AI-ready data is curated for model adaptation: examples are labeled for the task, reflect the production distribution, and are traceable. Operational records are not automatically training-ready; drifting categories, missing fields, and bot-closed tickets can undermine their usefulness. section 3.0 section 3.9
Assess readiness and prepare data by considering:
- Use the Levels 0–4 readiness spectrum to assess current data and identify steps toward fine-tuning readiness. section 3.9
- Combine acquisition strategies to fit the task: human annotation for high-quality seed data and evaluation sets, weak supervision for scale, production logs for freshness, and synthetic data for rare categories. section 3.9
- Gate synthetic examples for fidelity, format, and domain safety; check distribution alignment and cap the synthetic share at 30%. section 3.9
- Scrub personal data before examples reach annotators, annotation platforms, or training infrastructure; address both structured and indirect identifiers. section 3.9
Support-ticket records
A support team may have years of tickets, resolved threads, knowledge-base entries, and chat logs, yet still lack usable training data: categories may drift, severity fields may be empty, and some resolved tickets may have been closed by a bot. section 3.0
