Professional Publishers Sit on the Most Valuable AI Asset – and Are Giving It Away

The thesis

LLMs are brilliant at the general and mediocre at the specific. They know who Goethe was, but not which DIN standard changed last month, which tax-court ruling came out yesterday, or what the actual pharmacokinetics look like in patients with renal insufficiency.

This gap has been filled for decades by professional publishers: Beck, Haufe, Wolters Kluwer, Springer, Thieme, Elsevier, LexisNexis, the standards bodies. They curate, validate, version – and sell the result as a PDF, a loose-leaf binder, or a database login.

Exactly this asset – verified, editorially maintained domain knowledge – is more valuable in the AI era than ever before. And most publishers are currently giving it away by clinging to their old distribution model.

Why LLMs need professional publishers

A foundation model does not solve three problems, no matter how large it gets:

1. Freshness. Training data is frozen at the cut-off date. A ruling from yesterday does not exist for the model.
2. Verifiability. An LLM cannot explain where it knows something from. In regulated B2B (law, medicine, tax, engineering) that is a show-stopper.
3. Depth. General training data contains textbook knowledge. It does not contain the commentary on §613a BGB in its 19th edition with the last 47 rulings published since print.

This is where publishers sit on a treasure: curated, versioned, citable, continuously updated domain knowledge. Exactly what a productive B2B AI workflow needs – and exactly what no foundation model, however large, can hold itself.

Context Engineering – the discipline that solves it

Context Engineering is the discipline of putting the right knowledge into the model's context at the right time. Not everything at once (token budget), not too little (hallucination), but exactly what is relevant for the current task.

In practice:

User query
   ↓ intent detection
relevance routing
   ↓ which sources, which time slice, which jurisdiction?
retrieval (vector / GraphRAG / structured query)
   ↓ top-k snippets with citations
LLM call with curated context
   ↓
answer with source attribution

This only works if the source is machine-accessible – structured, granular, with stable IDs and version stamps. That is the handover point where publishers fail today: a PDF in a login portal is unreachable for an agent.

MCP – the distribution channel that was missing

The Model Context Protocol (MCP) is the standard interface through which LLM agents talk to external data sources. Instead of every publisher building its own API with its own SDK, there is one shared protocol – and the agent (in Claude, ChatGPT, Cursor, Lovable, in any in-house B2B tool) connects with a click.

For a professional publisher this means concretely:

• One MCP server per product line (e.g. "Beck Tax Law", "Thieme Internal Medicine", "DWA Standards").
• Authentication via existing subscriptions – no new billing model needed, the license login moves into the OAuth flow.
• Granular tools instead of document dumps: search_judgments(topic, jurisdiction, date_range), get_commentary(paragraph, edition), cite(document_id).
• Structured outputs with stable citation IDs, version stamps, confidence levels.

The agent in the law firm pulls the relevant commentary while drafting – with a citation that holds up in court. The hospital workflow queries the current guideline before suggesting a treatment plan. The engineering agent checks DIN conformity in real time.

From PDF sales to AI feed

The old model: document behind a paywall, human logs in, reads.

The new model: knowledge atom behind an MCP endpoint, agent retrieves, cites, gets billed per query.

Three implications for the business model:

1. Pricing becomes usage-based. Instead of annual per-seat licenses: fractions of a cent per retrieved knowledge atom. That scales with the customer's AI workflow, not with their headcount.
2. The brand stays cited. Every answer the agent gives carries the source attribution. "Source: Münchener Kommentar, 9th ed., §613a margin no. 47" is brand-building inside every single LLM output – something a generic foundation model can never deliver.
3. The competition is no longer publisher X, it is "the model already knows". That is exactly why speed matters now. Anyone without an MCP endpoint in 12 months will simply not appear in customers' workflows anymore – not because their knowledge is worse, but because it is not callable.

What publishers need to build now

Three building blocks, in this order:

1. Content API with stable IDs. Before MCP works, the content has to be addressable. Every paragraph, every section, every table needs a permanent ID, a version number, a change date. Many publishers have this in the backend already – it has just never been exposed to the outside.

2. Retrieval layer. Vector embeddings across the whole catalog, optionally a knowledge graph for structured relationships (commentary → paragraph → ruling → follow-up ruling). This is the layer that makes "semantic search" possible – the foundation of every agent workflow.

3. MCP server. Thin wrapper over auth + retrieval + tool definitions. Technically buildable in weeks once the first two steps are in place. This is the part that becomes externally visible – and where attention will focus.

The business hook

In the next two years, professional publishers have the chance to evolve from "distributor of knowledge documents" into "knowledge infrastructure provider for AI workflows". Those who do not make the move will be disintermediated – not because their content is worse, but because it simply does not exist for the new generation of workflows.

This exact model – content audit, ID strategy, retrieval layer, MCP server, billing and license integration – is what we build at Till Freitag for publishers and knowledge-driven B2B companies. If you are sitting on a curated knowledge base and want to make it AI-ready, do it now, not in 18 months.

The most valuable asset in your house is currently locked behind a paywall that no agent can open. Open the door – and put a token meter next to it.