Three autonomous Agentforce agents handling patient follow-up, appointment management, and care plan updates — deployed inside an existing Salesforce Health Cloud org in 8 weeks.
The client, a multi-clinic primary care network operating across 14 locations in Western India, had invested significantly in Salesforce Health Cloud two years prior. The CRM held comprehensive patient records, care plans, and appointment histories — but the processes built around it were still entirely manual. Care coordinators spent an estimated 60% of their working day on routine follow-up: post-appointment check-in messages, appointment reminders, care plan update notifications, and prescription renewal prompts.
The consequences were measurable. Patients who didn't receive timely follow-up were 3× more likely to miss their next appointment. Care coordinators, overwhelmed by volume, were prioritising by instinct rather than clinical risk — meaning high-risk patients weren't always getting the most timely attention. Staff turnover in the care coordination team was running at 34% annually, with burnout cited as the primary reason in exit interviews.
The ask was specific: use Agentforce to automate the routine, so care coordinators could focus on the cases that genuinely needed human judgment.
Rather than building a single broad-purpose AI agent, we designed three specialised agents with clearly bounded responsibilities. This approach — narrow by design — produces more reliable output, easier testing, and faster iteration cycles.
Triggers 24 hours after any completed appointment. The agent reviews the appointment notes and care plan context in Health Cloud, generates a personalised follow-up message (reviewed and approved by the coordinator before sending), and logs the interaction automatically. For patients with follow-up instructions, the agent drafts a structured summary referencing their specific care steps.
Coordinators reported that reviewing and approving agent-drafted messages took an average of 40 seconds — compared to 4–6 minutes to write them from scratch. For a coordinator handling 45 patients per day, this was the largest time saving of the three agents.
Handles appointment reminders, cancellations, and rebooking prompts autonomously within defined parameters. The agent sends reminders at 72 hours and 24 hours, handles standard cancellation acknowledgements, and identifies patients who have missed an appointment and initiates a rescheduling workflow — escalating to a human coordinator if three attempts fail or if the patient record carries a high-risk flag.
The escalation logic was the most critical design decision: the agent was built to be conservative, handing off to humans at the first sign of complexity. This made the clinical team comfortable trusting it with real patients.
When a clinician updates a care plan in Health Cloud, this agent generates a plain-language patient notification summarising what changed and why — based on the clinical note and plan delta. It also identifies any tasks the patient needs to take (e.g., schedule a new test, stop a medication) and adds them to the patient's task list in the patient portal.
This agent required the most iteration. Clinical note quality varied significantly across the 14 locations, and the prompt engineering had to be robust enough to handle sparse notes while conservative enough to avoid overinterpreting ambiguous language. We ran 200 test cases before declaring it production-ready.
At 90 days post-deployment, routine follow-up tasks had dropped by 52%. Care coordinators were spending the recovered time on higher-complexity patient interactions, proactive outreach to at-risk patients, and — notably — taking their full lunch breaks for the first time in months. Staff satisfaction scores improved by 28 points in the quarterly survey.
The biggest resistance we encountered at the start was from care coordinators who feared the agents would make errors that patients would experience. We addressed this by designing every agent output as a human-reviewed draft by default — AI generates, human approves, system sends. As trust was established, approval requirements were gradually relaxed for low-risk communication types.
The agents are not clinicians. They don't make clinical decisions. They handle the administrative surface area that was preventing coordinators from doing the clinical work they were trained for. That framing — protecting clinical judgment rather than replacing it — was the key to adoption.
"I used to feel like an email robot. Now I feel like a care coordinator again. The AI does the reminders — I do the care."
— Senior Care Coordinator, Healthcare Client