Temperature: complete definition in AI for SMEs

Temperature is a generation parameter of a language model that controls the randomness of responses. It typically ranges from 0 to 2: at 0, the model is deterministic and almost always takes the same output path (safe, reproducible); at 2, it explores more possibilities and produces more varied, creative but less predictable responses. It is a slider between coherence and creativity.

What it changes for an SME

Temperature tuning is a free, immediate quality lever:

• Data extraction (invoices, tickets): temperature 0 → extracted fields are identical from one pass to another;
• Content writing (emails, posts, marketing texts): temperature 0.7 to 1 → natural variants without coherence risk;
• Brainstorming or creative variations: temperature 1.5+ → more surprising ideas, to be shortlisted afterward.

The trap

Using the default temperature without thinking. For an accounting agent, a 0.8 temperature generates invisible discrepancies between two processings of the same document. In fractional AI leadership, we set temperature per use case in the agent specification: 0 for restitution tasks, a medium value for writing tasks.