Case Study: Prosper AI
Prosper AI is the AI platform that runs the entire patient journey, building voice agents that work both sides of the phone for healthcare operations. Over 60 outpatient groups and platforms trust Prosper AI to handle conversations about benefits verification, billing, and collections with patients and insurers. These are calls where the details have to be exactly right, and a dependable Speech-to-Text (STT) layer is central to every one.
Key Results
- 98.5% of user identity verifications completed successfully; 90% on the first attempt, with no spelling required
- 99%+ accuracy capturing benefits, registration, and other structured data on every call
- 99%+ QA-reviewed accuracy across post-call transcription processing
About
Prosper AI is the AI platform that runs the entire patient journey, building voice agents that work both sides of the phone for healthcare operations. Over 60 outpatient groups and platforms trust Prosper AI to handle conversations about benefits verification, billing, and collections with patients and insurers. These are calls where the details have to be exactly right, and a dependable Speech-to-Text (STT) layer is central to every one.
Challenge
Much of Prosper AI’s hardest work happens on the phone with insurers. When an electronic eligibility check comes back incomplete, an agent calls the payer directly and works through the phone tree and hold queues to reach a live representative who can verify benefits or move a prior authorization forward. Capturing the rep's answers accurately is critical to confirming coverage and updating patient records.
“The highest-stakes moments involve capturing names and alphanumeric identifiers like claim numbers, policy codes, and numerical amounts. A single misrecognition breaks every downstream workflow that depends on them (payer rejections, failed eligibility checks, incorrect billing, and manual rework).”
Arnau Saumell
Founding AI Engineer, Prosper AI
As call volume grew, Prosper AI put its STT solution through a rigorous evaluation on its own production audio rather than clean benchmark clips, judging each option on how it actually performed on live payer and patient calls. An earlier setup looked fine in pilots but did not hold up under production load, with latency that drifted through the day. On a live call, the agent might stall or the caller might get frustrated, landing the conversation back with a staffer, the exact manual work Prosper AI built its product to take on.
Solution
Prosper AI standardized its STT on Deepgram, running Nova-3 for real-time streaming during the call and high-fidelity batch transcription afterward. The batch pass is where Prosper AI extracts the structured data payer workflows depend on, so accuracy on alphanumerics is critical, and the team uses keyterm prompting to keep those identifiers and payer-specific vocabulary sharp. Deepgram runs as the transcription layer inside the agent stack, integrated over WebSocket for streaming and the SDK for batch, all in Python. A human QA team spot-checks transcripts, and the team watches sentiment and escalation rates as proxies for how well the speech layer is holding up on real calls.
What solidified Prosper AI’s partnership with Deepgram was how the speech layer behaved once it was carrying real traffic. Calls stayed responsive through daily peaks, and on the rare miss the errors were consistent enough that the team could design around them instead of firefighting each one.
“Deepgram stood out for delivering production-grade accuracy with consistently low latency and, crucially, a predictable failure mode. While other providers introduced instability and erratic behavior, Deepgram's issues were limited to occasional transcription inaccuracies.”
Arnau Saumell
Founding AI Engineer, Prosper AI
Results
For the calls that carry the most risk, verifying identity and gathering data from payers, Prosper AI now has transcription it can build on. Getting an ID or a code right is the difference between a call resolved automatically and a call that requires someone else to get involved. And for Prosper AI’s customers, the goal is to automate as many calls as possible so staff can focus on higher-value care.
The outcomes Prosper AI measures show how this impacts patients. Prosper AI completes 98.5% of identity verifications successfully, and 90% clear on the first attempt, without requiring the caller to spell a name or read back a member ID.
The structured data holds up too: benefits, registration, and other fields come through at 99%+ QA-reviewed accuracy across post-call transcription processing. For the patient, that means a benefits check or a prior authorization settled on the first call, without a transfer.
Prosper AI aims to become the industry standard for healthcare operations, front and back office, and expects Deepgram to remain the STT core as it scales.
“In healthcare, reliability is table stakes. Deepgram’s track record made it the clear choice, with consistently low latency, no hallucinated content, and accuracy that holds up in production. When every word in a patient interaction carries real consequences, you need a speech partner that gets it right every time, and Deepgram does.”
Arnau Saumell
Founding AI Engineer, Prosper AI
