An LLM does not read a chart, it reads structured text. Turning telemetry into a useful debrief is about framing, not the model.
Comparing a performance to a reference and explaining the gap in plain language is a task an LLM excels at, provided you do not hand it raw data.
The LLM as an analysis layer
You do not send a full trace. You extract features (braking zones, per-sector gaps, variability) and present them as a compact table. The model comments on already digested numbers, not a stream.
The prompt sets the role, the register and the limits: no invention, flag uncertainty, rank by impact. The output is structured to be fed back into the product.
Framing to avoid hallucination
The risk is that the model invents a plausible but wrong cause. You constrain it to the data provided, forbid extrapolation, and keep human evaluation on a sample.
The LLM is not the system’s intelligence. It is the layer that makes an upstream analysis readable.