Documentation is not a cost center. It's the retrieval layer your AI depends on.

Twenty years, three companies, zero missed deadlines

I've delivered structured documentation programs at three industry leading companies over 20 years and never missed a deadline. That is not a luck streak. It's what happens when content is governed instead of improvised. Every writer working from the same taxonomy, the same approved terms, and the same style rules produces predictable output. Predictable output ships on schedule. Most documentation programs miss deadlines because every writer is making their own decisions about structure, terminology, and tone on every single page. Governance removes those decisions before the writer ever opens the topic.

---

Structure beats prose: why DITA

I build documentation on structured authoring, specifically DITA. Instead of long unstructured pages, content is broken into topics: concepts, tasks, and reference material, each one a discrete, reusable, independently valid unit. That sounds like a technical preference, but it's really an economic one. A Word document or a wiki page is a monolith. Change one sentence and you're editing prose by hand, hoping you didn't break a reference somewhere else in the set. A topic is a component. It has an identity, it can be reused, and it can be governed on its own.

Governance is three things: taxonomy, term database, style guide

Taxonomy defines what categories exist and how they relate to each other, so content has a place to live instead of getting duplicated because nobody could find the original. A term database defines the approved vocabulary: preferred terms and their deprecated equivalents, so "sign in" doesn't quietly become "log in" in one product area while the rest of the set says something else. A style guide defines how sentences get written. On the programs I've run, that includes rules like using third person possessive voice instead of second person, keeping sentences under 25 words, and always rendering UI button labels in bold.

None of that is cosmetic. A consistent term database and style guide mean two documents written a year apart, by two different writers, read like they came from the same entity. That consistency does three concrete things: it gets translation memory closer fuzzy matches instead of near misses that still need a human translator, it gives large language models cleaner signal about what a given piece of text is actually referring to, and it strengthens the brand because the content stops reading like a patchwork.

---

Topic reuse: write once, reuse everywhere

Reuse is where governance turns into money. Write a topic once, govern it once, and reference it everywhere it applies instead of writing a slightly different version of the same procedure five times across five products. That cuts the time writers spend writing, it unifies the content because there is only ever one governed version of the truth, and it directly reduces localization cost, since every reused topic is translated once instead of five times.

I ran a CMS with 2.4 million topics across 35 languages. With a real reuse strategy in place, we cut content costs by between roughly $500,000 and over $1 million every year, because writers were referencing governed topics instead of rewriting the same content over and over. Translation memory added savings on top of that. A well governed, reused topic set produces the fuzzy matches translation memory needs to avoid paying for a full retranslation of content that barely changed.

---

Metadata and structured authoring are what make RAG actually work

Retrieval quality is not a model problem. It's a content architecture problem, and structured authoring is what solves it. Unstructured prose gets chunked for RAG by cutting it into arbitrary token windows that have no relationship to the meaning of the text, which is how you end up with a chunk that starts mid sentence and a model that has to guess the rest. A DITA topic is already the right shape for retrieval. It's a complete, self contained unit built around one concept or one task, so when it gets retrieved, the model gets a whole, coherent answer instead of a fragment.

Metadata is what turns that topic set into something a retrieval system can actually filter on: product, version, audience, and platform become real constraints instead of guesswork, so the system can rule out content that does not apply before it ever reaches the model. And because the term database keeps vocabulary consistent across the whole set, there is one governed way to refer to a given concept instead of three near synonyms competing for the same query and splitting the retrieval signal between them.

Case study: shipping the NetBackup Alta chatbot fast

We were able to move quickly on the NetBackup Alta chatbot because the documentation work was already done. The content was governed under a shared taxonomy, a term database, and a style guide, and it had years of reuse behind it, so it was already unified instead of scattered across inconsistent variants. Metadata was already in place to scope retrieval by product and version. None of that had to be retrofitted under a chatbot deadline. We plugged RAG into a content set that was already clean, and the retrieval results were clean because of it.

---

That's the throughline of everything I do on documentation readiness. The teams that can move fast on AI are the ones who did the unglamorous governance work first: taxonomy, term database, style guide, and a real reuse strategy. The teams still writing unstructured prose end up paying for that later, either in localization cost, in missed deadlines, or in a chatbot that can't find a clean answer because the content underneath it never had one.