The vocabulary of company AI, defined straight

An AI operating system is the governed layer that connects a company's knowledge, tools and AI agents into one system: it holds the agents' instructions, memory and access permissions in a single place, so AI behaves consistently, draws on the whole business's context, and stays under the business's control as it scales.

A company knowledge graph is a business's scattered knowledge, its documents, data, decisions and relationships, connected into one queryable structure that both people and AI agents can reason across, instead of searching one tool at a time.

Context engineering is the discipline of structuring what an AI system can see: selecting, organising and delivering the right business context to a model at the right moment, so its output reflects how the organisation actually works rather than a plausible guess.

Retrieval-augmented generation is the pattern where an AI model fetches relevant material from a knowledge source at question time and answers from what it retrieved, rather than from its training data alone. The model provides the reasoning; the retrieval provides the facts.

Data residency is where data physically lives and under whose jurisdiction: which country, which cloud region, and whose tenant. For AI systems it extends to prompts, outputs and telemetry, and it determines which regulators, laws and customers a deployment can satisfy.

A fine-tuned model is a base AI model further trained on an organisation's own examples, so it answers in the business's patterns, vocabulary and judgement. Done on an owned substrate, it is proprietary intelligence: an asset the company holds rather than a capability it rents.

An agent management platform is one console to register, monitor, govern and optimise every AI agent a business runs, whichever vendor or harness each agent lives in. It is a control plane, not another agent, and its defining property is neutrality: it manages the agents you bought, not just the ones you built.

An AI agent is an AI system that acts rather than only answers: it takes a goal, decides on steps, calls tools and other systems, and executes work such as sending messages, updating records or triggering workflows, with varying degrees of human oversight.

Agent sprawl is the uncontrolled accumulation of AI agents across a business, built, bought, embedded in SaaS, or wired up personally, each in its own console with its own access and its own bill, with no single place where they are all listed, measured or governed.

Shadow AI is AI used inside a business without the business's knowledge or governance: personal chatbot accounts fed with company data, unapproved AI features toggled on inside sanctioned tools, and agents running under individual logins that IT and leadership cannot see.

A control plane is the layer that manages a system without being part of the workload itself: it holds the registry, the policies, the monitoring and the controls, while the work runs elsewhere. In AI, a control plane oversees every agent in the estate without replacing any of them.

An agent register is the authoritative list of every AI agent a business runs: what each one is, who owns it, what data and actions it may touch, and how deeply it can be observed. It is the precondition for AI governance, because you cannot govern what you cannot list.

Human-gated approvals are a governance control where an AI agent's consequential actions, a payment, an outbound email, a record update, are held in a queue until a person approves them. The agent prepares the action; a human authorises it; an audit trail records who approved what.

A kill switch is the ability to pause or stop any AI agent instantly, from one place, regardless of which vendor or platform the agent runs on. It is the last-resort control that makes every other AI risk acceptable: whatever goes wrong, it can be stopped now.

Integration tiers describe how deeply an AI agent can be connected to a management platform, based on what the agent exposes: from inline gateway routing with full visibility, to native telemetry, to vendor admin APIs, down to plain registration for closed black boxes. The point of the tiers is that nothing is left ungoverned and nothing has to be rebuilt.

AI agent governance is the set of controls that keep AI agents accountable to the business: a register of what runs, boundaries on what each agent may touch, human approval of consequential actions, an audit trail of who did what, and the ability to stop any agent instantly.

Semantic model routing reads each AI request and sends it to the model that handles it best for the lowest cost: simple, mechanical work to fast cheap models, hard reasoning to frontier models. It replaces the default of sending everything to the most expensive model in the building.

Conversation-level cost tracking measures AI spend at the unit that matters: each conversation or workflow run, with its tokens, tool calls, model choices and latency recorded, across every vendor on one set of axes. It is the difference between knowing the invoice went up and knowing why.

AI enablement is getting a team to real, measured value from the AI it already pays for: hands-on training on tools like Claude, ChatGPT, Copilot and Gemini applied to actual work, reusable prompt and workflow systems, sensible access controls, and adoption tracked in hours given back rather than licences issued.

An AI opportunity audit maps how a business actually works, its tools, teams, workflows and data silos, and identifies where AI pays back first, ranked by impact and effort. Done properly it produces a map, a prioritised roadmap, and evidence, not a slide deck of possibilities.