Litware Health's Contact Centre Triage

Order matters because one step is irreversible:

1. Speech to text (batch transcription) with language identification first, so English and Spanish audio are routed correctly rather than force-transcribed in the wrong locale.
2. PII detection and redaction on the transcript, before anything is persisted or sent downstream. This is the step that must not be reordered. Redact names, dates of birth, MRNs and phone numbers, and write the redaction mapping to a separate, access-controlled store with its own retention rule if re-identification is ever needed for clinical review.
3. Custom text classification to produce the complaint label - see the next task for why it cannot be the prebuilt sentiment call.
4. Sentiment and opinion mining as supporting signal, not as the trigger.
5. Summarisation over the tagged set to build the weekly issue summary.

For the knowledge base, use question answering over the policy PDFs with the documents indexed in Azure AI Search, keeping the answer grounded in retrieved passages and returning the source page. Note the 15-minute target is comfortably met by batch transcription with polling, but the redaction step must be inside that window, not run later as a cleanup job.

Why it failed. Prebuilt sentiment evaluates the tone of text - positive, negative, neutral, mixed - at sentence or document level. 'Complaint' is a business concept, not a tone. A politely worded message ('I am writing to let you know the third rescheduling has still not been confirmed') reads neutral-to-positive and is a serious complaint; a patient venting about a car park may read strongly negative and is not. The 41% figure shows the model is mostly picking up general negativity from clinical conversations, and the misses show it cannot represent the intent at all.

What to build. A custom text classification model (multi-class, since a call is tagged complaint or not, or into a small set of issue categories), trained on a few hundred labelled transcripts per class drawn from the agents' existing manual tags. Feed it the redacted transcript. Keep sentiment as an additional feature, not the label.

How to evaluate. Judge it on precision and recall for the complaint class against a held-out labelled set, not on accuracy - complaints are 3% of calls, so the same imbalance trap applies. Also sample the newly tagged calls with agents before trusting the weekly summary.

Three layers, because redaction alone is not a control:

1. Redact on ingest. Run PII detection on the raw transcript, replace detected entities with tokens, and store only the redacted transcript in the analytics store. Classify the raw audio as high-sensitivity with a short retention window and restricted access rather than keeping it indefinitely.
2. Keep the mapping elsewhere. If quality review needs to trace a call back to a patient, store the token-to-entity mapping in a separate secured store with its own access list, audit log and retention period. The analytics users never see it; the review workflow requests a lookup.
3. Minimise what leaves the boundary. Send only redacted text to downstream services - summarisation, classification, the knowledge-base index. Do not let raw transcripts into a search index that agents can query.

Finally, be honest about the residual risk: PII detectors miss entities, particularly names in free conversation and Spanish-language variants. So sample-audit the redacted output for false negatives, and treat a measured false-negative rate as the evidence compliance needs - rather than asserting redaction is complete.