What should an AI assistant do when your company data is not enough to support a compliance decision?
The recording
Nordvind Energy is a fictional company created for this demonstration. The recording shows an AI assistant using Nordvind's stored company information to assess whether the company should report under CBAM this quarter. The assistant finds the recorded 40-tonne import, but not the threshold needed to decide. It names the gap and returns the decision to a person. See it full size.
The record
Four questionsWhat it used
The assistant used Mistral Large through the interface's EU route. It searched Nordvind Energy's stored company information with search_context, which reported ok. The recorded model identifier was mistral/mistral-large-latest.
What it checked
The audit record reported: checker · 3 figures trace to the tools.
The check traced three figures to retrieved company information. It did not verify that the figures were correct or that CBAM reporting applied.
Where it stopped
The assistant stated: “I don't have the CBAM tonnage threshold or de-minimis figure for this quarter in your context. Without it, I cannot confirm whether 40 tonnes triggers reporting.”
It then returned the decision: “That decision is yours to make. Confirm it with your accountant or advisor.”
What this does not prove
The recording shows one useful behavior: the assistant stopped when a compliance threshold was missing. Before a company could rely on it in daily operations, the same behavior would need to hold across real data, repeated use, and unusual cases.
Reading
Author's noteWhat I take from it
A useful business assistant should not turn missing company information into a confident answer. Here, the system showed what it knew, exposed the missing threshold, and left the reporting decision with a person. The source check did not guarantee that the answer was right. It made the system's limits visible.
Recording limit
The session did not follow its own model guidance
For questions involving exact figures, the software recommends DeepSeek V4 Flash, Kimi K2.6, or GPT-OSS 120B. This recording used Mistral Large. The source check and refusal run independently of the selected model, and both worked here. The recording does not compare the models.