Kihan: Authority before commitment

I’m a co-founder of Kihan, a new San Francisco technology company developing something we call Constitutional Computing.

We founded Kihan because we believe artificial intelligence has a serious problem that is becoming more dangerous as AI agents gain the ability to act independently: being capable of doing something is not the same as being authorized to do it.

An AI agent can now execute transactions, modify production systems, access sensitive corporate information, change permissions, and interact with dozens of enterprise applications. Its capabilities are expanding much faster than the mechanisms available to control what it is actually permitted to do.

That distinction matters enormously.

Most AI governance today relies on instructions, guardrails, access controls, or monitoring. These are useful, but they don’t fully address the fundamental problem. An agent can be instructed not to do something. It can be monitored while doing it. Its actions can be audited afterward. None of those measures, by itself, establishes whether a particular action was authorized before it happened.

And once an unauthorized transaction has been completed, a sensitive record altered, or confidential information exposed, an audit trail won’t undo the damage.

We started Kihan to change that.

Our approach is Constitutional Computing, a mathematically based architecture that separates what an AI agent proposes from what it has the authority to execute.

The principle is straightforward: an AI agent should never be able to grant itself permission to act.

Under our architecture, consequential actions must pass an independent authorization process before execution. The system evaluates the proposed action against established authority, determines whether it may proceed, must be held for approval, or must be refused, and preserves evidence of that decision.

We call this Authority Before Commitment™.

Our first product, Inktomi, is being developed to put this architecture into practice. It is designed to work alongside existing enterprise security infrastructure, rather than replace it, providing runtime authorization and evidence for AI agent activity.

We’re designing Inktomi for deployment on customers’ own infrastructure, with particular attention to environments using AI coding agents, the Model Context Protocol (MCP), and enterprise applications where autonomous actions can have real operational, financial, or regulatory consequences.

This is not another chatbot, an AI observability dashboard, or an attempt to make language models behave through increasingly elaborate prompts. We’re addressing the authority to act, independently of the model generating the proposed action.

Who is Kihan for?

Our initial focus is on organizations where mistakes have substantial consequences: financial services, healthcare, legal services, insurance, government, and other regulated industries.

We’re particularly interested in hearing from CISOs, security architects, enterprise AI developers, compliance professionals, and the engineers responsible for integrating AI agents into production environments.

But the underlying problem isn’t limited to regulated businesses. Any organization giving an AI agent access to tools that can change real-world systems will eventually confront the question of what that agent should be permitted to do, under whose authority, and how that authority can be demonstrated.

We believe this will become one of the defining architectural requirements for enterprise AI.

Here’s where we’d particularly value your feedback.

If you’re deploying AI agents today, how are you preventing them from taking actions beyond their intended authority? Are you relying on prompts, tool permissions, human approvals, or something else?
What is the first consequential action you would want an independent authorization system to control? Financial transactions? Code deployment? Access to sensitive information? Changes to production systems?
For enterprise developers and security teams, where would a mandatory authorization check create unacceptable friction? What would you need to see before trusting such a system in production?
If you work in a regulated industry, what evidence would your compliance or audit team require to demonstrate that an agent’s actions were properly authorized?

We’re interested in the difficult answers, not just the agreeable ones. If you think our approach has a weakness, we’d particularly like to hear about it.

Kihan is a young company, and Constitutional Computing is an ambitious undertaking. We’re building it because we believe that as AI becomes increasingly autonomous, the industry needs to establish a fundamental distinction between intelligence and authority.

An AI system may know how to do something. That doesn’t mean it has the right to do it.

That’s the problem we’re working to solve.

https://kihan.ai