The documents your AI reads were never treated as data

A sales rep asked the company's internal AI assistant for the current price of an enterprise tier and quoted the answer to a customer. The assistant had cited a pricing sheet from the shared drive; the current version sat in the same folder, both had been indexed, and nothing in the index said which one was real. Finance caught it during deal review two weeks later. When I traced it back with the team, the question that stalled the investigation was not technical. Nobody could say which version was authoritative, because nobody had ever been asked.

The instinct is to treat this as a retrieval problem: better chunking, a newer model, a reranker. It is not. Both pricing sheets were genuine documents, and the AI assistant retrieved one of them correctly. What was missing was a decision about which one counted, and that decision belongs to data governance rather than to the AI stack. The pricing sheet simply never passed through the data platform, so the governance never reached it.

At most startups, the most valuable input to the planned AI features is unstructured: contracts, support tickets, call transcripts, pricing sheets, internal documentation. The data platform was built for tables. The governance gap does not sit inside the data platform. It sits next to it.

The documents nobody classified as data

Data platforms were scoped around systems of record with schemas: the CRM, billing, product events. Documents lived in tools owned by business functions, such as legal's drive, support's helpdesk and sales' call recordings, and nobody classified them as data. They have no data owner, no data contract, and no retention rule aligned to how they are now used.

When an AI prototype needs them, the access path is whoever has a shared-drive link. In almost every AI readiness conversation I have, the data catalogue lists every data warehouse table and none of the unstructured data sources the AI roadmap depends on. The functions that own those documents do not know they have become data producers.

Authority and currency

For a table, data quality means completeness, freshness, and valid values. For a document, it means two things no table-oriented check can express: authority, meaning whether this is the signed version or a draft, and currency, meaning whether this is the policy in force or the one it replaced. A folder can hold both, and without context they look identical.

The default build skips the question. The AI team exports a folder snapshot, indexes it, and ships. Three versions of the same policy sit side by side in the index, drafts next to signed copies, and the AI assistant answers from whichever matches the question best. It does so with full confidence, because confidence is not a property of the source document.

The snapshot also starts ageing the day it is taken. The refresh is manual, so a corrected pricing sheet changes the shared drive but not the answer. Pointing a RAG tool directly at the drive solves the refresh and leaves the authority question untouched: no tool knows which of three pricing PDFs is the real one.

The three-party gap

Each party is behaving reasonably. The data team treats documents as outside its remit, since they never entered the data platform. The business function that owns the pricing sheet does not know it is now a production data source for an AI system. The AI team owns the data pipeline but not the content. Between the three, nobody decides which document collections are authoritative, who keeps them current, or whether the access rules on the drive still apply once the content is searchable through the AI assistant.

The leadership mandate

The CTO owns the decision that unstructured data feeding AI is data, and that it falls inside data governance even when the storage stays in Google Drive or the helpdesk. Without that decision, the gap persists by default, because every party has a sound reason the responsibility sits elsewhere. With it, the data team's remit extends to the unstructured data sources the AI roadmap depends on, and the business functions that own them become data producers with obligations they can see.

The path forward

The starting point is an inventory, made before anything new is indexed: the unstructured data sources the AI roadmap depends on, each with an owner, an authoritative location, a version rule, a refresh trigger, and an access rule. Most teams find it is a short list. The top three collections are worth treating as data products, with a named owner accountable for what the AI assistant is allowed to treat as current.

The test for whether it works is a single question: if this document changes tomorrow, how long until the AI answer changes? In the pricing sheet case, the honest answer was never, and finance found out before anyone in engineering did.