Most firms arrive at an AI conversation with a tool in mind and a problem loosely described. The vocabulary does not help: assistant, chatbot and agent are often used to mean the same thing, which makes it hard to tell what is being proposed and what it is allowed to do.
A working distinction
- Chatbot: a scripted or narrowly scoped conversation. It handles a defined set of intents and escalates everything else. Low risk, low flexibility.
- Assistant: answers questions from an agreed knowledge base, with citations. It can be wrong, but it is checkable. No external action.
- Agent: given a goal, chooses steps and uses tools to reach it. The most useful and the most consequential, because it can act.
None of these is inherently better. The right choice depends on whether the problem needs an answer, a recommendation, or an action.
The question to ask before choosing
Ask what the system is allowed to do when it is not sure. A chatbot says 'let me connect you to a person'. A knowledge assistant says 'I could not find this in the knowledge base'. An agent with no guardrails does something plausible and tells you afterwards.
The interesting design question is not what the AI can do. It is what it may do unsupervised.
Where most firms should start
Start with the assistant. It has the shortest path to value and the lowest risk of an invisible mistake: the firm's own documents, a citation on every answer, and a clear behaviour when nothing relevant is found. It teaches the team how to work with a model on their own material before anything is connected to clients.
Then add one workflow with a human approval gate. Research prepares, a person commits. That single pattern covers proposal drafting, onboarding, reporting and most repetitive internal processes without handing over judgement.
Where agents earn their place
Agents are worth the added risk when the path is genuinely variable: the steps depend on what is found along the way, and a fixed workflow would need constant manual branching. Even then, the useful default is a limited tool set, logged actions and an approval step before anything external.
If someone proposes twelve agents before a single workflow is in production, the vocabulary is doing the selling rather than the problem.