The CX universe takes on healthcare with agentic AI
Agentic AI is moving from theory into day-to-day healthcare. Providers, life sciences teams and payers are testing agents to fix wait times, reduce leakage, match patients to trials and support underfunded SDOH work - all while staffing is tight and costs keep rising.
It's still early. Most teams are figuring out where agents fit in real workflows, how to connect them to messy data, and how to measure outcomes without adding risk.
Adoption reality: interest is high, usage is uneven
Salesforce reports roughly 12,500 customers using Agentforce out of about 150,000. Interest is broad, but many leaders are still deciding which use cases are worth operationalizing.
As Marc Benioff put it, this is a moment where tech is outrunning customer adoption. The message: listen harder, build for outcomes, and earn trust one workflow at a time.
What's working now: access, scheduling and call deflection
Las Vegas-based Steinberg Diagnostic Medical Imaging runs MRIs late into the night. That only works if referrals, benefits and scanner availability line up without bottlenecks.
With Genesys-hosted agents, SDMI now handles about 4,000 more calls per month, dropped abandonment from 10% to under 3%, and books appointments 24/7. During business hours, the virtual agent answers basic questions so humans can focus on edge cases and higher-stakes conversations.
As Rachel Papka noted, a large share of calls are simple status checks. Agents clear that queue without making older patients wrestle with a portal.
Workflow first, tech second
Healthcare leaders care about leakage, no-shows, clinician utilization and closing care gaps. Those are workflow problems.
Tara Mahoney from Genesys put it plainly: no nurse wakes up asking for an "agentic AI platform." They want discharge, triage and post-care workflows that actually work.
SDOH and clinical documentation support
Adobe Population Health connects patients on Medicare, Medicaid and ACA plans to care and to SDOH resources like transportation and food. Funding is tight, so efficiency matters.
The team is building a Salesforce-based agent to handle documentation during hour-long risk assessments. That frees social workers and nurses to focus on the person in front of them - and craft better plans - instead of typing into a screen.
Industry agents and the data glut
Salesforce is building agents into Healthcare Cloud and Life Sciences Cloud to automate common processes. It's a practical move in a sector where a typical hospital system touches dozens of data sources and most of the information is unstructured.
The takeaway: agents help when they're pointed at clear business goals and have access to the right data with the right guardrails.
Commercial teams: smarter prep, better recommendations
Haleon plans to use agentic AI so reps can walk into pharmacies and dental offices with product recommendations grounded in local demand and prior behavior. Today, that prep takes too long and varies by rep.
Automating the research makes each visit tighter and more relevant, and it scales what the best reps already do well.
Patient experience, stitched across episodes
Qualtrics acquired Press Ganey Forsta, bringing patient-experience expertise and a large healthcare dataset under one roof. That matters because HCAHPS and related surveys tie directly to reimbursement and quality improvement.
Authoritative overview: HCAHPS Program.
Qualtrics is working with Stanford Health Care on agent use cases like language support and removing conflicting instructions. Picture the basics handled before the visit: a note that a patient wants help from the parking lot, or that a caregiver needs a translated after-visit summary. Fewer surprises, less friction.
Faster access to therapy and financial help
Deloitte sees strong traction in patient support programs: verifying benefits, matching patients to assistance, onboarding them quickly and making sure the pharmacy actually has the drug in stock.
Prebuilt agents are landing on the Salesforce AppExchange to speed this up. Some organizations are going enterprise-wide with centralized approvals; others start with a small pilot and expand once the metrics look good.
Your 90-day plan
- Pick two workflows with clear ROI (e.g., scheduling + post-discharge outreach). Define the success metric before you build.
- Map data access and PHI boundaries. Decide what the agent can see, write and escalate.
- Choose a runtime that plays well with your stack (EHR, CRM, contact center, knowledge bases). Keep the first integration list short.
- Ship a pilot in under eight weeks. Cap scope, script handoffs to humans and log everything.
- Train staff on prompts, failure modes and escalation. Put a name and purpose to the agent so patients aren't confused.
- Review outcomes weekly. Keep what works, cut what doesn't, then expand one adjacent workflow.
Guardrails you cannot skip
- HIPAA/PHI controls: data minimization, encryption, access logging and clear data-retention rules.
- Grounding and citations: retrieve from approved sources, cite them, and block free-form answers where precision matters.
- Safe handoffs: confidence thresholds with instant transfer to a human. No dead ends.
- Prompt security: filter inputs, sanitize outputs, and monitor for jailbreak attempts.
- Auditability: store prompts, sources, responses and user feedback for every interaction.
- Model quality: evaluate for accuracy, toxicity, bias and reading level across key populations.
- Accessibility: multilingual support, plain-language outputs and screen-reader friendly content.
- Vendor diligence: data residency, subprocessor lists and Business Associate Agreements where required.
- Incident playbook: rollbacks, alerting, patient notification and remediation steps.
Helpful reference: NIST AI Risk Management Framework.
Metrics that matter
- Access: call abandonment, time to first available appointment, referral-to-visit time.
- Throughput: calls handled, self-service completion rate, average handle time.
- Leakage and no-shows: in-network retention, confirmation rates, same-day fill.
- Clinician time: minutes spent documenting, time from visit end to note sign-off.
- Quality: documentation accuracy, instruction consistency, language match rate.
- Cost: cost per contact, cost per scheduled appointment, cost per support case.
- Experience: HCAHPS top-box where relevant and average time to resolve patient feedback.
- Therapy access: time-to-therapy, prior-auth cycle time, adherence over 30/60/90 days.
Build the stack with intent
- Channels: phone, SMS, web, portal - meet patients where they are, not where your org chart lives.
- Identity and consent: clear opt-in, consent tracking and caregiver access flows.
- System connections: EHR, claims, CRM, contact center, knowledge bases, and where applicable PACS or specialty systems.
- Safety layers: policy guardrails, red-team checks and human review for high-risk actions.
- Analytics: real-time dashboards with drill-downs to transcripts and source documents.
- Change management: frontline input, simple playbooks, fast feedback loops.
Agentic AI isn't magic. It's operations with new tools. Start with one painful workflow, wire in the data you trust, measure the lift and expand from there.
If your team needs structured upskilling on agent workflows by role, explore curated options here: AI courses by job.
Your membership also unlocks:
