Bunkerhill Health has raised $55 million to expand Carebricks, a platform that lets health systems build AI agents for clinical, operational and administrative work. The Series B attracted Khosla Ventures alongside existing backers including Sequoia Capital, Felicis, Optum Ventures and Y Combinator.
The investment story is less important than the deployment evidence. Healthcare has no shortage of promising models that never leave a pilot. Carebricks is already supporting agents inside Cleveland Clinic, Intermountain Health and the University of Texas Medical Branch.
UTMB shows what production deployment looks like
UTMB has more than 20 agents live across its organisation. They review imaging, track follow-up, prioritise referrals and automate administrative tasks that consume scarce staff time. That footprint suggests a platform becoming part of daily operations rather than remaining an isolated experiment.
In its first month, a coronary-calcium agent using an FDA-cleared algorithm reportedly flagged a patient at imminent risk of a heart attack. Clinicians confirmed the concern and performed a triple bypass. The case is powerful, but it is one event—not evidence of performance across a population.
A nephrology triage agent ranks patients by severity and routes less urgent cases towards telemedicine. UTMB reports that average specialist waits fell by more than 50%. A lung-nodule agent follows incidental CT findings to appropriate care, with an 80% faster response for urgent cases and twice the rate of guideline-concordant follow-up.
These figures come from Bunkerhill and the health system rather than an independent audit. Still, they concern live production processes, which makes them more informative than synthetic benchmark scores. They also reflect one organisation’s staffing, data and workflow design and should not be assumed transferable without local testing.
A platform expands opportunity and responsibility
Carebricks differs from a fixed application because hospitals can create agents around ideas proposed by their own teams. One department might detect early cardiac risk; another might manage prior authorisations or update a disease registry.
That flexibility addresses the gap between what clinical leaders know could improve care and what overextended staff can execute. It can also multiply risk. A department that configures its own triage logic inherits responsibility for thresholds, monitoring and unintended disparities.
Hospitals need a formal path from idea to production. It should define the clinical owner, intended population, data sources, performance target and human intervention point. Validation must examine false positives and false negatives across demographic groups, not only average time saved.
Every agent requires a failure mode. If imaging data is incomplete, an integration breaks or confidence falls below an agreed level, work should return visibly to a person. Silent degradation is especially dangerous when clinicians begin trusting that the system will surface every important case.
Bunkerhill says the funding will expand Carebricks into more use cases and strengthen governance, monitoring and safeguards. Those functions are not supporting features; they determine whether a 20-agent deployment remains manageable when it becomes 100.
Scale should be measured in outcomes, not agent counts
Health-system boards need clear answers about liability and escalation. Who owns a missed finding? How often is performance reviewed? What happens when an agent and a clinician disagree? Can a department alter logic without central approval, and can every decision be reconstructed later?
Operational metrics such as wait time and coordinator workload matter because they show whether an agent relieves a real constraint. Clinical metrics matter more: follow-up completion, time to treatment, adverse events and patient outcomes. Both need baselines and long enough observation to identify new failure patterns.
Human capacity should be redirected, not merely removed. When an agent handles registry updates, staff can contact patients who are difficult to reach or resolve cases that do not fit standard pathways. If automation only leads to thinner staffing, the system may have less resilience when exceptions accumulate.
Carebricks gives Bunkerhill a credible position in healthcare AI because it has crossed the boundary into working hospitals. The next test is whether decentralised creation can coexist with consistent institutional control. A platform becomes valuable when it helps clinicians operationalise more good ideas; it becomes trustworthy only when each agent remains observable, contestable and accountable after the novelty has passed.
