Skip to content

AI agents explained for managers: what to ask first

An AI agent plans steps and calls your systems; a chatbot only replies. Learn the differences and the questions to ask before approving one.

6 min readBy AI Guru AcademyFacts last checked

In short: an AI agent is a system where a foundation model breaks a request into steps, then calls your software or data sources to carry those steps out. A chatbot replies with text. An agent can change something in a business system, so the approval questions are about permissions, oversight and failure, not only answer quality.

What is an AI agent, and how is it different from a chatbot?

An agent acts; a chatbot talks. AWS describes Amazon Bedrock agents as ones that extend foundation models to understand a request and break the work into smaller steps, collect extra information from the user through conversation, make API calls to company systems to fulfil the request, and query data sources to improve accuracy.

The AWS documentation gives two examples of agent tasks: helping customers process insurance claims and helping customers make travel reservations. In both, the output is a completed action, not a paragraph.

A plain chatbot has one job: produce a reply. It has no route into your claims system or booking platform unless someone builds one, and once someone does, you are running an agent.

How do agents differ from RAG assistants and fixed workflows?

The difference is who decides the next step. Three common patterns sit on a line from fixed to flexible.

How do agents differ from RAG assistants and fixed workflows?
PatternWho decides the stepsTouches business systems?Typical risk
Fixed workflowA person, in advanceYes, in a set orderRigid; breaks on cases nobody planned
RAG assistantNone; it retrieves, then answersReads onlyA wrong or ungrounded answer
AgentThe model, at run timeYes, through toolsA wrong action, not just a wrong answer

A RAG assistant retrieves relevant material and uses it in a response. AWS documents this for Amazon Bedrock Knowledge Bases: when a query is made, the knowledge base searches your data to find relevant information to answer it. Retrieval informs the answer; it does not act.

An agent can use retrieval too. The AWS agent documentation says agents query knowledge bases to supplement the information they use when taking actions. So a RAG assistant is often one component inside an agent, not a rival to it.

How does an agent actually work?

An agent follows a loop of interpreting, deciding, calling tools and returning a result. The Amazon Bedrock documentation describes the agent as interpreting the user input, invoking action groups, querying knowledge bases and returning output to the user. An action group defines the actions the agent can perform for the user, such as the API operations it can call.

Orchestration is the coordination across those parts. AWS says agents orchestrate interactions between foundation models, data sources, software applications and user conversations. For larger jobs, the multi-agent collaboration page describes designating one agent as a supervisor and associating one or more collaborator agents with it.

For a manager, the useful point is that the list of tools is the list of things the agent can do. Scope the tools and you scope the risk.

What should a manager ask before approving an agent?

Ask five questions, and do not approve until each has a written answer.

  1. Which tools can it use? List every system, and whether each is read-only or can change data.
  2. Who approves actions? Decide which actions need a person to confirm first, such as payments, deletions and messages to customers.
  3. How is it monitored? Name the person who reviews what the agent did, and how often.
  4. What happens when it fails? Define the fallback, such as handing the case to a human, and how a wrong action is reversed.
  5. What limits its output? Check which safeguards sit around the model.

On the fifth question, Amazon Bedrock Guardrails can apply content filters and denied topics to model inputs and outputs. The documented filter categories are Hate, Insults, Sexual, Violence, Misconduct and Prompt Attack. Guardrails screen content. They do not replace the first four questions, which cover what the agent is permitted to do.

Worked example: a decision rule for scoping an agent

Illustration only; the company and figures are invented.

A mid-sized retailer, "Northfield Home", wants an agent to handle refund requests. The team lists three tools: read the order history, check the returns policy, and issue a refund.

Apply this rule: grant read tools freely, and gate write tools behind a person until the agent has a track record. Northfield Home lets the agent read orders and the policy, and prepare a refund recommendation. A staff member approves each refund. After a review period with few corrections, the team may allow small refunds without approval, with a written limit and a named owner.

The same agent without that scoping could issue refunds on a misread request. The model is identical; the permissions are the difference.

Are the AWS agent products still available to new customers?

Check the current documentation before you plan around a specific product. The AWS Bedrock agents page states that you can no longer sign up for Amazon Bedrock Agents, now called Amazon Bedrock Agents Classic. Existing users can continue to use it. For code-defined agents, the page points to Amazon Bedrock AgentCore and the Strands Agents SDK.

This affects planning, not the concepts. Autonomy, tool use, orchestration and oversight apply whichever product you choose, and product availability changes over time. Always confirm against the current AWS documentation rather than a guide that may be out of date, including this one.

Key takeaways

  • An agent plans steps and calls tools; a chatbot only replies, and a RAG assistant retrieves and answers.
  • The tools you give an agent define what it can do, so scope them first.
  • Read access is low risk; write access needs human approval until trust is earned.
  • Ask who approves actions, who monitors them and how failures are handled before sign-off.
  • Guardrails filter content but do not replace permission design.
  • AWS lists Amazon Bedrock Agents Classic as closed to new sign-ups, so check current documentation.

FAQ

What is the difference between an AI agent and a chatbot?

A chatbot generates replies. An agent breaks a request into steps and calls tools, such as APIs to company systems, to complete it. That ability to act is what raises the approval stakes.

Is a RAG assistant an agent?

Not by itself. A RAG assistant retrieves information and uses it to answer. An agent may use retrieval as one of several tools, alongside actions that change things.

Do agents need a human in the loop?

For actions with financial, legal or customer impact, a person approving the action is a sensible default. You can relax this once monitoring shows reliable results and limits are written down.

Can I still start with Amazon Bedrock Agents?

AWS documentation says new sign-ups for Amazon Bedrock Agents Classic are closed, and it points to AgentCore for code-defined agents. Read the current AWS page before deciding.

Sources

Last checked 2026-11-16.

AI Guru Academy is an independent training provider and is not affiliated with or endorsed by Amazon Web Services. AWS service names are trademarks of Amazon.com, Inc. or its affiliates.

Next step: take the free AIB-C01 diagnostic at /diagnostic, then see the course at /courses/aib-c01-exam-prep.

AI agents explained for managers: what to ask first | AI Guru Academy