Glossary
AI Governance
The policies, roles and controls that decide which AI systems an organisation runs, who approves them and how they are monitored once live. Frameworks such as the NIST AI RMF, ISO/IEC 42001 and the EU AI Act give it a recognised shape.
General definition
AI Governance is the management layer around AI: the decisions, documents and checks that make sure an organisation knows what AI it is using, why, with what data and with what risks. It sits above individual technical measures such as AI guardrails and above legal AI compliance; governance is what decides which guardrails to require and which laws apply to which system. In most organisations it is owned jointly by a risk or compliance function, security, legal and the engineering teams that ship the systems.
A working programme contains the same building blocks whatever the industry:
- Inventory: a register of every AI system, model and vendor in use, including the LLM features inside other products and the agents employees have set up themselves
- Risk classification: tiers based on impact (a chatbot that answers FAQs is not the same as a model that scores loan applications), which set how much review each system gets
- Approval and change control: who signs off a new system, a new model version or a new data source before it reaches users
- Policies and training: an AI acceptable use policy for staff, data rules, disclosure rules and the AI literacy the EU AI Act now expects
- Monitoring and incident response: logging, evaluation in production, a route for users and staff to report problems, and a plan for withdrawing a system that misbehaves
Three external frameworks give the programme a recognised structure. The NIST AI Risk Management Framework (voluntary, United States) organises the work into Govern, Map, Measure and Manage. ISO/IEC 42001 turns it into a certifiable management system. The EU AI Act makes parts of it law for anyone placing AI on the EU market or using it there, with obligations that scale with the risk tier. Most organisations map their programme to one of these rather than inventing a structure of their own.
Governance fails most often not for lack of policy but for lack of visibility: models change under a vendor’s API, staff adopt tools nobody registered, and the logs that would show what an agent actually did were never kept. The practical test of a programme is whether it can answer, for any AI output a customer complains about, which model produced it, from what inputs, under whose approval.
In the Ethora ecosystem
Ethora does not run your governance programme, but a self-hosted chat and agent platform removes several of its blind spots. Every AI agent in the AI SDK is a named object with its own system prompt, model, knowledge base and attached tools, so the inventory is a list you can read from the admin panel rather than a spreadsheet reconstructed afterwards. The model is chosen per agent, and because the AI service talks to any OpenAI-compatible endpoint you can point it at a cloud model or at a self-hosted model whose version you control.
Monitoring and evidence come from the messaging layer: every exchange between a person and an agent is a message in a room, covered by the server-side archive, the compliance audit trail with immutable log export, and the Trust and Safety system that can block, quarantine or report content. Data boundaries are set by the deployment model. On a dedicated server or a self-hosted install, prompts, transcripts and knowledge stay inside your own cloud account or data centre, which is the answer most governance committees want to the question of where the data goes.