Kim OlsenAI consultant in Copenhagen. I set up Claude agents for sales and customer teams.
The one-page rulebook that tells your AI sales agent which emails go to a person.
A discount ask hidden in a meeting request. Doubt about the renewal tucked into a thank-you. A how-to question from one of your key accounts. These are the emails an agent answers politely and gets wrong. This rulebook names them in six lines, and the test that comes with it lets you check what still slips through on your own mail.
What you get.
- The rulebook in English and Danish, as a block to paste into your agent’s instructions
- The 140 invented test emails, Danish and English, with the answer key for every one
- The script that runs a rulebook over the emails with your own API key
- The script that scores a run against the key: what slipped through, what was handed over for no reason
- A README with the commands and what the test cannot tell you
What the test showed.
40 of 40
fresh emails sent the right way by the English rulebook. The emails were written after the fix.
39 of 40
with the Danish rulebook. The one that slipped through is printed in the README as a known limit.
0
easy emails handed to a person for no reason, on the 40 fresh emails, in both runs.
A test, with a person in the loop.
This was a test of the sorting, not a setup to copy as it is. I do not recommend letting an agent send replies to your customers on its own. In real use the agent sorts the mail and suggests what to do, and a person approves every reply before it reaches a customer.
Read how the test was built and what it found: Which emails your AI sales agent must hand to a person →