AI Can Make Mistakes
The message box of the AI Assistant carries a short notice: AI can make mistakes. This page explains what that means and what to do about it.
The assistant writes its answers with a language model. A model predicts a likely answer; it does not look the answer up. Most of the time that answer is right. Sometimes it is wrong, out of date or incomplete, and it can read just as confidently as a correct one.
Check what matters
Treat an answer as a draft, not as the record. Before you rely on one, check it against the platform itself, above all when it involves:
- dates, deadlines and durations
- amounts, rates and totals
- people, clients and the documents they belong to
- anything you are about to send to a client
Open the job, estimate or document the answer is about and confirm the values there. The assistant only reads the context listed under Agent and context, so an answer can miss something that is on a screen it was not given.
If an answer looks off, Ask again on the message asks for a new one. Rephrasing the question, or adding the document as context, often helps more.
Approve before it changes anything
Actions that change data, such as creating a job or moving a document to another stage, stop at an Approval required card. The card lists the arguments the assistant is about to use. Read them before you click Approve, and use Make changes to correct any of them first. Nothing on that card runs until you approve it.
Which actions ask for approval is set per agent. See Tools.
Report a wrong answer
Every finished answer has two buttons below it: Good response and Bad response. Click Bad response, pick what went wrong (for example Incorrect answer), and add a comment if it helps. When you send it, the question and the answer are shared with the team so the assistant can be improved.