Concurrence is running clinical AI agents at a roughly 1.2 trillion input token annualized run rate using Databricks Unity Gateway as a compliance-gated inference control point. The company processes approximately 100.8 billion input tokens and 11.2 million LLM calls every 30 days—a 5x growth from late 2025 to July 2026—across patient-facing workflows including care coordination, ambient documentation, and medication adherence outreach.
The stack consolidates three Databricks components: Lakebase for operational agent and conversation state, Unity Catalog for governing data and AI assets, and Unity Gateway for centralized AI access and security. Events stream through Zerobus Ingest into governed Delta tables, including 2.7 million world-model events per month and 90,000 at peak daily load. Concurrence computes patient state from immutable event history rather than overwriting records, preserving source and provenance of each data point. This architecture allows agents to call or text patients about care gaps and medication adherence, then record learnings back into patient state for future workflows.
Testing happens at scale before production deployment. Simulation and evaluation traffic is approximately seven times greater than production traffic on the new platform. Because patient state is computed from immutable events, teams can replay that state and test different agent paths without changing real patient records. Agent traces stream into Delta tables alongside clinical data, then scheduled jobs using Databricks-hosted Claude score interactions for conversation quality and safety, extract memory, and write results back to Delta.
Compliance gates model routing before cost or performance considerations. Concurrence is HIPAA-, GDPR- and SOC 2-compliant today, with HITRUST and ISO 27001/42001 in progress. Each healthcare organization gets its own schema and service principal with access controls and audit trails governed through Unity Catalog. For batch AI workloads, Concurrence runs jobs on Databricks-hosted Claude under a Business Associate Agreement; the endpoint resolver only permits models within the BAA-covered namespace, preventing protected health information from routing to uncovered models. The same covered path runs safety classification for self-harm, suicidal ideation, and medical emergencies.
Real-time patient and clinician inference remains on Concurrence's existing provider infrastructure today, though the company has already built and feature-flagged its Unity Gateway integration for real-time inference with a synthetic canary continuously testing it end-to-end. Production traffic can move to Unity Gateway as required compliance coverage becomes available.
Concurrence routes all coding-agent model and tool traffic through Unity Gateway's CLI. In July, 14 individual users generated 35.85 billion input tokens through Unity Gateway, of which 95.37% were cache reads. Since rollout on July 10, coding agents have generated approximately 360,000 requests and 61 billion cumulative input tokens. Every request is attributed to the engineer who made it, allowing individuals to monitor usage and the organization to track models in use, token consumption, cache rates, and spend by person and team. Claude Opus 4.8 and GPT-5.6 Sol account for most coding-agent model usage, with Opus 5 usage growing.
Production inference currently runs on 14 models from a governed catalog of 46. Most production volume runs on smaller, faster models, with frontier models reserved for complex reasoning. Concurrence is developing clinical reasoning benchmarks to determine which models perform best across different healthcare tasks and has begun testing Unity Gateway Smart Routing against healthcare-specific routing approaches it is developing. Because model eligibility in healthcare starts with compliance, those evaluations will assess how intelligent routing can optimize model choice within the boundaries established for each workload.
For teams scaling clinical AI in regulated environments, the takeaway is straightforward: make compliance the first gate in your routing logic, not the last, and build your data foundation on immutable events so you can test and replay without corrupting production state.