Chatbot or AI agent? Know what you are buying.

Start with the job, not the label

Imagine a plumbing business receiving the same questions every day: which areas it serves, whether it handles emergency work and how to request an appointment. A conversational assistant connected to approved information may answer those questions well. If that is where the repeated work ends, a broader agent team may add little value.

Now follow the enquiry further. Someone needs to identify the customer, establish the job type, check service coverage, record the request and route it to the right person. Those steps involve systems and decisions beyond the reply. Write them down before choosing a product. A supplier should be able to show which steps its setup actually completes, which require a person and which are outside scope.

A practical way to understand the difference

Anthropic distinguishes workflows with predefined routes from agents that can choose their next steps and tools. For a business buyer, that distinction is more useful than the appearance of a chat interface. A chatbot may include tools and agent capabilities; an agent may work without a visible conversation at all.

The following comparison is a buying aid, not a universal classification system. Product names vary. Ask the provider to demonstrate the behaviour in your actual tools rather than treating any label as a guarantee.

Follow one request all the way through

Consider an illustrative appointment change. The system first identifies the correct customer and existing booking. It checks the permitted calendar, finds suitable alternatives and asks the customer to choose. Before making a change, it confirms the chosen slot is still available. After saving, it verifies the new booking and sends the appropriate confirmation.

If the calendar is unavailable, the customer cannot be identified or a policy exception applies, the useful outcome is a clear handoff. It should preserve the details already collected and tell a named person what remains unresolved. Guessing a time or saying the booking changed when it did not would create more work than the automation removed.

Ask for a demonstration that includes failure

A polished demonstration usually shows an ordinary request with complete information. Your business also receives incomplete names, duplicate records, vague messages and unusual requests. Include those cases in the demonstration. Ask what happens when a tool rejects a change or a record has been edited by someone else.

Look for evidence rather than narration. If the agent says it updated the CRM, inspect the record. If it says it sent a message, check the delivery result. A useful activity history separates intended actions, attempted actions and confirmed results. Your team should not have to reconstruct that distinction from a confident paragraph.

• Use a realistic request from your business, with private details removed.
• Check the resulting record in the destination tool.
• Repeat with missing information and an unavailable connection.
• Ask who receives the exception and what information they receive.

Measure the work removed, not the messages generated

For a support workflow, track successful resolutions, staff corrections and repeat customer contacts. For appointment coordination, track completed booking changes, errors and time spent handling exceptions. The most interesting number is often the amount of human handling still required after the system has responded.

A faster reply can be valuable, but speed alone is not completion. Equally, an agent that performs many actions may be wasteful if those actions are unnecessary. Compare a representative set of requests before and after implementation. Record the volume and type of work so a quieter week does not look like an improvement caused by the system.

Choose the smallest complete solution

If approved answers solve the problem, start there. If a fixed process can reliably move the work, use that. Add agent judgement where requests vary enough to justify it. The goal is a complete customer or staff outcome with less manual handling, rather than the largest collection of bots.

Before approving a proposal, ask for five things in writing: the job the system owns, the tools it can access, the actions it may take, the situations it must escalate and the evidence you will use to judge success. Those details make it easier to compare providers and understand what your investment is intended to change.