Choosing an LLM for the enterprise: an evaluation you run on your own work

Stanford's HELM evaluates models across many scenarios and metrics beyond accuracy, including calibration, robustness and efficiency ^[1]. LMArena ranks models by crowd-sourced head-to-head preferences ^[2]. Both are useful for a shortlist. Neither knows your documents, your edge cases or what a wrong answer costs you.

Collect 100 to 500 real inputs for the task, with the answer a domain expert would accept. Include the hard cases: ambiguous requests, missing data, inputs in other languages. Decide how each answer is graded, by exact match, a rubric, or a model grader checked against human grades. Anthropic and OpenAI both publish guidance on building such test sets ^[3][4].

• Task quality: score on your evaluation set, including how often the model says it does not know.
• Cost per completed task: tokens in and out, retries and any tool calls, at expected volume.
• Latency: time to first token and to completion at the 95th percentile, not the average.
• Safety and security: resistance to prompt injection and data leakage, mapped to the OWASP Top 10 for LLM applications ^[5].
• Data control: where data is processed and stored, retention, training use, and whether a private or regional deployment is available.
• Compliance: obligations under the EU AI Act for your use case ^[6] and how the vendor supports the documentation you will need, using a framework such as NIST's AI RMF ^[7].

Models change quickly, so the architecture matters more than the first choice. Keep prompts, tools and the evaluation set independent of any one vendor's SDK, route by task so a cheaper model handles simple steps, and re-run the evaluation whenever a model version changes.

Found this useful?

Share the key finding on X.