Machine-accessible expertise.
Expertise has always reached you through a person or a document. This page defines a third route, expertise an AI system can query directly and audit, and the site you are reading is the working demonstration rather than the brochure.
Precise enough to be tested.
Machine-accessible AI expertise is the practice of converting expert judgment, methodology, regulatory knowledge and decision frameworks into structured, governed interfaces that humans and AI systems can query, evaluate, apply and audit.
The four verbs are the specification. Query: an addressable interface answers a specific question, rather than a document hoping to be read whole. Evaluate: the answer arrives with source, status, and caveats, so its quality can be judged rather than assumed. Apply: the frameworks are published in executable detail, so the judgment can be used, scored, and acted on without its author present. Audit: the chain from question to answer to source can be reconstructed later, by someone who was not there.
The term is used on this page as an emerging practice. No claim is made that it is an established category, and the adjacent work in § 05 is named precisely so the claim can be compared rather than taken on trust.
What changes when the asker is a machine.
| Expertise through a person | Expertise through a governed interface | |
|---|---|---|
| Who can ask | Whoever books the expert. | Any person, and any AI system with an MCP client. No key, no account. |
| What comes back | An opinion, quality unknown until tested. | An answer carrying the framework applied, the source it came from, its status, and its caveats. |
| Cost of checking it | A second expert. | One request: the answer names the page it came from. |
| When knowledge runs out | Hedging, or confident improvisation. | A stated refusal naming what does not exist, asserted by tests rather than by tone. |
| The hard case | Is the engagement. | Escalates: the interface identifies the engagement shape, lists real slots, and can book the conversation. |
The conversion chain the definition describes, end to end:
Five properties, or it is something else.
A chatbot over documents is retrieval. A RAG pipeline is infrastructure. Neither is machine-accessible expertise until it clears all five of these, and the fourth is the one most implementations fail.
Structured
Judgment is converted into addressable concepts, frameworks with formal records, definitions with one canonical URL, before any retrieval happens. Retrieval over an unstructured pile returns the pile.
Governed
The interface has bounded authority. Read tools read the published corpus and nothing else; the tools that write, book or send require explicit confirmation and are annotated so clients prompt their user first.
Evidenced
Every answer carries its source, its status, and its date. An answer that cannot name where it came from is an assertion wearing the expert’s name.
Honest about absence
Where the knowledge ends, the interface says so, mechanically. This corpus states that there is no SAMA banking AI directive and no CBB AI regulation, and tests assert those refusals, because a fluent description of a nonexistent instrument is the most dangerous output an expertise interface can produce.
Escalating
The interface knows what it is not: hard cases route to the human expert, with the routing shown. Expertise without escalation is a FAQ; escalation without expertise is a contact form.
One request
Each property above is checkable against this site in a single request, and checking is the intended response. The inventory of what to check is § 04.
This site is the reference implementation.
Every number below is derived at build time from this repository or the live server card, never typed, so the page cannot quietly outlive its own evidence.
| Surface | What exists | Check it in one request |
|---|---|---|
| The corpus | 66 canonical pages, every one also served as Markdown to any client sending Accept: text/markdown, with a 406 rather than the wrong representation when both are refused. | Any URL on this site, with that header. |
| The callable layer | 11 MCP tools, public, no credential: 9 read the corpus, 2 can book a real consultation or send a real message, both requiring confirmation. | tools/list at the endpoint on /concylium. |
| The judgment | 4 framework specification pages, and the MESA assessment itself callable: start_assessment returns the published questions, score_assessment scores the answers and names the weakest controls. | /frameworks/mesa, or the two tools. |
| The regulatory knowledge | 44 maintained instruments across 6 jurisdictions, each carrying regulator, status, and date, including the rows that state a regulator has issued nothing. | /ai-governance/regulatory-tables, or lookup_regulation. |
| The escalation | The interface can identify the fitting engagement, read real calendar slots, and book the conversation with a real confirmation email. The commercial shape of that ladder is its own page. | identify_relevant_service, then list_consultation_slots. |
Who else is near this, named.
A definition offered without its neighbours is advocacy. Three serious efforts approach the same territory from different directions, and none of them does what this page defines, which is the observation that justifies the term.
Toloka’s Tendem makes human experts callable by agents over MCP: a marketplace routing questions to thousands of verified specialists. The interface is machine-accessible; the expertise stays inside people, priced per consultation, with the marketplace as the governance layer. Credo AI’s GAIA puts a governance assistant inside an enterprise governance platform: machine-delivered process, a platform’s workflow rather than a named expert’s published judgment. Microsoft’s Agent Governance Toolkit governs the agents themselves at runtime, policy enforcement and sandboxing; it is infrastructure for the asker, silent on the quality of what gets asked.
The intersection left open, and occupied here, is narrower than any of those: one named expert’s published judgment, converted into a governed interface that refuses honestly, carries provenance, and ends in the expert. Whether that intersection becomes a recognised category is a question evidence will settle. The definition above exists so the evidence can be compared.
MCP made judgment a supply chain.
Two things happened close together. Agents crossed from answering into acting, which moved governance inside the architecture, the argument of governed production AI. And MCP standardised how an agent reaches outside itself, which turned external judgment into a dependency an agent can acquire at runtime. From that point, the question where does your agent get its regulatory knowledge, and how would you know if it were wrong stopped being rhetorical: it is a supply-chain question, and it has better and worse answers. An interface that carries provenance, states its refusals, and escalates is the better answer. The governance of the connection itself is treated at MCP governance.
What this page does not claim.
Read this before you cite the page
- The term is offered as an emerging practice, defined so it can be tested. No claim of category invention is made, and the adjacent work in § 05 is named so the comparison can be checked.
- The interface answers from the published corpus only. It does not improvise beyond it, and it does not constitute professional advice; engagements are where advice with responsibility attached lives.
- Every count on this page is derived at build time from this repository and the live MCP server card, and is accurate as of the date in the title block, not perpetually.
- Adjacent-work descriptions are drawn from those organisations’ own public materials as read in August 2026, and are one sentence each by design; their sites are the authority on their products.
What readers ask first.
Is this just RAG with better marketing?
RAG is one mechanism this practice can use, and the least important part of it. The bar in § 03 is what separates them: structure before retrieval, bounded authority, provenance on every answer, mechanical honesty about what does not exist, and escalation to the human. A retrieval pipeline has none of those by default. The refusals are the sharpest test: ask a generic RAG system about a regulation that does not exist and it will usually find you one.
Why would an AI system consult this instead of its own knowledge?
Recency, provenance, and accurate negatives. A model’s trained knowledge has a cutoff and no sources; this corpus is dated, cited, and states where regulators have issued nothing. For a question like which instrument binds an AI system in the UAE today, the difference between a trained recollection and a dated row with a source is the difference between plausible and defensible.
The practical answer is simpler: assistants are increasingly instructed to check rather than recall, and this gives them somewhere specific to check.
What does using the interface cost?
Nothing. Public endpoint, no key, no account, no rate-limit tier. The commercial ladder, what is free and where the paid engagements begin, is the subject of the expertise-as-a-service page.
Can my firm have its own expertise made machine-accessible?
Yes, and the fastest way to scope it is to see how this one is built: the explain_this_setup tool returns the architecture of this site and server, component by component, because the method is part of the published work. The conversion of a practice’s judgment into a governed interface is an architecture engagement, and it starts with the Fit Call.
Check the claim before you cite it.
The fastest audit of this page is one request: ask any MCP client to call explain_this_setup at the endpoint on the Concylium page and compare what comes back with what you just read. If you want the same conversion done for your own practice or institution, that is an engagement, and it starts with the Fit Call.
Ask your AI assistant instead.
This page is a snapshot, accurate at the release it cites. The same corpus is callable, publicly and without a key, so an assistant can query it live and return an answer carrying the source it came from. For this page that is explain_this_setup and search_knowledge, which do what this page describes rather than describe it again: the first returns how this site's machine layer is actually built, component by component, and the second queries the corpus behind this page and returns matches with the URL each came from. The page states the practice; the tools are the practice.
claude mcp add --transport http concylium https://mcp.nabeelkhan.com/api/mcpClaude Desktop, ChatGPT, Cursor, VS Code and Gemini CLI take the endpoint on its own: https://mcp.nabeelkhan.com/api/mcp. No key, no account, nothing to sign. Setup for every client.
“Using Concylium, call explain_this_setup and tell me whether this site actually implements what its machine-accessible-ai-expertise page claims.”
A category page that survives being audited by the reader's own assistant is doing something a brochure cannot.