AI for Documentation — What It's Good At and What It Isn't

AI is genuinely useful for documentation in ways that differ from its usefulness in other writing forms. Documentation writing is often blocked not by lack of ideas but by the friction of converting tacit knowledge — the expertise a practitioner carries in their head — into explicit, structured, written form. AI dramatically reduces this friction. A practitioner who can describe a process verbally can now use AI to convert that description into a structured first-draft SOP in minutes. The expertise that would have taken days to document can be captured in hours.

This is a real acceleration. It also creates a real risk: documentation that looks complete and is inaccurate. AI that generates documentation from a practitioner's description will fill gaps in the description with plausible-sounding content that the practitioner may not notice is invented rather than accurate. The SOP that includes a step the practitioner did not describe but that sounds like it should be there is an SOP with an error that will persist until a practitioner encounters it in practice — potentially after following it incorrectly.

The useful AI documentation workflow has three phases. Phase one: capture. The practitioner describes the process — verbally, in notes, in a rough outline. Phase two: structure. AI converts the raw description into the documentation template, filling in what was described and flagging what was not. Phase three: verify. The practitioner reviews the AI-generated document against the actual process, correcting errors, filling gaps, and confirming that the procedure as written would produce the intended outcome.

The key insight is that AI owns the conversion step — taking messy human description and producing clean structured documentation — not the verification step. Verification requires the practitioner to engage with the document at a level of specificity that catches errors. The practitioner who skips verification because the AI output looks clean has introduced unverified content into the documentation system.

A specific capability worth developing: using AI for documentation gap analysis. Once a procedure or playbook is drafted, prompt AI to identify situations the current documentation does not cover. 'Here is an SOP for our customer onboarding process. What customer situations or system states does this procedure not address?' The AI will generate a list of gaps — some obvious, some non-obvious — that the practitioner can then evaluate and address. AI as a documentation reviewer is useful precisely because it approaches the document without the practitioner's assumptions about what is obvious.