Disclosure or it didn't happen — and the enterprise-AI disclosure to interrogate is the phrase "grounded in your own data." It is the most common promise in the category, and the mechanism behind it has a name: retrieval-augmented generation. Understanding the IP clarifies what vendors are actually selling when they make that promise.

What did the patents actually claim? US12517941B1, "Retrieval-augmented generation for large language models" (issued 2026-01-06, assigned to POMA AI GmbH), and US12536449B1, "Self-supervised retriever optimization via attention-derived feedback in retrieval augmented generation systems" (issued 2026-01-27, assigned to Intuit Inc.), both protect the retrieval-and-grounding machinery — fetching the right documents and feeding them to the model — rather than the model itself.

“A document preparation method involves creating a hierarchical representation of an input document without summarizing or omitting any content.”— U.S. Patent No. 12,517,941 source

The core idea in the POMA AI grant is a "lossless hierarchical regurgitation process," and the word that carries the business value is lossless. Claim 1 describes reproducing an input document as a hierarchical representation "in which all text elements from the input document are preserved verbatim," with the language model explicitly instructed "to create the hierarchical representation without summarizing or omitting any of the text elements." That is the opposite of the usual summarize-then-embed pipeline, and it is a direct answer to the enterprise objection that AI "makes things up": if nothing is dropped or paraphrased on the way in, the system can quote the source exactly on the way out. A dependent claim adds that each chunk "maintains reference information indicating the chunk's source location" and that the model "is instructed to include the reference information when quoting" — citations, in other words, as a structural feature rather than a hope.

The claims also confront the practical limit that makes RAG necessary in the first place: the model's context window. The regurgitation process is iterative, generating "a partial hierarchical representation until a context window threshold of the generative language model is reached," then compressing prior partials "horizontally and/or vertically" and concatenating them with "the unprocessed remainder of the input document" for the next pass. The retrieval side builds "chunksets," each of which "indicates a traversal path through the structure of the input document" and "maintains pointers to the chunkset's constituent chunks," assembled by "collapsing adjacent chunks at the same depth level that fit within a token limit" and "recursively prepending higher-level parent chunks." Further claims handle the messy inputs real enterprises have: tables become a "tabletree representation" that "preserves all data," and images are described by "a vision-capable language model" or OCR and inserted "at a position corresponding to the non-textual element's location." This is the unglamorous data-engineering layer that determines whether an enterprise deployment actually works.

The Intuit grant attacks the other half of the problem — not how documents are prepared, but how the retriever learns which ones to fetch. Its title points to "attention-derived feedback," meaning the system improves the retriever using signals from the model's own attention rather than hand-labeled training data. Retriever quality is the difference between an assistant that surfaces the right policy clause and one that confidently cites the wrong one, and it is exactly the kind of self-improving loop that compounds into a durable moat.

Both grants also describe the live query path, which is where the prepared data finally earns revenue. The POMA AI claims describe receiving a query, optionally embedding it, "identifying chunksets relevant to the query by comparing the embedded query with stored chunkset embeddings," then "assembling a context document from chunks referenced by the identified chunksets" and handing it to the client model. There is even an incremental-update claim — detecting "an update to the input document," identifying the affected portions, and "selectively regenerating only the affected portions" — so a customer's knowledge base stays current without reprocessing everything. That maintenance story is the unglamorous detail that separates a demo from a deployment a business will pay a subscription for.

The business implication is where value accrues, again. If the defensible IP is the retrieval pipeline and the data integration, then a vendor can build a durable product on top of a model it licenses rather than owns. That is the enterprise-AI business model in one sentence: own the grounding and the data plumbing, rent the intelligence. The patents are the evidence that this is where companies are staking claims — POMA AI on lossless document structuring, Intuit on retriever self-optimization.

Distinguish GAAP from guidance from pitch. None of this appears as a segmented "RAG revenue" line in an audited filing; it is product strategy visible in patents and sales narratives. The grants describe methods, not market share, and neither one discloses how many customers or dollars the technique attracts. So the appropriate posture is to treat RAG IP as a signal of competitive positioning, not as a disclosed financial result. The grants are real; the revenue attribution is not yet in any 10-K.

The exacting takeaway: when an enterprise vendor sells "AI grounded in your data," the thing it can actually defend is the retrieval-and-grounding layer, not the model. POMA AI's lossless-structuring grant and Intuit's retriever-optimization grant make that concrete. Judge the pitch by the quality of the retrieval IP and the data integration — verbatim preservation, source citations, table and image handling, self-improving retrieval — because that is the part the documents say is theirs.