AI in Healthcare: A Practical Guide for Small Practices.

It can be, but it depends entirely on the implementation. Cloud-based AI services like OpenAI, Anthropic, and Microsoft can be HIPAA-compliant when the vendor signs a Business Associate Agreement (BAA) and the data flow is architected appropriately. Some specific tiers of these services explicitly support PHI; others don't. For workloads where data sensitivity is highest (sensitive case discussion, document analysis, clinical notes), local LLMs running on the practice's own hardware are the cleanest path because PHI never leaves your network. Most healthy compliance postures end up as a hybrid: cloud AI for non-PHI logistics (scheduling, marketing, public-facing content), local LLMs for anything sensitive.

Five things, mostly operational and admin: (1) AI receptionists that capture every inbound call including after-hours; (2) automated treatment plan and case acceptance follow-up sequences; (3) automated recare and recall to keep patients on schedule; (4) no-show reduction with smart reminders and automated waitlist filling; (5) AI-assisted documentation and insurance verification. Clinical AI (diagnostic imaging, decision support) exists but is largely the domain of hospital systems and large groups. The operational layer is where independent practices win.

A focused operational AI install for a small practice typically lands in the mid four figures to low five figures, depending on call volume, integration complexity, and which modules you start with. AI receptionist alone is usually low to mid four figures. Adding treatment plan follow-up, recare automation, and no-show reduction brings it to mid five figures. Local LLM installations for sensitive workloads are quoted separately based on hardware sizing and use cases. Most healthy practices see the install pay for itself inside the first quarter through recovered missed-call appointments and case acceptance lift alone.

Modern voice AI is genuinely good. Patients usually realize they're talking to an AI, but they almost never object when the agent is competent, helpful, and gets them what they need quickly. Configuration matters: voice tuning, pacing, escalation logic for complex cases, and clear handoff to a human when the conversation moves into anything that needs one. The agent should never attempt clinical or medical advice; that boundary is configured firmly by design.

Case acceptance is overwhelmingly an operational problem, not a clinical one. The case is presented well. The patient is interested. They walk out without booking because they want to think about it. What happens next determines acceptance: if there's a polite, persistent, automated follow-up sequence over the next 30-60 days addressing common objections (cost, recovery, scheduling), acceptance rates climb 5-15 percentage points. Without that sequence, most undecided cases drift into never. AI handles the sequence so your treatment coordinator doesn't have to manually chase every undecided case.

Yes, especially for repetitive verification work. Daily morning verification batches, eligibility checks, and routine claim follow-up can be automated with appropriate compliance safeguards. The work that requires human judgment (denied claim appeals, complex coding decisions, AR triage) stays with your team, but the rote 60-70% of the work the team currently does can be handled in the background.

Cloud AI (Claude, OpenAI, Microsoft) is faster to deploy, more powerful, and appropriate for non-PHI workflows or when the vendor signs a BAA covering specific use cases. Local LLMs (Llama, Mistral) run on hardware in your practice. They're slower to set up, require infrastructure investment, but PHI never leaves your network. Most practices end up with a hybrid: cloud for scheduling, marketing, public-facing chat; local LLMs for any workflow involving sensitive clinical or financial PHI.

Take the AI Opportunity Score. It's a 60-second quiz that uses healthcare-specific KPI benchmarks (case acceptance, recare retention, no-show rate, response time) to estimate your annual recovery opportunity based on your actual numbers. No signup required. Most healthy practices find that operational AI represents 5-20% of current annual revenue in recoverable leakage.

Operational AI for small practices is real and shipping today. The technology has been ready for two years; the install playbooks are mature; the ROI math is consistent. Clinical AI (diagnostic imaging, decision support) is real but mostly limited to hospital systems with the data and budget to deploy it. The operational layer is what's accessible right now to a Plano dentist, a Frisco vet, or a Highland Park plastic surgeon.

AI receptionist, almost always. It captures every inbound call (including after-hours), it integrates cleanly with most practice management systems, and it produces measurable ROI inside the first month through recovered missed appointments. Once that's running, the natural next moves are automated case/treatment plan follow-up and recare automation. Each install builds on the data and infrastructure of the previous one.