Healthcare · Specialty Pharmacy
Preventing misdelivered medicines with NLP.
Fortune 10 healthcare enterprise
An NLP system that extracts delivery addresses from call transcripts at ~95% accuracy and validates them against member records — saving roughly $2M a year and protecting specialty medicine deliveries.
Healthcare · Specialty Pharmacy
Challenge
What was the challenge?
Specialty medicines are among the most expensive products a pharmacy ships, and many are temperature-controlled or time-critical. A package sent to an outdated or misheard address doesn't just inconvenience a patient — it can mean a spoiled shipment, a delayed treatment, and a costly re-dispense.
The delivery addresses in question lived inside free-text call transcripts: conversational, imperfect speech-to-text where an address might be spelled out, corrected mid-sentence, or fragmented across turns. Manually reading transcripts and cross-checking addresses against member records was slow, inconsistent, and impossible to scale to real call volumes.
Approach
How did we approach it?
We built a layered NLP pipeline rather than betting on a single technique. Named entity recognition finds address candidates in noisy conversational text; part-of-speech tagging helps assemble fragmented components into complete addresses; and targeted regular expressions catch the specific patterns that specialty-medicine conversations produce, which general-purpose models miss.
Extraction alone wasn't the goal — verification was. Every extracted address is automatically compared against the member's active address of record in the enterprise member database. Agreement confirms the shipment; disagreement raises a flag for intervention before anything is dispatched. The design principle: the system's job is to stop wrong deliveries, not just to read transcripts.
Delivered
What did we deliver?
- An NLP pipeline that extracts delivery addresses from call transcripts with approximately 95% accuracy.
- Automatic validation of every extracted address against the member's active address in the enterprise record system, flagging mismatches before a package ships.
- Specialty-medicine-aware extraction rules tuned for the vocabulary and conversation patterns of specialty pharmacy calls.
- A production deployment processing transcripts accurately and efficiently as part of daily operations.
Mechanism
How does it work under the hood?
- 01Transcripts are processed with named entity recognition to locate address mentions in conversational text.
- 02Part-of-speech tagging assembles street, unit, city, and postal components into structured candidate addresses, even when they arrive fragmented across sentences.
- 03Regex patterns tuned for specialty-medicine calls catch edge cases — corrections, spelled-out segments, non-standard formats — that statistical models alone miss.
- 04Each candidate address is matched against the member's active address of record; mismatches are flagged for human intervention rather than silently shipped.
Outcome
What was the outcome?
The system reached approximately 95% extraction accuracy in production and contributes roughly $2M in annual savings by preventing misdeliveries, spoiled shipments, and re-dispenses — while removing a manual review bottleneck from the delivery workflow.
Beyond the savings, it added a safety layer to specialty pharmacy operations: every delivery now gets an automatic address cross-check that previously depended on a human catching the discrepancy.
Learnings
Key learnings
Layered techniques beat one big model on messy real-world text — NER, POS tagging, and targeted regex each catch what the others miss.
Extraction is only half the job; cross-checking against a trusted source of record is what turns a model output into a safe operational decision.
Design for the failure path: flagging mismatches for humans, rather than auto-correcting, is why operations trusted the system in production.
Related stories
Healthcare · Pharmacy Benefits
Digitizing complex pharmacy contracts with generative AI
An end-to-end GenAI pipeline that reads dense pricing contracts — tables, legal language, and billing rules — and turns them into audit-ready structured data, unlocking roughly $64M in annual savings.
Healthcare · Pharmacy Network
Automating document intake for 700+ pharmacies
A hybrid LLM + machine-learning system that classifies and extracts data from roughly 1.6 million faxed pharmacy forms a month at 94% accuracy, feeding clean structured data straight into order creation.
Work with us