Three questions, three answers, every one with its source.
Numbers, dates and quotes in an answer come from what it looked up for that question, never from memory of an earlier chat. Anything that would leave the house, like a note to a supplier, is drafted first and sent only when you confirm.
You
What is still open with Halden Steel?
fluctor
Two invoices, €50,400 in total. invoice_2041 is 37 days overdue, invoice_2088 is due on 30 June. One delivery from order #88 has no invoice yet.
invoice_2041.pdf · page 1 · invoice_2088.pdf · page 1 · delivery_88-1.jpg
You
Did we agree a price increase with them this year?
fluctor
Yes. On 14 March, by email: plus 4 percent from April. It has not appeared on any of the nine invoices since.
RE: pricing from April · email of 14 March · invoice_2041 … invoice_2077
You
Draft a note to Dave Corbett about the missing units on PO 4471.
fluctor
Draft ready, with the delivery note and his own email attached. It goes out only when you confirm.
delivery_4471-1.jpg · RE: delivery week 22, PO 4471
LIVE · ASSISTANT
/chat
Product screenshot
QUESTION · ANSWER · SOURCES SHOWN ALONGSIDE
The answer on the left, the documents it came from on the right. The line it quoted is marked.
It would rather ask than guess.
A name that matches nothing gets a question back. A list that was cut short says so. A claim with no document behind it is not made.
No match
“I can’t find a customer called Halden Ltd. Did you mean Halden Steel Ltd?”
Only part of the answer
“Showing 50 of 214 matching orders. The total below covers these 50 only.”
Nothing to back it up
“There is no document in your data that says this, so I won’t state it.”
Answers in your language. Ask in German, it answers in German.
Asks about a date. “What did we know about order #88 on 1 May?” gets the state on 1 May, not today’s.
Follows the thread. “And the one before that?” works, because it keeps the conversation in mind.
The same chat builds a board, or an agent.
Describe what you want to keep an eye on and a dashboard is drafted. Describe a task you do every week and an agent is planned, with the cost named before it runs.
“A board with overdue invoices by customer and the trend over 90 days.”
→ three tiles and one chart, each checked against your live data before it is saved.
Dashboards →“Every Monday, remind customers with invoices more than 30 days overdue.”
→ four steps and an expected cost, then a first run that waits for your approval before anything is sent.
Agent →Ask it the question you could not answer last week.
Type it live in the call, on your own documents, and watch where the answer comes from.
No integration project to start · Your data stays in the EU