The generative wave made every AI problem look like a prompt, but the fleet of specialized AI services keeps winning on the tasks it was built for: cheaper per call, faster, deterministic contracts, and often better accuracy. Today, the tour of the ones that matter and the framework for choosing them over an LLM.

The Roster

Language: PII detection and redaction (the workhorse of every compliance conversation this month, scrubbing transcripts and tickets before they reach models or logs), named entity recognition, sentiment and opinion mining, summarization, custom text classification and NER trained on your labels, and conversational language understanding for intent routing. Speech: transcription with real time and batch modes, speaker diarization, custom speech tuned to your domain vocabulary (the difference between medical or ATC transcription working and not), neural text to speech with custom voice for the brands that license one, and speech translation. Vision: image analysis with captions, tags, and OCR, custom classification and object detection through the vision training tools, liveness and face capabilities behind their gated access program. Translator rounds it out with text translation the Speech and Language services compose with. Each has the same operational shape by design: a resource, an endpoint, Entra auth with managed identity, private endpoints, diagnostics, one governance pattern across the fleet.

Prebuilt or Prompt: The Decision

The framework has four questions. Volume and unit cost: transcribing a call center or scanning millions of comments prices catastrophically through a frontier model and routinely through the purpose built service. Latency and determinism: the specialized services return structured results with stable schemas in tens of milliseconds, no prompt drift, no parsing a model’s creative JSON. Accuracy on the narrow task: PII detection, diarization, and OCR are benchmarked specialties where the dedicated models hold their ground or better. Flexibility: the moment the task needs reasoning across context, judgment, or generation, the LLM earns its cost. The architectures that win combine them, and the call center pattern is the canonical example: Speech transcribes and diarizes, Language redacts PII and tags entities, and only then does the LLM summarize and extract the insights, each layer doing what it is cheapest and best at, the composition the Document Intelligence post already modeled for paper.

Governing the Fleet

Multi service resources versus per service is a governance choice: single service resources give per workload cost attribution and least privilege, the multi service account buys convenience for exploratory teams, and the FinOps tagging decides how much convenience you can afford. Commitment tiers exist for the high volume services and belong in the same portfolio review as the reservations post. Containers deserve the closing note: many of these services ship as containers running on your AKS or at the edge, disconnected scenarios included, which matters for data residency conversations and for the IoT arc starting next week, the same model scoring at the site where the data is born. Route everything through the content and data handling review your responsible AI process defines, log the calls like any dependency, and the specialized fleet becomes what it should be: boring, priced, and quietly doing most of the AI work in the building.

Cheers
Osama