Document AI is just plumbing. The plumbing is the product.

On two of the projects I have shipped — the Extrema technical-drawings pipeline and the Adhoc SCF portfolio-analysis system — the model was the easy part. Picking Document AI or Gemini and writing the extraction prompt took a week. Making the system reliable enough that an analyst would trust it without re-checking every field took six months.

The hard work happens in three layers nobody puts in the slide deck. The first is document detection. At Adhoc SCF, the same client portfolio arrives as eleven distinct KID templates from eleven banks, plus six legacy Excel formats from the previous advisor, plus scans of paper statements going back to 2018. Before you can extract anything, you must answer: what am I looking at? We trained a small classifier on the first page only — title-line position, header colour, footer copyright string — and it routes the document to the right extraction prompt. The classifier is two hundred lines of Python. It is the single highest-leverage component in the system.

The second layer is reconciliation. The same security has three different names across three documents. ISINs are sometimes missing, sometimes wrong by a digit, sometimes hidden inside a paragraph. We built a fuzzy-match step that combines string distance, optional ISIN look-up against the public Refinitiv API, and a final human-confirm row for anything below the threshold. Reconciliation is where every Document-AI demo collapses on real client data. The trick is to refuse confidently: if the system cannot identify a row, it surfaces a "needs review" card with the source snippet, the candidate matches, and a one-click choose. Analysts forgive uncertainty. They do not forgive silent errors.

The third layer is structured "I don't know". Every extracted field comes with three things: the value, the source-page coordinates, and a confidence score. The UI lets the analyst click the value, see the original PDF region highlighted, and either confirm or overwrite. That click is logged. The classifier and the extraction prompts retrain on that log every Sunday. Each week the system gets a little harder to disagree with.

The lesson is that Document AI feels glamorous because you can demo it on a single PDF and look like a magician. It earns its keep on the long tail — the scanned page that's 4° rotated, the Excel cell that says "TBD", the KID where the cost table spans two pages. The model is 5% of the work. The plumbing — classification, reconciliation, confidence, review UX, retrain loop — is the product.

This is what investors and product managers consistently underestimate. The cost of a Document-AI feature is not the model; the model has been a commodity since GPT-3.5. The cost is building the four boring layers around it that turn 88% extraction into 99.5% extraction. Skip those and your "AI feature" is a demo that survives Tuesday.