AB-100 key terms

A structured review of a business problem to judge whether AI can solve it, which data and systems it needs, and whether the benefit justifies the effort. Candidate use cases are ranked by value, feasibility and risk.

Agreed, measurable targets that show whether a solution delivers the intended business outcome, such as hours saved per case, deflection rate or user satisfaction. They are set with business owners before build and reviewed after launch.

The financial benefit a solution produces compared with what it costs to build and run, usually expressed as a percentage or a payback period. It combines quantified benefits with licensing, build and operating costs.

The full cost of owning a solution over its life: licences, model consumption, integration, support, training and eventual retirement. It is compared with expected benefits to decide whether funding is justified.

A decision on whether to configure an existing product, such as a packaged copilot, or build a custom solution, weighed on cost, time to value, differentiation, control and long-term maintenance.

A check of whether the data a solution needs exists, is accurate, complete, permitted for the intended use, and reachable from the systems the agent will call. Gaps found here change scope and timeline.

Policies, roles and controls that define who owns data, what it may be used for, how long it is kept and how quality is maintained. AI solutions rely on it to keep grounding content accurate and compliant.

Commitments such as fairness, reliability and safety, privacy and security, inclusiveness, transparency and accountability that guide how AI systems are designed, tested and operated.

A documented review of how a proposed solution could affect people and the organisation, covering bias, safety, privacy and compliance risk, plus the mitigations and owners agreed before build starts.

A cross-functional team that sets standards, reusable components, review gates and shared services for AI projects so teams do not rebuild foundations or repeat the same mistakes.

A low-code Microsoft service for building conversational agents that use topics, knowledge sources and actions, then publish to surfaces such as Microsoft Teams, a website or a contact centre.

An agent described mainly by instructions, knowledge sources and actions that runs on the Microsoft 365 Copilot orchestrator, so you define behaviour without hosting the runtime or model yourself.

An agent whose runtime, orchestration or model you host and control, connected to Copilot or Teams channels. Chosen when you need custom models, custom code or tighter control over data flows.

A designed conversation path in Copilot Studio with trigger phrases, questions, conditions and responses that handles one kind of request in a predictable, testable way.

A feature where the agent searches configured knowledge sources and has a model compose the reply at run time, instead of returning a fixed scripted response.

A connected store of content an agent may search when answering, such as a website, SharePoint site, file collection or business table, subject to indexing and user permissions.

A pattern where relevant content is retrieved from an index and inserted into the model prompt, so answers are based on your organisation's current facts rather than only on training data.

A numeric representation of text that lets a system find passages with similar meaning, stored in a searchable index and used to supply relevant content to a model.

Supplying a model with authoritative content, such as retrieved documents, database records or system instructions, so responses are tied to verified facts and can cite their sources.

The standing instructions that set an agent's role, tone, boundaries and rules for using knowledge and tools. They apply to every turn of the conversation.

A defined capability an agent can invoke, such as a connector, REST endpoint or Power Automate flow, letting the model read or change data outside the conversation.

The logic that decides which topic, knowledge source or tool handles a user message and in what order. Generative orchestration lets a model make that choice rather than relying only on trigger phrase matching.

A design where the agent hands a conversation or decision to a person when confidence is low, the request is sensitive, or policy requires approval, passing context to that person.

Model-level controls that detect and block harmful categories such as hate, violence, sexual content and self-harm, and can flag prompt injection attempts in user input or retrieved documents.

Choosing a model for a scenario by balancing answer quality, latency, context size, supported features and cost, while confirming regional availability, data residency and capacity.

A prebuilt or custom adapter that lets an agent or flow read and write data in another system, such as Dataverse, SharePoint or a line-of-business API, using managed authentication.

The plan for separate development, test and production environments so changes to agents, flows and data can be built and validated without affecting live users, with security policies applied per environment.

A Power Platform package that carries an agent, its flows, tables and settings between environments in a versioned, repeatable way, supporting controlled releases and rollback.

The step that makes a finished agent available to users by choosing surfaces such as Teams, a website or a contact centre and applying the authentication, branding and settings each requires.

A representative collection of questions with expected answers or behaviours, used to measure accuracy, groundedness and safety before release and after every change.

Usage and quality data a deployed solution emits, such as session volume, resolution and escalation rates, latency, errors and user feedback, used to prove value and find problems.

Estimating the request volume a solution must serve and provisioning enough model capacity, such as provisioned throughput or suitable quotas, to hit latency targets without overpaying.

A time-boxed release to a small group of real users with agreed success criteria, used to confirm value, quality and supportability before wider rollout.

The activities that help people understand, trust and use a new AI solution, including training, champions, communications and adjusting processes so the tool fits how work is done.

Rules that limit which connectors and data sources may be combined and what content may leave the tenant, applied to agents to stop sensitive information reaching unapproved services.

The ongoing cycle of reviewing telemetry and feedback, updating knowledge and prompts, retesting and republishing, so quality holds as content, models and user needs change.

Confirming that only the right people and systems can reach an agent, its knowledge and its actions, using single sign-on, role assignments and least-privilege credentials for connectors.

Keeping track of released versions and having a tested way to restore a previous one, so a faulty update can be reversed quickly without losing data or leaving users unsupported.