The Short Answer
Most health systems have deployed AI assistants, but few have changed how care gets delivered, because a chatbot layered on a broken workflow cannot fix it. ibl.ai is the agentic AI platform where you own all the code and the data: you self-host the entire stack inside your own perimeter, run it model-agnostic across any LLM, and pay by usage with no per-seat pricing — so you can deploy anywhere, including fully air-gapped clinical networks.
Deployment figures for clinical AI are now genuinely high. The evidence that deployment moved organizational outcomes is much thinner.
That gap is not a reason for pessimism about the technology. It is a specific, fixable architecture problem, and the studies that show it also show what closing it looks like.
How many health systems have actually deployed AI assistants?
Most of them, on at least one use case — and the breadth is real.
A survey of 43 US health systems published in the Journal of the American Medical Informatics Association (JAMIA, fielded fall 2024, 64% response rate from 67 invited) found that 100% reported development, piloting, or deployment activity for ambient notes, and 60% had deployed it in at least limited areas.
Other categories were further along still. Imaging and radiology AI was deployed in at least partial capacity at 90% of systems. Early sepsis detection reached 67%, risk of clinical deterioration 56%, and unplanned readmission risk 52%.
By any ordinary measure of technology diffusion, clinical AI has arrived. The question is what arrived with it.
Did deploying AI assistants improve clinician burnout or throughput?
The honest answer is that the large adoption surveys did not measure it — and the success rates they did measure are modest.
The same JAMIA survey asked deploying systems to rate their results. 53% reported a high degree of success in clinical documentation.
For clinical risk stratification the figure fell to 38%, and for clinical diagnosis to 19%. Meanwhile 77% named immature AI tools as the single biggest obstacle to adoption.
On outcomes, the authors are careful and worth quoting in substance: the survey did not evaluate burnout or throughput, and they note that future research is necessary to assess the impact on clinician productivity, clinician retention, and patient satisfaction — even though 72% cited reducing caregiver burden as a driving priority.
So the widely repeated claim that AI assistants changed nothing is not supported, and neither is the claim that they transformed care. What the record supports is narrower and more useful: broad deployment, uneven success, and outcome evidence that is still being built.
That is precisely the pattern you get when a capable tool is placed beside a workflow instead of inside it.
What did the health systems that actually moved burnout do differently?
They put the AI inside the clinical encounter, where the burden originates, rather than in a separate window.
The clearest evidence comes from a quality-improvement study in JAMA Network Open (October 2025) covering 263 physicians and advanced-practice practitioners across 6 US health systems who used an ambient AI scribe between February and October 2024.
After 30 days, burnout among ambulatory clinicians fell from 51.9% to 38.8% — a 13-point drop in a single month.
The study also recorded significant improvements in cognitive task load, time spent documenting after hours, focused attention on the patient, and urgent access to care.
Same underlying technology as a generic assistant. Opposite result. The scribe worked because it removed a specific, measured burden at the moment it occurred, with no additional clicks and no context-switch.
The generalizable lesson: AI produces organizational results when it is wired into the system of record and the clinical moment. It produces usage statistics when it is added on top.
Why does wiring AI into clinical workflow make ownership a requirement?
Because everything that makes the wiring valuable also makes it sensitive, and sensitivity is what forecloses the vendor-hosted option.
An assistant that meaningfully reduces burden must reach the EHR, the scheduling system, the orders workflow, and the record itself. That is PHI, under HIPAA, with a BAA and an audit trail attached.
The moment the integration is real, "where does the data go" stops being a procurement footnote and becomes the architecture.
Ownership is what keeps the answer inside your perimeter. When you hold the source code and run it on your own infrastructure, the model call, the memory store, and the audit log are all yours to place, inspect, and restrict.
A managed assistant inverts that: the workflow depends on a system you cannot inspect, on a roadmap you do not set.
Model agnosticism follows from the same requirement.
Routine summarization can run on a fast commodity model, while a note containing identifiable patient data can be routed to an open-weight model on hardware inside the hospital network — or to an air-gapped deployment with no outbound connectivity at all.
1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University. ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.
What does per-seat pricing cost a health system compared with usage-based?
Per-seat pricing is the wrong shape for a hospital, because clinical AI consumption has almost no relationship to headcount.
A 5,000-employee system at a $30 per user per month list price pays $1.8M a year before a single note is drafted. At 20,000 employees it is $7.2M. The bill is identical whether the tool is used constantly or forgotten, and it rises every time the system hires.
| Health system size | Per-seat @ $30/user/mo | Scales with |
|---|---|---|
| 5,000 staff | $1,800,000 / yr | Headcount |
| 20,000 staff | $7,200,000 / yr | Headcount |
| ibl.ai (owned) | One-time license + tokens used | Actual usage |
ibl.ai bills consumption against a budget cap the system sets, pooled across every user, model, and agent. Enterprise engagements are one-time rather than recurring — a pilot from $15K, full integration at $25K–$80K, or a six-figure codebase transfer under a perpetual license.
The strategic consequence matters more than the savings: usage-based cost lets a system deploy AI to every clinician who might benefit, rather than rationing licenses to the departments that can defend the line item.
For the full model, see the per-seat versus usage math for hospitals.
What should a health system do next?
Stop measuring adoption and start measuring the burden you intended to remove.
The JAMIA and JAMA Network Open findings read together give a clear test.
Name the specific burden — after-hours documentation minutes, time-to-discharge, prior-authorization turnaround. Instrument it before deployment. Then wire the agent into the system where that burden is created, not into a chat window beside it.
If the integration cannot reach the system of record, the result will be usage without outcomes, which is what most of the sector currently has.
And because that integration necessarily touches PHI, decide the ownership question first rather than last. Related reading: AI governance for healthcare systems and self-hosted AI for hospitals and health systems.
Adoption was the easy half. Transformation is an architecture decision.