Taking action
Outcomes
A doorway into another system
Knowledge answers questions; tools take action. A tool is a doorway from an AI Employee into another system, reaching beyond conversation into the places where real work happens. Through one doorway it can check live calendar availability. Through another it can look up an order in a client's online store, or pull a product's current price, or create a ticket in a job tracker. Booking the appointment at the end of a conversation, rather than just chatting about one, is a tool at work.
The division of labor is precise:
- The AI decides when to use a doorway: it reads the conversation and recognizes the moment a lookup or an action would help.
- The doorway decides what is possible: a tool does one job in one system, and nothing beyond it.
A tool that checks calendar availability can check calendar availability. It cannot send email, edit the website, or touch the point of sale, because no doorway to those exists unless someone builds one. The employee's reach is exactly the set of doorways it has been given, and that set is a deliberate choice.
When information wants to be a tool
The Knowledge Base holds facts that hold still: services, policies, descriptions. Some information refuses to hold still: today's open slots, current stock, the status of an order placed yesterday. You could keep re-uploading documents to chase it, and the signal that you are doing so is unmistakable: if you find yourself updating knowledge constantly, that information wants to be a tool. A doorway to the live system answers with the current truth every time, with nothing to maintain in between.
That gives you the full sorting picture across the anatomy: rules go in instructions, stable facts go in knowledge, live data and actions go through tools.
How actions stay safe
Sooner or later a client asks the reasonable question: "so the AI can just do things?" The accurate answer has three parts, and each one is a design fact:
- Every action goes through a doorway you chose. No tool, no action. The employee cannot improvise a connection to a system nobody gave it.
- Each doorway is scoped to one job. The availability checker checks availability. Adding a second job means deliberately adding a second doorway.
- The AI never holds the keys. Credentials for connected systems stay outside the conversation and outside the model. The employee uses the doorway; it never sees what unlocks the door.
Judgment stays where it belongs, too: with you. Which doorways exist, and which employees get them, is a configuration decision a person makes, and reviewing how they are used is part of running a workforce. The Explanation view covers this side as well: it shows which tools an answer called and what came back, so every action is as auditable as every fact.
Name one system a client uses every day: their booking software, their point of sale, their online store. Now write the single question a receptionist could answer instantly with a doorway into it. That question is a tool waiting to be configured.
Building doorways is builder work: connecting a real system involves its address, its parameters, and its credentials. When you or your technical person are ready, the tools reference walks through it. The concept is what matters here: knowledge for what holds still, tools for what moves.
Knowledge Check
Three quick questions on what tools do, when to reach for one, and how actions stay scoped.