The words around AI routing, defined so they can be quoted.

An LLM router is a component that sits between an application and several large language models and decides, for each request, which model should answer it - by reading the request and scoring the candidates on quality for that task, cost, latency and the caller's rules.

AI routing is the practice of deciding, per request, which AI capability should handle it - a model, an agent, a tool, a skill, a workflow or a person - under the caller's policy, with a record of why.

An AI gateway is a single API endpoint placed in front of several AI providers that handles keys, rate limits, retries, logging and - in a routing gateway - the decision about which model answers each request.

Model selection is choosing which AI model should handle a given request. Automatic model selection makes that choice per request from the request's own signals and the candidates' measured quality, cost and latency, instead of once in configuration.

Agent routing is deciding, per request, whether an AI agent - rather than a plain model, a tool or a person - should handle it, and which agent; it ranks agents next to the other targets under the same policy and records the choice.

MCP, the Model Context Protocol, is an open protocol through which AI assistants and agents discover and call tools, read resources and use prompts exposed by a server; an MCP router is an MCP server whose tools make the routing decision, and a router that imports other MCP servers as routing targets.

LLM cost is what an application pays to call language models, almost always priced per million input and output tokens with rates that differ by one to two orders of magnitude between models; LLM cost optimisation lowers it by sending each request to the cheapest model that meets the quality required.

AI governance is the set of controls that decide where AI requests may go, what they may contain, what they may cost and who may send them - and the records that prove the controls were applied. Enforced in a routing layer, it is a control on the path of every request rather than a document beside it.

Model fallback is what a router does when the chosen model fails, times out or is saturated: it re-routes the request to the next best candidate under the same rules instead of returning the error, and records both the failure and the recovery.

Semantic routing decides where a request goes from what the request means - its task, domain, complexity and sensitivity, read from the text - rather than from a field the caller set or a keyword it contains.

Prompt injection is an attack in which text that reaches a language model - in the user's message, a retrieved document, a web page or a tool result - contains instructions the model follows instead of, or in addition to, the ones its operator gave it.

PII redaction is the replacement of personal data in a request - names, e-mail addresses, phone and card numbers, identifiers - with placeholders before the request leaves the organisation, so a model can do its work without seeing who the data belongs to.

LLM observability is the ability to see, for every request an application sends to a language model, what was asked, which model answered, what it cost, how long it took, whether it succeeded and why that model was chosen - and to aggregate those facts across traffic.

A context window is the maximum number of tokens a language model can take in one request - the instructions, the conversation so far, any documents and the answer it produces, together - and it decides whether a request fits a model at all.

An OpenAI-compatible API is an HTTP interface that accepts the request and response shapes of OpenAI's chat completions, embeddings and models endpoints, so any client written for OpenAI can use a different service by changing its base URL and key.

Agent Skills are packaged sets of instructions, examples and resources - a folder with a SKILL.md at its root - that an AI agent loads when a task calls for them, so expertise is written once, versioned and shared instead of pasted into every prompt.

A2A (Agent-to-Agent) is an open protocol for agents built by different teams or vendors to discover each other through an agent card, exchange tasks and messages, and report progress, so one agent can delegate work to another without sharing code or internals.

A routing trace is the record a router keeps of one decision: the signals it read from the request, every candidate with its scores and price, the rules that removed or favoured candidates, the target chosen with its confidence, and later the outcome - enough to explain and audit the choice.

The routing learning loop is the mechanism by which a router improves from its own traffic: each reported outcome updates the estimated quality of the target that answered for that kind of request, so later decisions shift toward the targets that worked.

Structured output is a model response constrained to a declared shape - a JSON schema, an enumeration, a fixed format - so a program can parse it directly instead of extracting an answer from prose.

A reasoning model is a language model that generates intermediate thinking tokens before its answer - working through the problem step by step - which raises its accuracy on multi-step tasks and raises its cost and latency per request, because the thinking is generated and billed like output.