The criterion is the process, not the position
The useful question is not "who needs an agent" but "which process repeats and scales badly". The signs of such a process: typical inputs, a stable form of the result, the need to check the output and a queue of people waiting for the material. If the process is one-off — no agent is needed.
The executive: metrics without a queue
The Database consultant answers a question by voice or text: it builds an SQL query to the allowed views of PostgreSQL and returns a number with an explanation of the calculation and the source table. The executive does not wait for the analyst's report and does not learn SQL.
The expert: lawyer, analyst, engineer
For the expert the agent saves hours of search: the Legal RAG navigator finds clauses in the array of contracts with verbatim quotes and details, the Requirements Analyst collects scattered notes into a specification with acceptance criteria, the mail assistant prepares drafts of replies based on the company's knowledge base.
The product team: artifacts instead of chat
The team that makes decisions needs verifiable materials: the CustDev Agent builds an opportunity map with links to quotes, Super Oracle builds a contradiction map and the directions of the solution, Muse builds an interface audit with evidence and alternatives. This is one end-to-end process, not three separate tools.
Operations teams: email and meeting minutes without repetitive work
For operations services the agent handles recurring incoming work: the Mail assistant sorts the emails and prepares drafts based on the knowledge base without auto-send, the Local meeting scribe turns the recording of a meeting into minutes with timecodes — on your computer and without sending the data.
When an agent is not needed
- The task is one-off — manual work or chat is cheaper.
- The process is unstable: the rules change every week.
- There is no data and material for the input.
- Every decision requires expertise, and a person would have to redo the work anyway.
Questions and answers
Does a small team need an agent?
Usually yes: in a small team the queue to the only analyst or lawyer hurts more. A subagent covers one function without hiring.
How is the result of an agent different from a chatbot answer?
The result is a structured artifact: a report, minutes, a specification, a contradiction map — with sources and limits, not a dialogue history.
Where to start the choice?
With the catalog: three agents and nine subagents cover research, invention, critique, mail, meetings, documents, data and requirements.
When is an agent not needed?
When the task is one-off, exploratory and the context is easier to manage manually in a dialogue — a chatbot is enough.
What are the signs that an agent will fit?
The process repeats, the result must have a stable structure, several people check it, and the link between conclusions and sources matters.
Is a big team required to launch?
No: one responsible person with the right role is enough — the control points are set before integration.
Which process do implementations start with?
With a repeating routine that has a clear artifact: mail, meeting minutes, search in contracts, metrics, requirements.
What if there is little data so far?
Start with the demo on synthetic data and a survey; the quality of the agent's output depends on the quality of the materials.
How is an agent different from a chatbot in practice?
The agent organizes the work around the goal: inputs, stages, permissions and a verifiable artifact; the chatbot maintains the dialogue.
Who is responsible for the result?
The person: the agent prepares the artifact, the decision is made by the employee with the right competence.