Shift handover with Gemini Enterprise: the CareShift AI case at a mid-sized hospital

TL;DR A mid-sized Brazilian hospital (240 beds) deployed CareShift AI on Gemini Enterprise to support nursing shift handover. In 90 days: clinical omissions -62%, handover time -40%, team NPS +33 pts. Approved by CCIH, ethics committee, and CRM.

-62%

Clinical omissions
18% → 6.8%

-40%

Handover time
45 → 27 min

96%

Team adoption
after audio TTS

+33

Nursing NPS
41 → 74

Shift handover is the clinical moment where the most serious errors occur. The Joint Commission points to communication failures in up to 70% of sentinel adverse events. Notebook + verbal + WhatsApp dominate — and information is lost at every shift.

This case shows how Autenticare deployed an agent that does not replace the handover — it preserves and verifies it.

The starting point

240 beds, 6 inpatient units + adult ICU + pediatric ICU.
3 nursing shifts, ~80 nurses and 200 technicians per day.
Manual handover (notebook + verbal) averaging 45 min per unit.
Internal audit identified 18% of critical information omitted between shifts.
Communication-related adverse events: 14 in the last year.

What the agent does (and does not do)

Reads the medical record — TASY/MV via API: progress notes, prescriptions, vital signs, pending exams, isolation alerts.

Builds SBAR per bed — Situation, Background, Assessment, Recommendation. International standard adapted to the Brazilian context.

Highlights changes since the last handover — new medication, dose change, abnormal lab result, clinical status change.

Presents on tablet at the bedside with optional audio (TTS) for nurses with occupied hands.

Collects confirmation from the responsible nurse with an audited digital signature.

Flags discrepancies — if the verbal report diverges from the medical record, alerts before closing.

The agent does not decide — it presents and verifies. Clinical decisions remain 100% with the professional.

Architecture

Gemini Enterprise Plus with Vertex AI Agent Builder.
Vertex AI Search indexing: institutional protocols, medication manuals (Bulário), guidelines (CFM, COREN, medical societies).
Dedicated connector to the electronic medical record via Apigee (TASY HL7 + MV REST).
Tools: reading progress notes, prescriptions, vitals, labs; recording signatures; generating handover PDFs.
Tablet app as PWA, integrated with the hospital's Workspace SSO.
BigQuery for audit and quality analysis.

Results in 90 days

Metric	Before	After	Delta
Critical clinical omissions	18%	6.8%	-62%
Average handover time	45 min	27 min	-40%
Communication adverse events (proj./year)	14	5	-64%
Nursing team NPS	41	74	+33 pts
Adoption (handovers with agent)	—	96%	new
Nurse time with medical record	22 min/shift	9 min/shift	-59%

The most important figure for clinical management was not time — it was the drop in adverse events. Every avoided event is a life preserved (and reduced legal exposure).

Healthcare governance: what needed to be ready

1. Ethical approval

The hospital's research ethics committee approved the use after reviewing the protocol. This does not replace CEP/CONEP for research, but is an institutional requirement.

2. Anonymization in logs

Operational logs contain no patient-identifiable data — only bed + event. Identification remains exclusively in the medical record.

3. DPIA focused on sensitive data

Health data is sensitive under Art. 11 of the LGPD. Detailed DPIA following our template — see DPIA for Gemini Enterprise projects.

4. CFM/COREN compliance

CFM Resolution No. 2.314/2022 and COREN on AI use in healthcare. Principles: AI as support, not substitute; final decision with humans; transparency to the patient where applicable.

5. Business continuity plan

Tablet without network or agent offline: handover reverts to the manual process without disruption. Periodic training maintains competency.

6. CCIH audit

Hospital Infection Control Committee validates that the agent alerts for isolations, contact precautions, and ongoing antibiotic therapy.

⚠️ What we did NOT do (intentionally) We did not automate prescriptions (would require separate certification). We did not implement autonomous Manchester triage (still handled by human nurses). We did not use it for diagnosis (the agent presents what is in the record, it does not interpret clinically). These limits were negotiated with the ethics committee before the design phase. Breaking any one of them would have delayed the project by months.

What we learned along the way

SBAR needed to be regional

The international model did not capture nuances of the Brazilian context (claim denials, insurance type, SUS/private status). We adapted it with the team over 4 cycles.

Audio changed adoption

Nurses with occupied hands (gowning, bed bathing) consume the summary by audio. Adoption rose from 78% to 96% after adding TTS.

Digital confirmation became culture

The digital signature created a sense of individual responsibility. It coincided with the drop in adverse events.

Physicians started using it

They were not the target audience. But after 6 weeks, on-call physicians requested access for quick rounds. A natural expansion.

The agent did not replace nurses — it returned 13 minutes of medical record time per shift so nurses could be with the patient. That, multiplied by 80 professionals and 3 shifts, is an entire team back at the bedside.

CareShift AI for your hospital

Want to replicate these numbers at your institution?

The CareShift standard applies to any hospital with an electronic medical record (TASY, MV, Soul MV, Pixeon, Cerner). Implementation: 60-90 days for the first unit, 30 days per additional unit. Includes DPIA, CCIH, business continuity plan.

Schedule assessment → Gemini Enterprise for healthcare