These four terms often enter a meeting as if everyone shares a definition. Usually they do not. Use the distinctions below to pin each term to observable behavior before anyone approves a connection or relies on an answer.
This article provides general operational education. It is not legal advice or a compliance determination. No attorney-client relationship exists, and none of its protections apply.
Retrieval-augmented generation
Retrieval-augmented generation, or RAG, combines information retrieval with an AI-generated response. The system finds material and supplies it to the model as context for answering.
In a fictional legal-team example, an assistant retrieves sections of an approved contracting playbook before preparing a draft comparison. The playbook provides context that the model might not otherwise have.
Retrieval does not establish that the material is current, complete, or relevant. The system might retrieve an old clause, miss an exception, or generate an answer that goes beyond the passage.
Microsoft's RAG documentation describes the retrieval-and-generation pattern and its limits. It explicitly recognizes that grounded answers can remain inaccurate.
Ask which sources are included, how versions are managed, and whether the reviewer can open the exact supporting passage. Test a question whose answer is absent from the approved source set.
For the playbook example, the useful outcome is an explicit uncertainty or escalation route. A fluent answer to every question is not an acceptance condition.
Connector
A connector links software to another service or data source. The term alone does not tell you what the connected tool can read or change.
For example, Microsoft Copilot connectors can make external data available through indexed or live retrieval. Power Platform connectors expose actions and triggers. These are different product mechanisms.
In a fictional matter-intake workflow, one connection might retrieve an approved form. Another might create a record in the destination system. Review their permissions and consequences separately.
Ask which account supplies the connection, which records it can reach, and whether it can create, update, delete, or send anything. Check whether the data is copied or fetched when needed.
Test the permitted action and a prohibited one using approved test material. Then confirm how to disconnect the tool and what happens to queued work.
“Connected” is a setup description. The access boundary still needs to be checked in the actual product, account, and workflow.
AI agent and automated workflow
An automated workflow can follow predefined steps. An AI agent may choose steps or tools as it works toward a goal. An automated workflow can also contain AI without giving the model control of the whole process.
Anthropic uses this distinction in its engineering guidance. Definitions differ across vendors, so treat the label as a starting point for questions.
In a fictional routing process, fixed rules send requests to known destinations. An agentic version might choose which source to search, whether to ask a follow-up question, and what tool to use next.
That flexibility changes what needs testing. Inspect allowed tools, stopping conditions, and the point where a person approves a consequential action.
A narrow task may work with fixed rules or a single AI step. There is no need to add model-directed decisions unless they solve a demonstrated problem.
Start with the workflow-selection article. The operating question is what the system can decide and do, regardless of its product label.
Grounding, citations, and verification
Grounding supplies context for an answer. A citation points the reader toward a source. Verification checks whether that source supports the specific statement being used.
In a fictional policy summary, an answer links to the right handbook but gives an outdated approval threshold. The link works. The proposition still fails verification.
Open the source and find the supporting passage. Check its date, scope, exceptions, and relation to the sentence. For a quotation, compare the exact wording. For a calculation, reproduce it.
Microsoft's RAG documentation distinguishes retrieving context from evaluating the resulting answer. Check the evidence even when the interface supplies a citation.
Decide how uncertainty reaches the reviewer. If the source is unavailable, contradictory, or outside the approved set, the workflow should use its correction or escalation route.
See human-review placement for a checkpoint that gives the reviewer the source and the proposed output before reliance.
If the terms are clear but your own team's answers are not, the free AI readiness assessment surfaces the gaps in 10 questions.
Sources used
- Microsoft Foundry RAG documentation, updated August 21, 2026; checked September 19, 2026.
- Copilot connectors overview and Power Platform connectors overview, checked September 18, 2026.
- Building effective agents, December 19, 2024; checked September 18, 2026. Used for terminology, not current tooling recommendations.
