Market Analysis

The AI stack and the missing knowledge layer

MCP standardises access. Orchestration coordinates agents. But no layer yet owns trusted knowledge. Here is why the knowledge layer is the most important unsolved problem in the AI stack.

Build your first knowledge brain

Create a brain

The AI infrastructure stack is being built at speed. Three layers are already well-funded, well-understood, and converging on standards. One layer is missing. It is the most important one.

The three layers that are solving themselves

The Stack — 01

The emerging AI stack

Three layers are standardising. One remains unsolved.

🔌

MCP

Standardised connectivity to tools, data, and actions.

→

📦

MCI

Tool aggregation, packaging, discovery, and caching.

→

🤖

Orchestration

Agent coordination, workflow execution, memory, routing.

→

❓

Knowledge layer

Trusted context, source authority, provenance, conflict resolution.

MISSING

→

🎯

Application

The outcome the user or business is trying to achieve.

The first three layers have strong ecosystems. Anthropic, OpenAI, and the open-source community are solving connectivity, tooling, and orchestration. Nobody has solved what sits between orchestration and raw data sources.

Each of these layers has competing implementations already converging on standards. MCP is winning the connectivity layer. MCI extends it with tool management. LangGraph, CrewAI, OpenAI Agents, and others are competing on orchestration. All three will be commoditised within two years.

What none of them solve

Trusted knowledge: which source is the authoritative one for a given topic
Expert curation: who decided this information is correct, and when
Context ownership: who is responsible for this knowledge asset
Source authority: when two sources conflict, which one wins and why
Provenance: where did this fact come from and is it still current
Conflict resolution: the CRM says active, billing says suspended — which answer should the agent act on

MCP standardises access, not truth. Orchestration coordinates agents, not knowledge. The layers that are being built assume that high-quality, trusted context already exists somewhere. It does not.

The context hallucination problem

When an AI system returns a wrong answer, the instinct is to blame the model. But in most enterprise failures, the model is not the problem. The context is.

Consider a realistic scenario: the CRM says a customer is active, the billing system says the account is suspended, a Slack thread says the renewal was completed last week, and a cached agent response reflects the status from three days ago. The model receives all four signals. It has no way to determine which one is authoritative.

The Problem — 02

Context conflict, not model failure

When AI output is wrong, the default diagnosis is usually incorrect.

❌What people assume

— The model hallucinated

— We need a bigger, smarter model

— Better reasoning will fix it

— The AI is not ready for production

~80%

of "hallucinations" traced to
missing or conflicting context

✅What is actually happening

→ Multiple sources returned conflicting facts

→ No source has declared authority over the topic

→ The model picked the most plausible answer

→ The context layer does not exist yet

existing standards for
source authority or provenance

The model is not wrong. It is doing exactly what it was designed to do: synthesise the available information into the most coherent answer. When the available information is contradictory, the answer will be wrong regardless of model quality.

This distinction matters more as AI moves from answering questions to taking actions. A wrong answer costs you time to verify. A wrong action — cancelling the wrong account, sending the wrong response, routing the wrong ticket — costs significantly more.

What MCP solves and what it does not

Solves: standardised protocol for connecting AI to external tools and data sources
Solves: consistent interface for exposing read and write actions to any compatible client
Solves: tool discovery so agents can find what capabilities are available
Does not solve: which data source should be trusted when two return different answers
Does not solve: data quality, knowledge validation, or expert review
Does not solve: provenance, ownership, or conflict resolution

MCP is excellent infrastructure. It gives AI systems access to information. It has no opinion on whether that information is correct, current, or authoritative.

What MCI adds

MCI (Model Context Interface) sits above MCP and adds a management layer: tool aggregation, declarative tool definitions, configuration management, caching, and simplified MCP server creation. It makes the connectivity layer easier to work with. It solves packaging, not truth.

What orchestration frameworks handle

Agent coordination: multiple agents working on related tasks in sequence or in parallel
Prompt routing: directing requests to the right model or tool based on intent
Workflow execution: multi-step processes with branching logic and state management
Memory management: short and long-term context across sessions
Monitoring and observability: tracking what agents did and why

Orchestration frameworks are excellent at coordinating agents. They have no opinion on which source to trust when two agents return contradictory facts. They assume the knowledge problem is already solved.

The three unsolved problems

The Gap — 03

Three problems no current layer solves

Every AI system operating at scale will hit all three.

⚡

Context
conflict

—

Multiple sources, no authority

When CRM, billing, Slack, and a cached agent all disagree, the model picks an answer. Nobody decided which source was authoritative. The error is invisible until someone acts on it.

🔍

No
provenance

—

Where did this fact come from?

MCP delivers the data. It does not tell the agent who produced it, when it was verified, whether it is still current, or who is responsible for keeping it accurate.

🏆

No
authority

—

No ranking, no resolution

When sources conflict, something must decide. Today that is the model, using probabilistic inference. The correct answer is a knowledge layer with explicit authority rules and conflict resolution logic.

These are not model problems. Throwing a larger model at context conflict, missing provenance, or absent authority rules does not fix any of them. They require a structural solution at the data layer, not the reasoning layer.

Most organisations do not have a retrieval problem. They have a trust problem. The data exists. The question is which version to believe, who decided that, and whether the decision is still valid.

What the full stack looks like with a knowledge layer

The Solution — 04

The complete AI stack

With a knowledge layer in place, every other layer works as designed.

⚡ Request flow with a knowledge layer

User or application sends a request to the agent

Orchestration layer routes the request to the right agent and tools

Knowledge layer (KBrain) supplies verified, curated, authoritative context — before any live data is fetched

MCP / MCI fetches real-time data from source systems — complementing, not replacing, the knowledge layer

Agent answers with grounded, traceable output — sourced from verified knowledge, not probabilistic inference

The knowledge layer is not a replacement for MCP or orchestration. It is the missing piece that makes everything else reliable. Without it, the stack is functional but not trustworthy.

Where KBrain fits

KBrain is the knowledge layer. Not a chatbot, not a search engine, not an orchestrator. It captures expert knowledge, assigns ownership, tracks provenance, and exposes structured, verified context that agents can trust across sessions and across models.

Expert knowledge capture: curated brains built from documents, expertise, and structured data
Ownership and provenance: every brain has an owner and a source record
Authority ranking: brain descriptions tell agents which brain covers which topic, before retrieval
Cross-model portability: the same brain works on Claude, ChatGPT, and any MCP compatible client
Reusable domain expertise: build once, share with any agent or user who needs it
AI-ready representation: structured and indexed for agent retrieval, not human keyword search

The AI stack is almost complete. Connectivity is solved. Tool management is being solved. Orchestration is being solved. The knowledge layer is the last piece — and the most strategically valuable one. It is also the least crowded.

Build your first knowledge brain

Subscribe to KBrain, create a brain from your expertise or your data, and make it available to Claude, ChatGPT, or any MCP compatible assistant.

Create a brain

Frequently asked questions

What is the missing knowledge layer in the AI stack?

The knowledge layer is the part of the AI stack responsible for trusted context: who curated this information, who owns it, which source is authoritative when sources conflict, and how provenance is tracked. MCP, MCI, and orchestration frameworks all assume this layer exists. It largely does not yet.

What is context hallucination?

Context hallucination is when an AI system receives contradictory information from multiple sources and synthesises a plausible but incorrect answer. The model is not reasoning incorrectly. It is working with conflicting inputs and has no mechanism to determine which source is authoritative.

How is KBrain different from an orchestration framework?

Orchestration frameworks like LangGraph or CrewAI coordinate how agents work together. KBrain provides the knowledge those agents work from. Orchestration is about workflow. KBrain is about context authority. They are complementary layers, not competing tools.

Does KBrain replace MCP?

No. MCP solves how AI connects to tools and data. KBrain solves which knowledge should be trusted. KBrain uses MCP to expose brains to compatible AI clients. The two layers work together: MCP handles connectivity, KBrain handles context quality.

Why is the knowledge layer commercially important?

The connectivity and orchestration layers will be commoditised by large platform companies. The knowledge layer requires domain expertise, curation, and ownership — things that cannot be automated away. Whoever controls the trusted knowledge layer controls the most durable part of the AI value chain.