The Integration Surface: Where Healthcare AI Goes to Die
The integration surface is where healthcare AI goes to die. Not the model, not the demo, not even the contract — the part where an agent has to read and write inside a live EHR without breaking HIPAA, without duplicating records, without forcing a nurse to copy-paste between screens. That surface is different at every hospital. Most vendors pretend it isn't.
Wedge, founded in 2025 by Devraj Gopal and Steven Segawa out of Y Combinator's Summer 2025 batch, reached 60-plus locations in its first week of piloting, a pace that suggests the market has been waiting for this model. Wedge's bet: the only way through is to put senior engineers inside the building. Not solutions engineers. Not consultants. Forward-deployed engineers who sit in standups, own the integration surface, and ship against a measurable workflow number.
The problem isn't new. Olive AI raised nearly $1 billion automating revenue-cycle tasks before shutting down in 2022; UiPath's RPA bots still choke on HL7v2 feeds that vary site to site. Both approaches assumed a standard integration layer that doesn't exist. Epic dedicates a technical team to each hospital implementation for a reason — every instance is a snowflake of custom modules, local vocabularies, and firewall rules. A point-solution vendor selling the same agent to dozens of health systems hits dozens of different walls.
Gopal built health AI at Stanford and Hopkins Medicine; Segawa led ML at NVIDIA and shipped production systems at Rivian and SoFi. Their process starts with an automation identification audit: shadowing staff, mapping back-office tasks from payment reconciliation to medical coding and AI governance. Each agent gets a defined ROI target: revenue recaptured, costs cut, hours saved. Then an onsite engineer tailors the agent to that specific Epic or Cerner instance, handling bidirectional sync, conflict resolution at the resource level, bulk FHIR pipelines.
Early traction bears this out. Wedge's website reports $250,000 per year saved per agent, and its data shows nearly two million patients served across dozens of sites.
Epic and Cerner's Native Tools Leave a Gap Wedge Fills
Epic and Oracle Health (Cerner) together serve three-quarters of U.S. hospital admissions. Epic says it's building roughly 200 AI features; Oracle's agent runs in over 30 specialties and cuts documentation time by almost a third. Yet a 2025 Deloitte survey found only 31 percent of hospitals achieved seamless EHR interoperability. The gap is not model quality. It is workflow fit.
Epic's integration runs through FHIR R4 APIs and HL7 v2 standards; Cerner leans on its CDS Hooks framework plus FHIR endpoints. Both pathways work for first-party features. They work less well for anything that sits outside the vendor's roadmap. Epic customers must use App Orchard, an app store that imposes tight integration fees and review cycles. Cerner customers, via Oracle Cloud, can tap a broader menu of cloud-hosted AI services, but Oracle's next-generation EHR is currently available only for ambulatory providers, with acute care functionality slated for 2026. Epic's AI tools already span inpatient and outpatient. That split leaves health systems with a choice: wait for the vendor's timeline or build workarounds that inherit technical debt.
Workflow misalignment drives failure. Nearly two-thirds of AI implementation failures in hospitals resulted from workflow misalignment, not technical failure, a 2024 JAMA Network Open study found. When AI drops into an EHR without clinical redesign, clinicians hit alert fatigue: low-confidence, irrelevant alerts that get ignored until the system provides no net benefit. Early surveys show clinicians worry less about data security than about AI distracting them or adding false alerts. Trust forms when tools clearly save time and errors stay rare. Native tools, which vendors build for the average workflow, rarely meet that bar for every department.
Hidden costs compound the problem.
| Category | Description | Monthly Range | Annualized Range | Notes |
|---|---|---|---|---|
| Integration Engineering | Initial integration engineering cost | — | $140,000–$280,000 | One-time |
| Internship Compensation | Software engineering internship (San Francisco) | $4,000–$8,000 | $48,000–$96,000 | Pro-rated monthly |
Per-seat licensing adds thousands more; retraining and governance inflate budgets by roughly half. Health systems that have done this before move 42 percent faster because they've already paid the learning tax. Leading systems now budget millions a year for eight to a dozen concurrent deployments.
Wedge's forward-deployed engineers change the calculus. They sit inside the hospital, audit the actual workflow — not the vendor's idealized version — and build agents that plug into the existing FHIR or HL7 layer without waiting for App Orchard approval or Oracle's ambulatory-only rollout. The agents automate back-office operations: payment reconciliation, records retrieval, claims management, medical coding, reception, AI governance. Because the engineers co-design with the staff who use the tools, the agents fit the department's rhythm. Non-disruptive integration models achieve nearly nine in ten clinician adoption within 90 days, versus about one in three for systems requiring workarounds or parallel documentation, multi-site analyses show.
Compliance is the other lever. Epic's closed ecosystem and Oracle's 2025 data breach — affecting more than 80 hospitals, prompting an FBI investigation and class-action suits, have sharpened buyer scrutiny. Texas filed an antitrust suit against Epic in December 2025 alleging monopolistic practices and data access restrictions; Particle Health filed a federal suit earlier.
The regulatory stack is hardening beneath both vendors. Since August 2025, the FDA had moved from one-time device clearance to a lifecycle oversight model built around Predetermined Change Control Plans, requiring manufacturers to pre-specify how AI models would be retrained, recalibrated, and monitored post-deployment. ONC has raised the certification floor for every certified EHR: HTI-1 and HTI-4 now mandate native FHIR R4 API support, SMART on FHIR 2.0 for third-party integrations, and real-time patient data access by July 2026. Starting in 2027, vendors must report interoperability metrics. CMS-0057-F sets a hard deadline: FHIR-based prior authorization APIs go live January 1, 2027, and the rule bars AI from making coverage determinations without documented individual patient circumstances.
HHS has codified its own governance framework through a five-pillar strategy released in January 2026 covering governance, infrastructure, workforce, research, and care delivery. The agency's "OneHHS" model centralizes decision-making across divisions, and a compliance plan requires all divisions to apply risk management practices to high-impact AI by April 3, 2026, with metrics drawn from a use-case inventory. An RFI that HHS issued December 23, 2025 (comments due February 23, 2026) signals the agency intends to accelerate adoption while tightening the guardrails.
State enforcement is moving faster than federal rulemaking. Colorado's AI Act takes effect June 30, 2026, requiring disclosure, impact assessments, and anti-bias controls. Utah has enforced disclosure and a $2,500-per-violation penalty since May 2025. Texas and California added their own mandates this year. Five more states have draft bills targeting utilization-management AI and disclosure, with enforcement expected through 2027.
The liability architecture is hardening in parallel. Clinicians carry personal liability for over-reliance on AI outputs that breach the standard of care. Health systems face institutional exposure for poor oversight, inadequate training, or weak vendor management. Vendors remain liable for software defects, safety flaws, and capability misrepresentations — and cannot contract away responsibility for personal injury. Non-compliance with FDA or HIPAA rules only strengthens the case against them. In response, leading health systems are rewriting contracts with AI-specific indemnity clauses, data-breach SLAs, audit rights, and change-management triggers that force renegotiation when models are updated.
Wedge positions its operating system as a trust layer that governs every agent's data access, audit trail, and liability boundary. That layer travels with the agent, not the EHR vendor's roadmap. Hospitals are not rejecting native AI. They are layering Wedge on top where the vendor's roadmap stops short of their operational reality. The vendor builds the platform. Wedge builds the fit.
Growing the Engineers Who Can Live Inside a Hospital
Wedge's forward-deployed model only works if the engineers it sends into hospitals can speak both languages: the clinical workflow vernacular of revenue-cycle managers and the technical dialect of agentic AI systems. With a founding team of two, the company cannot hire its way out of this talent gap. It has to grow the talent itself.
A rolling software-engineering internship, posted continuously on Y Combinator's job board and LinkedIn, serves as the mechanism. The role, based in San Francisco, tasks interns with the design, development, and maintenance of software systems that support AI agent deployment in healthcare settings.
Interns at Wedge ship code that runs inside the hospitals where Wedge is piloting. Every agent the company deploys — each targeting a back-office function, must integrate with legacy EHRs, satisfy HIPAA and emerging HHS transparency rules, and survive the scrutiny of compliance officers who can shut down an automation on a single audit finding.
That environment forces a compressed learning curve. The co-founders know this terrain from years inside academic medical centers and regulated enterprises, and they use the internship as a controlled exposure program. Interns work through the same audit-workflow-shadowing process that full-time forward-deployed engineers run with hospital staff, mapping back-office tasks before they write a single line of agent code.
The pipeline doubles as a retention filter. Engineers who complete the internship have already operated inside the constraints that make healthcare AI fail: fragmented APIs, vendor-specific Epic customizations, state-level privacy statutes, and the cultural resistance of staff who have seen "AI solutions" break their workflows before. Those who stay become the bench for Wedge's department-by-department expansion strategy: each new hospital engagement needs an engineer who can land on day one and start the comprehensive automation identification audit that defines ROI targets for every agent.
Compensation sits below typical Big Tech intern packages, but the trade-off is direct ownership of agents that go live in production healthcare environments, not sandbox demos. For a two-person founding team backed by Y Combinator, the internship is the primary talent-acquisition channel for the forward-deployed engineering corps that differentiates Wedge from point-solution vendors and outsourced consultancies alike. The next cohort will ship the agents that automate the next back-office department (and the one after that) under a single operating system that Wedge monitors and maintains "forever," per the founders' own language.
Compliance Is Now a Continuous Engineering Discipline
This regulatory stack creates a compliance burden that generic AI platforms cannot absorb. A vendor selling a model and walking away leaves the hospital owning every gap: PCCP maintenance, FHIR interoperability proof, prior-auth documentation, state-law disclosure, vendor-contract remediation, and clinician-training records. Wedge architected its trust layer for exactly this stack: forward-deployed engineers embed in the health system to map workflows to regulatory requirements, build agents that produce auditable decision trails, maintain PCCP-compliant retraining loops, and keep vendor contracts current as rules shift. The platform's governance module tracks model versions, data lineage, and access logs across every AI touchpoint that processes PHI, satisfying HIPAA's extension to training pipelines and third-party MLOps tools.
The federal deregulatory current — Trump executive orders in January and December 2025 directing agencies to reduce AI barriers, and ONC's proposed HTI-5 rule that would cut more than half of current certification criteria, does not erase the compliance obligation. It creates a dual-track reality: federal floors may drop while state ceilings rise, and hospitals operating nationally must satisfy both. Wedge's model treats compliance as a continuous engineering discipline, not a one-time certification. That distinction is why health systems piloting at dozens of sites are treating the trust layer as infrastructure, not an add-on. The next regulatory deadline is always six months away. The only sustainable posture is a team that lives inside the workflow and owns the audit trail.
Working in frontier tech? Zero G Talent tracks the openings: see every open ASML role, browse frontier tech jobs, openings at Stripe, and the people building the field.