The Document-First Principle for Enterprise AI

Advice is different from execution

There's a meaningful line between an AI system that tells you what to do and one that goes and does it. Much of the current excitement about "agentic AI" is really excitement about erasing that line -- letting a model not just recommend a change but make it, commit it, deploy it. That's a legitimate direction for some tools. It is also, for a lot of enterprise work, exactly the wrong default.

The alternative is what we've come to think of as the document-first principle: an AI assistant's job, for anything consequential, is to understand the situation, investigate the evidence, compare options, and produce a reviewed artifact a human can act on -- not to reach into a live system and act on it directly. Advice, however well-reasoned, is not the same thing as execution, and treating them as interchangeable is where a lot of AI incidents actually come from.

Plans create a review boundary

A plan sitting in a document is inherently reviewable in a way that an autonomous action already taken is not. Before anything changes, there's a moment where a human can read what's proposed, ask why, push back, or simply say no -- and that moment only exists because the AI stopped at "here's what I'd do" instead of continuing on to "and I did it."

This isn't a defensive design because AI is untrustworthy in some special way. It's the same reason a junior engineer's pull request gets reviewed before it merges, or a proposal goes to a committee before budget moves. The review boundary is what makes an action accountable rather than merely explainable after the fact.

Humans remain accountable

Somebody has to own the consequences of a change, and "the AI decided" is not an answer anyone actually wants to give when something goes wrong. Keeping a human in the loop for consequential actions isn't a limitation bolted onto the AI to make people feel better -- it's a straightforward statement of where accountability actually lives. The AI can do excellent analysis. It cannot be held responsible for what happens next. Only a person can be, so a person should be the one who decides.

Documents preserve rationale

A side benefit of document-first design that's easy to undervalue: the artifact itself becomes a record of why, not just what. Six months later, "why did we do it this way" is a much easier question to answer when there's a document that laid out the evidence, the alternatives considered, and the reasoning behind the choice -- rather than only a diff or a changelog entry that shows the result with none of the thinking behind it.

That rationale is worth almost as much as the decision it supports. It's what lets the next person -- human or AI -- pick the work back up without re-litigating settled questions from scratch.

Autonomous action should require explicit authority

None of this is an argument that AI should never act directly on a system. It's an argument that it shouldn't do so by default, silently, or by accident. Direct action is a capability that should be explicitly granted, scoped, and visible -- not something that falls out of an assistant simply being capable enough to attempt it. The default posture for anything with real consequences should be "produce something a human reviews," with direct execution reserved for cases where that authority has been deliberately and narrowly extended.

The boundary can expand gradually as trust is earned

Document-first isn't meant to be a permanent ceiling on what AI is allowed to do -- it's a sensible starting boundary that can move as trust is actually earned, one narrow capability at a time, rather than all at once because a model demo looked impressive. An organisation that's watched an AI system produce consistently sound, well-evidenced plans in one specific area has a real basis for cautiously extending it more direct authority there. An organisation that hasn't watched anything yet does not.

Start with advice. Let plans do the work of building trust. Expand execution deliberately, narrowly, and only where the evidence supports it. That ordering -- not the reverse -- is what makes AI in serious operational settings something people can actually rely on, instead of something they have to constantly double-check.