Sarathi Principle
Delegate judgment. Retain authority.
Delegate reasoning, never accountability.
The Sarathi Principle is the operating principle at the heart of the Charioteer Framework. It explains how a human and an AI should share decision-making when both can contribute meaningful judgment, while human intent, authority, and accountability remain explicit.
The Sarathi
In the chariot metaphor, the passenger or warrior determines the destination and purpose. The sarathi, or charioteer, reads the terrain, directs the horses, and makes the many local decisions required to move safely and effectively. The two roles are different, but neither is merely decorative.
Charioteer applies that relationship to human-AI collaboration. An AI should not be treated only as a command executor that waits for every small instruction. A capable model can interpret context, identify options, make a plan, surface missing information, and suggest a better course when the human’s current view is incomplete.
That contribution deserves respect. It does not require the AI to become the final authority.
Two Kinds of Decisions
The Sarathi Principle separates decisions by their nature, not by a simplistic rule that the human always decides and the AI always executes.
| Decision area | Primary responsibility | Examples |
|---|---|---|
| Purpose and stakes | Human | Define the goal, decide what success means, set acceptable risk, and choose what must not happen. |
| Bounded judgment | AI | Interpret context, compare options, plan a sequence, detect inconsistencies, and recommend the next move. |
| Authority and enforcement | Human and system | Grant permissions, approve consequential actions, enforce policy, and set budgets and stopping conditions. |
| Execution and recovery | System, with human oversight | Record durable state, perform authorized actions, provide audit trails, and make intervention or recovery possible. |
The AI can make many micro-decisions that would otherwise slow the human down: how to structure a draft, which information to retrieve first, how to sequence an investigation, or when a plan needs revision. The human stays focused on the larger objective, the constraints, and the trade-offs that carry real consequences.
Collaboration, Not Deference
A useful sarathi does more than follow instructions mechanically. When the terrain changes, the charioteer notices. When an approach is weak, the charioteer can say so. Likewise, an AI collaborator should be allowed to question assumptions, explain uncertainty, and propose a correction when the available evidence points elsewhere.
The human should not outsource judgment blindly. They should use the AI’s perspective as an additional source of reasoning, then decide how much of that reasoning to accept. This makes the relationship collaborative rather than deferential in either direction.
Accountability Does Not Move
Delegating a decision is not the same as delegating responsibility. The human who sets a system in motion remains accountable for its purpose, its permissions, and the consequences of its use.
That accountability is not a reason to reduce the AI to a passive assistant. It is a reason to design the collaboration well: make the boundaries clear, give the AI enough context to contribute meaningfully, and preserve the records and controls needed to understand what happened.
Applying the Principle
Use the Sarathi Principle when designing a human-AI workflow:
- Define the human’s goal, authority, and non-negotiable constraints.
- Identify the local decisions where AI judgment can improve speed, breadth, or quality.
- Give the AI the context and tools required to make those decisions well.
- Make consequential permissions, durable state, and recovery paths explicit in the system.
- Expect the AI to challenge weak assumptions, but keep the human responsible for the final direction.
The aim is neither total human micromanagement nor unchecked AI autonomy. It is a partnership in which a human can make better use of AI capability while remaining clearly in charge of the system’s purpose and responsibility.