Start with what the bot must not do
A community member testing Fin wrote that the responses were not yet good enough to turn it on for all users. Before you switch yours on, write down the cases it must refuse, the cases it must hand to a person and the things it must never say, such as a price, a refund or a policy you have not written. That list becomes the test.
Run it more than once
Policy-following agents often fail to repeat a task: an early model scored about 61% on single retail runs and about 25% when the same task had to pass eight runs in a row. Run each case at least three times and call a case that passes only some runs flaky and fix it before go-live.
Check the label and the way out
Article 50(1) of the EU AI Act makes providers of such systems ensure people are told they are talking to an AI system unless that is obvious, from 2 August 2026. Zendesk's terms put notification on the customer where the law requires it. Check on a test chat that the label and a way to reach a person are visible in every state.
Which paid job fits
The test pack runs a written set against your bot in a sandbox and names the fix for each fail. The handover job checks that each trigger reaches a named person with the conversation. The hostile-input test is for bots that can look things up or send messages.
A monthly check re-reads a sample once the bot is live. None promises a resolution rate, and you switch the bot on yourself.
Sources and limits
- Intercom Community: how do you train Fin AI to respond (13 Dec 2024) Checked 2026-10-12.
- A community member wrote that they had been testing Fin and were not yet satisfied enough with the responses to turn it on for all users.
- tau-bench: benchmarking tool-agent-user interaction (arXiv 2406.12045) Checked 2026-10-12.
- Policy-following agents often fail to repeat a task: an early GPT-4o scored about 61% on single runs in retail and about 25% when the same task had to pass eight runs in a row.
- EU AI Act Article 50 Checked 2026-10-12.
- Article 50(1) requires providers to ensure people are told they are talking to an AI system unless that is obvious, and applies from 2 August 2026.
- Zendesk AI services addendum Checked 2026-10-12.
- The customer must notify end users or obtain consent where law requires, and the customer deploys and owns the output of the AI services.