.
Giving no meaningful constraints
The AI cannot account for preferences it has not been told, such as urgency, budget, or quality. Generic inputs therefore produce generic or poorly fitted outputs.
Using AI as the decision owner
AI can gather and synthesize options, but the human remains accountable for the final choice. Delegating accountability creates risk when the model's evidence or reasoning is weak.
Is it for you?
Best for
It is best for people delegating research, writing, comparison, or creative production to AI.
Not ideal for
It is not ideal for decisions that cannot be safely reviewed or corrected by a knowledgeable human.
From the transcript
“we're all sort of gonna become directors, where we start the input with a creative thought, and then AI does all the work, finds the…”
“I assume there's always gonna be a touch point of human, AI, human, and that will probably be the interface.”
“humans are always gonna be in the loop, and it's gonna be a human sandwich, I suppose, in that sense of using AI.”
From the episode
This OpenAI Update Is Going To Replace Google Search with Ben Tossell
Ben Tossell