Choose the right approach — plain prompting, RAG, or an agent — for realistic workplace scenarios and frame the task for it.
Choosing the right tool for the job
Plain prompting, RAG, or agent — when to use which
Not every workplace task needs an agent or a RAG pipeline. Choosing the right approach depends on where the information lives, how many steps the task requires, and what happens if the model gets it wrong.
Plain prompting is the right starting point when the information needed is general knowledge already in the model's training data, the output is a draft (text, code, a template) that you will review and edit, or no external system needs to be queried or changed. Drafting a first-pass project proposal, explaining a concept to a non-expert, or brainstorming meeting agenda items are all good fits.
RAG is the right choice when the answer must come from your specific documents — a policy manual, a knowledge base, a set of reports — that the model was not trained on. It makes responses checkable because they cite the retrieved source. Use it when accuracy to your internal data matters and you can afford to verify the cited passages. Common workplace examples: answering questions from an HR policy manual, surfacing the right clause from a contract library, or querying past project retrospectives.
An agent is appropriate when the task requires multiple steps, calls to live external systems (databases, APIs, calendars, code execution), or autonomous decision-making across those steps. Examples: pulling this week's support tickets from a CRM, summarizing them, and drafting a report; or checking inventory across three systems and flagging discrepancies.
A critical warning: neither RAG nor agents remove the need for human verification. RAG can retrieve the wrong passage; agents can chain together several plausible but wrong steps. The more consequential the output — financial decisions, client-facing communications, legal interpretations — the more explicitly you must build in a human review step before anything is sent or acted upon.
Lesson notes
Plain prompting, RAG, or agent — when to use which
Not every workplace task needs an agent or a RAG pipeline. Choosing the right approach depends on where the information lives, how many steps the task requires, and what happens if the model gets it wrong.
Plain prompting is the right starting point when the information needed is general knowledge already in the model's training data, the output is a draft (text, code, a template) that you will review and edit, or no external system needs to be queried or changed. Drafting a first-pass project proposal, explaining a concept to a non-expert, or brainstorming meeting agenda items are all good fits.
RAG is the right choice when the answer must come from your specific documents — a policy manual, a knowledge base, a set of reports — that the model was not trained on. It makes responses checkable because they cite the retrieved source. Use it when accuracy to your internal data matters and you can afford to verify the cited passages. Common workplace examples: answering questions from an HR policy manual, surfacing the right clause from a contract library, or querying past project retrospectives.
An agent is appropriate when the task requires multiple steps, calls to live external systems (databases, APIs, calendars, code execution), or autonomous decision-making across those steps. Examples: pulling this week's support tickets from a CRM, summarizing them, and drafting a report; or checking inventory across three systems and flagging discrepancies.
A critical warning: neither RAG nor agents remove the need for human verification. RAG can retrieve the wrong passage; agents can chain together several plausible but wrong steps. The more consequential the output — financial decisions, client-facing communications, legal interpretations — the more explicitly you must build in a human review step before anything is sent or acted upon.