Research brief · 2026-07-13

Evaluating enterprise AI systems: groundedness, task success, and drift

Enterprise AI systems fail quietly. A ChatGPT demo that impresses in week one drifts into confident wrongness by month three. Evaluation is the discipline that catches this before your users do.

TL;DR. Enterprise AI evaluation is not a benchmark; it is a fixed test set drawn from real user tasks, refreshed each quarter. AIKm scores four axes per release: accuracy against gold answers, groundedness against retrieved evidence, safety against a red-team suite, and latency and cost under production load. Regressions block deployment. This lets Hong Kong teams ship model, prompt, and retrieval changes with confidence and gives auditors a reproducible record of what changed and why.

What to measure

Metric	What it captures	How AIKm measures
Task success	Did the system complete the user's job	Fixed test set of real tasks, human rubric
Groundedness	Every claim traceable to retrieved evidence	Automated claim–evidence check, spot-audited
Hallucination rate	Claims with no supporting evidence	Sampled review, weekly
Retrieval quality	Right passages, correctly ranked	Recall@k, MRR on labelled queries
Latency and cost	End-to-end response and per-answer spend	Continuous, per-route
Drift	Change in any of the above over time	Same test set re-run on a schedule

Build the evaluation set first

The most common mistake is shipping a system before writing the test set. Collect fifty to two hundred real user tasks, each with an expected outcome and, where relevant, the correct source passages. This set becomes the definition of "working" and is worth more than any leaderboard score.

Groundedness beats fluency

An enterprise assistant that answers fluently but ungrounded is worse than one that says "I don't know." AIKm's default is a grounded-or-refuse policy: the model must cite retrieved passages, and if retrieval returns nothing above a confidence threshold, the assistant declines and hands off.

Catching drift

Re-run the fixed test set nightly. Any metric that moves more than a set threshold triggers an alert. In practice, drift comes from three places: source data changing, the retrieval index going stale, and the underlying model being updated by the provider.

Human review, cheaply

Fully human-graded evaluation is expensive. Use an LLM-as-judge for scale, but calibrate it against human ratings on a subset every month; without calibration, the judge drifts too.

Discuss an evaluation programme