How to choose between pre‑built services and custom models?

Use pre‑built services for common tasks and fast time‑to‑market. Choose Azure ML for niche domains, higher metrics or data/control requirements.

Which metrics should I track?

Precision, recall and F1 for classification/NER; BLEU/COMET for translation; perceived quality and end‑to‑end latency/costs.

Enable PII detection/masking where available, use compliant regions, retention policies and access controls.

Text Analytics, NER, sentiment, classification, translation and Q&A with Azure Cognitive services.

Detect sentiment, language, key phrases, PI/PII and insights from text & documents.

Extract entities such as people, organizations, places, tax IDs, IBANs (where supported).

Ticket routing, moderation, and content categorization (single or multi‑label).

Neural translation in real‑time with custom glossaries and format preservation.

Build knowledge bases from FAQs/documents and enable Q&A in chat and portals.

Use regional endpoints with keys/role assignments. Text Analytics example:

POST /text/analytics/v3.2/analyze
Ocp-Apim-Subscription-Key: <key>
Content-Type: application/json

Event‑driven with Functions/Service Bus, batch via Data Factory; for custom models move to Azure ML.

Service	When to use	Output
Text Analytics	General insights from text	Sentiment, key phrases, language, PII
NER	Extract structured fields	Entities + types + offsets/confidence
Classification	Routing, moderation, categories	Labels (single/multi) with scores
Translation	Localization & multilingual	Translated text + glossaries
Q&A	Self‑service information	Answers from KB/FAQs

Normalize encoding, remove boilerplate, handle language and domain specifics.

Precision/recall/F1, error analysis and A/B tests on real samples.

PII masking, logging, rate limiting, and drift/cost monitoring.

Multilingual support?

Yes, many services support dozens of languages; check the ones you need and validate quality with custom glossaries.

Can I train a custom classifier?

Yes, with custom classification where available or with Azure ML for full pipelines.

How to handle sensitive text?

Enable PII detection, apply masking and access controls; use compliant regions and limited retention.