Assistants
Chatbot, or agent?
The assistant Nexora built into Extrack can take 29 different actions on its owner's behalf, live at myextrack.com. That is the whole distinction: a chatbot produces sentences, and an agent produces changes in a system. Everything else about the comparison follows from it.
What can an agent do that a chatbot cannot?
Change something. A chatbot can tell a customer their invoice is overdue; an agent can find the invoice, send the reminder and record that it did. Extrack's assistant has 29 such actions, each wired into the product properly, and that wiring is the part that takes engineering rather than prompting.
Is an agent just a chatbot with better prompts?
No, and this is where most projects go wrong. The conversation is the easy half. The work is defining what the assistant may do, what each action needs before it can run, and what happens when one fails — none of which a prompt can supply.
How do you stop it inventing things?
By giving it tools instead of latitude. An assistant with 29 defined actions can do 29 things and say plainly when the answer is none of them. One with a prompt and a hope will fill the gap, because filling gaps is what the underlying model does well.
Which one does lead qualification need?
An agent, if the answers are to be worth anything. Asking the questions a salesperson would ask is chatbot work; writing the answers where the team already looks, so the conversation has actually started when a person picks it up, is not.
Questions we get asked
- Can we start with a chatbot and add actions later?
- You can, and the second half is where the cost is. The actions, their preconditions and their failure paths are the build; adding them later means doing that work later rather than avoiding it.
- Does an agent need access to our live systems?
- To be useful, yes — that is what separates it from a chatbot working off a copy. Which systems, and which operations within them, is the first thing to scope.
How Extrack's 29 actions were built