# AI Agent or Chatbot: What You Are Actually Buying for Training A chatbot is defined by its form, a conversation, while an AI agent is defined by control: it plans and carries out the steps of a task with tools, and one product can be both. This post is for owners and managers comparing training products sold under either label. It shows how to find out what actually runs behind the chat window. ## What makes something a chatbot? The Consumer Financial Protection Bureau defines chatbots as computer programs that mimic parts of human conversation. Its 2023 report separates rule-based chatbots, which follow a decision tree or match keywords, from more complex ones built on large language models. The definition is about the form of the exchange: a person types or speaks, and the program answers. It says nothing about what happens behind the answer. Some chatbots only reply. Others are connected to other systems and can look up a record, open a ticket or start a task. So the word chatbot tells you how you will talk to a product, not how much it can do. ## What makes something an AI agent? NIST, the federal standards agency, describes AI agents as systems that can perceive and take actions in an environment. In the common design today, a general-purpose model sits inside software that lets it use tools and act beyond producing text. The defining trait is control over the work. The system chooses the next step, uses a tool, checks the result and continues until the task is done or it stops. An agent can use a chat window as its front door, or run with no conversation at all, from a schedule or a trigger. In training, its steps might include drafting a lesson file, writing quiz questions or placing people in a group, if it has permission. ## Where do the two overlap? One label describes the interface and the other describes control, so a product can be a chatbot, an agent, both or neither. A chat window that only answers from documents is a chatbot without agency. A chat window that drafts a lesson and assigns it is a chatbot interface on top of an agent. A background process that redrafts files on a schedule is an agent with no chat at all. That is why the label on the box settles little. What matters is which steps the product can take, with which permissions, and where a person approves. ## What can each setup do in a training program? Most training work is a mix of answering, building and deciding. The table compares two common setups rather than two labels: a chat that only answers, and an agent that runs steps, with or without a chat window. Use it to read a vendor's feature list without taking the label on faith. Two setups compared by training job | Training job | A chat that only answers | An agent, behind a chat or a trigger | A person still | |---|---|---|---| | Answer a policy question during a shift | Replies from the documents it can reach | Replies, and can log the question for the source owner | Decides which document is the standard | | Build a lesson from a procedure | Writes lesson text into the conversation | Creates the lesson as a file in a training system | Checks the facts and approves it | | Check understanding | Asks questions in the conversation | Creates a quiz and records each score | Decides what a passing answer looks like | | Assign training by role | Points people to a lesson; assignment happens in another system | Assigns lessons to a group, if given permission | Chooses who needs what, and when | | Keep material current | Answers from the new version once someone swaps the document | Proposes a redraft when a source changes, if set up to watch it | Approves the new version before release | | Decide someone is ready | Quizzes a person and can pass the answers to a manager | Collects evidence such as scores and finished lessons | Watches the work and makes the call | ## What can go wrong with each setup? An answering chat can be wrong, and it can sound sure while being wrong. The CFPB report notes that chatbots sometimes give wrong answers and can leave people in loops with no way to reach a human. In training, a wrong answer teaches the wrong habit to one person at a time. An agent adds a second risk: a wrong action. A lesson built from an outdated file, or assigned to the wrong group, spreads the error to everyone who receives it. The fix is similar for both. Limit what they work from, and put a person between the draft and the people who learn from it. - Ask which sources the tool can read, and who decided that list. - Ask what the tool can change, and in which systems. - Ask where a person approves, and whether that step can be skipped. ## How do you test what runs behind the chat? Labels change faster than products, so test behavior instead of names. Bring one real procedure to the demo, and ask the vendor to run these steps in front of you. Insist on your file rather than theirs, and on a live run rather than a recording. - Change one line in the procedure. Does the tool only answer differently next time, or does it propose an updated lesson? - Ask what the tool did since your last session without being asked. An answering chat will have done nothing; an agent may have acted, which is fine only if you allowed it. - Ask for the record: who approved the lesson, which source it came from and which version staff saw. - Ask what the tool may change outside its own workspace, and how that permission is removed. - If the vendor cannot show the steps, treat the product as an answering chat and judge its price on that basis. ## What should a training tool never do without a person? Whatever the setup, some actions should always wait for a human decision. Write this list into your purchase notes, and ask the vendor to show where each one is blocked or held for approval. If a vendor cannot show it, assume the tool can do it. - Publish a new or changed lesson to staff. - Mark a person as ready to work alone. - Change a score, a completion record or a training history. - Send messages to customers or to people outside the company. - Change records in your HR, payroll or scheduling systems. ## Is Elvina a chatbot or an agent? Neither label fits Elvina neatly today. Elvina is designed as an AI employee for training, built around a defined job rather than a conversation. Drafting lessons, job aids and quizzes in Creation Studio from sources approved in Company Knowledge is available now. Assigning those lessons by group on the Learning Platform is available now. Approval by a named person before staff see any piece is available now as well. Answers to employees in Chat, drawn only from approved knowledge, are planned, not available today, so Elvina does not answer your staff in a chat today. Elvina working through the job on her own, the agent-like part, is planned, not available today; today our team runs these steps with you on the platform. Elvina does not act inside your other systems. ## What does a side-by-side test look like? A fictional car audio and remote-start installer wants help training its new installers. The owner sits through demos from two vendors and brings the same vehicle check-in checklist to both. - The first product answers questions about the checklist well. After a change to the dashboard photo step, it answers from the new file once someone uploads it, but no lesson changes and nobody is told. - The second product proposes a revised lesson and a new quiz question, and holds both for the shop manager's approval. - The owner asks both vendors what their tool may change outside its workspace, and writes the answers down. - The shop manager approves the revised lesson, and only then do the installers see it. - The owner chooses based on the record each tool keeps, not on which demo felt smarter. ## What we would measure Whatever you buy, measure what changes for the people being trained and for the person who approves the work. Start the log before the tool goes live. - Questions new hires still bring to a manager about tasks the tool covers. - Time from an approved procedure change to the updated lesson reaching staff. - Drafts the approver sends back, and the reason each time. - Actions the tool took that nobody asked for, which should be none unless you allowed them. - First-week checks a manager signs off on for the covered tasks. ## Questions people ask **Q:** What is the difference between an AI agent and a chatbot? A chatbot is defined by conversation: you ask, and it answers. An AI agent is defined by control: it plans and carries out the steps of a task with tools. The two overlap, because many agents use a chat window. In training, ask which steps a product can take and where a person approves. **Q:** Is a chatbot enough for staff training? An answering chat can help with questions, but a reply is not a training program. Someone still has to build lessons, assign them by role, check understanding and keep material current. If an answering chat is all you have, make sure it answers only from documents a named person has approved. **Q:** Can an AI agent change things in our other systems? Some can, if you give them permission, so ask exactly which systems and which actions. Elvina does not act inside your other systems. It works on its own platform, where a named person approves each lesson before staff see it, and that approval step is available now. ## Sources Public sources, checked on the date shown. - The CFPB defines chatbots as programs that mimic human conversation and separates rule-based chatbots from complex ones built on large language models. Source: https://www.consumerfinance.gov/data-research/research-reports/chatbots-in-consumer-finance/chatbots-in-consumer-finance/ (checked 2026-10-02). - The CFPB report says chatbots sometimes get answers wrong and can trap customers in loops without a path to a human. Source: https://www.consumerfinance.gov/data-research/research-reports/chatbots-in-consumer-finance/chatbots-in-consumer-finance/ (checked 2026-10-02). - NIST says AI agents perceive and act, usually a general-purpose model in software scaffolding that lets it use tools beyond text output. Source: https://www.nist.gov/news-events/news/2025/08/lessons-learned-consortium-tool-use-agent-systems (checked 2026-10-02). No launch date is committed for planned features.