AI for post-acute and home-based care
Practical agents that do the first pass on repetitive work while your clinicians, coordinators and billers make the calls. Organized by line of business, because Hospice and PCS don't share the same problems.
Home Health
- Referral intake triageSummarize referral packets, pull diagnoses and orders, and flag what's missing before acceptance.
- Documentation reviewCheck visit notes against the plan of care and flag gaps for clinical review before billing.
- Coding supportSurface likely diagnosis and coding gaps for your coders to confirm. The coder decides.
- Visit utilizationTrack planned versus completed visits by episode and alert before a missed visit becomes a problem.
Hospice
- IDG meeting prepSummarize each patient's changes since the last interdisciplinary group meeting.
- Eligibility documentationOrganize the clinical record so that recertification support is easy to find and review.
- Cap and length-of-stay analyticsWatch cap exposure and length-of-stay trends by program and referral source.
- Family and bereavement outreachScheduled, personal follow-up calls and messages, with staff reviewing anything sensitive.
Personal Care and CDS
- Medicaid eligibility batchesCheck every client's coverage on a schedule and route lapses to the right person.
- EVV exception handlingClassify missed clock-ins and mismatches, and draft the fix for staff to approve.
- Authorization and unit trackingUnits used versus authorized, by client and payer, with alerts before you over-deliver.
- Shift fill outreachVoice and text agents contact available caregivers when a shift opens.
Private duty care and nursing
- Caregiver matchingRank available nurses and aides by skills, location, continuity and hours.
- Hours versus authorizationKeep pediatric and adult PDN hours inside what's authorized and billable.
- Family communicationConsistent updates on schedules and changes without adding staff time.
- Private-pay invoicing checksCatch billing mismatches before they reach the family.
Revenue cycle
- Claim scrubbingCatch missing data, authorization mismatches and payer-specific rules before submission.
- Denial classificationSort denials by root cause and fixability, then route each to its owner.
- Payment reconciliationMatch remittances to claims and surface underpayments.
- Payor-side supportPrior authorization review support and encounter data quality checks for health plans.
Post-acute operations and analytics capability map
The measures and workflows I build across service lines, from the first referral to the last dollar collected. Every item here is something I've delivered in production.
Enterprise measurement
- Process optimization metrics
- Process flow metrics
- Cross-service-line benchmarking
Operational value chain, referral to cash
Sales
- Referrals
- Non-admits
- Calls
- Referral source conversion
Intake
- Admissions
- Holds and resumes
- Recertifications
- Discharges
- Average minutes to schedule
- Eligibility and benefit verification
- Authorization capture
Scheduling
- Planning
- Tracking
- Management
- Resource loading
- Verification
- Authorization optimization
- Missed and unstaffed visits
Care management
- Visit performance
- Documentation timeliness and tracking
- Plan of care optimization
- Census
- LUPA rate
- OASIS submission timeliness
- Hospice HOPE submission
Billing
- Billed dollars and hours
- Unbilled dollars and hours
- Payments
- Adjustments
- NOA and NOE timeliness
- Claim submission lag
Collection
- A/R
- Appeals and grievances
- Collections
- DSO
- Denial root cause
- ADR and TPE response
Supporting functions
Resource management
- Hires
- Terms
- Open positions
- Turnover rate
- Credentialing and licensure expiry
Payroll
- Utilization
- Travel time
- Mileage
- Overtime and agency spend
Finance
- Revenue
- Cash flow
- Cost per visit and patient day
- Margin by service line
Compliance foundation
- Auditing
- PEPPER target area monitoring
- Face-to-face encounter compliance
- Documentation integrity
AI agents I've built
Seventeen production agents, described generically. Each one follows the same pattern: shared configuration, secrets in Key Vault, a statistics-only or dry-run mode, and an LLM only where it earns its place.
Intake and documents
Referral packet parser
Reads faxed and PDF referral packets with OCR, then extracts demographics, insurance, orders and diagnoses into structured intake data.
Azure Document Intelligence, Tesseract, LLM extraction
Human review
Care plan and authorization extractor
Pulls care plan and authorization details from spreadsheets whose layouts drift file to file, using a rules pass first and an LLM to settle the hard cases.
Python, Claude on Microsoft Foundry, config-driven fields
BAA-covered model
PHI redaction agent
Detects and redacts patient identifiers in PDFs using an editable policy, combining an LLM pass with deterministic pattern matching.
LLM plus regex, local model option
Dry-run mode
Workbook unifier
Merges inconsistent worksheets into one clean load file. The LLM sees only column headers, never patient rows.
pandas, LLM schema mapping
Headers only to LLM
Clinical risk
Hospice transition risk scorer
Scores active home health patients for hospice appropriateness using decline, comorbidity and diagnosis features, with drivers and suggested case manager actions.
Snowflake, feature engineering, LLM scoring
Clinician decides
LUPA risk agent
Explains why 30-day periods fall under their LUPA threshold (scheduling gaps or lost visits) and flags active periods at risk while there's still time to act.
Python, statistics, case-mix-adjusted odds ratios
Aggregates-only output
Revenue and growth
PDGM period analyzer
Profiles 30-day periods, detects revenue issues, compares episodic and non-episodic payers, and writes an executive summary with charts.
Python, pandas, LLM summary
Statistics-only mode
Census growth planner
Turns an annual growth goal into monthly referral and admission targets by agency, using conversion, discharge and seasonality history. Re-plans mid-year against actuals.
Python, SQL, Power BI
Writes to the planning model
Data quality and master data
Volume variance checker
Scores every service line, agency, branch and measure against its own history and reports outliers, missing weeks, stale feeds and sudden zeros.
Python, statistics
No LLM
Power BI anomaly detector
Queries semantic models on a schedule and flags spikes, drops, gaps and outliers. Fails the job when something critical turns up.
DAX, Power BI REST API
No LLM
Duplicate patient matcher
Finds likely duplicate patient records and classifies each pair with a confidence score. Merges happen only after review.
SQL, LLM matching
Human review
Provider NPI matcher
Fills in missing referral-source NPIs from the CMS NPI Registry. Auto-updates single high-confidence matches and queues the rest for review.
NPI Registry API, LLM scoring
Review queue
Document integrity checker
Verifies that stored clinical PDFs are complete and readable, and queues broken ones to be fetched again.
Python, PDF validation
Dry-run by default
Analytics
Natural-language Power BI analyst
Answers business questions by writing DAX against the live model, correcting its own query errors, and returning a written analysis.
Claude, DAX, Streamlit UI
Read-only
Operational performance dashboard
Builds a weekly census, admissions and discharge dashboard for Home Health, Hospice and PCS, with flagged issues and year-end projections.
Python, SQL, self-contained HTML
Aggregates only
Voice and engineering
Outbound call agent
Places personalized outbound survey and confirmation calls with a conversational voice agent, then writes transcripts and outcomes back to the call record.
Voice AI, Twilio, Azure Functions
Staff handoff
.NET error remediation agent
Reads application error logs, analyzes exceptions against the source code, proposes fixes, and opens pull requests only after approval.
Claude, GitHub, Docker, Prometheus
Approval workflow
The shared pattern behind them
- Pluggable models. Claude on Microsoft Foundry, Azure OpenAI or local models through one provider layer, so PHI stays with BAA-covered or on-premises services.
- Safe defaults. Dry-run, review queues and "no LLM" modes come first. Anything that writes or deletes needs an explicit flag.
- Configuration, not code. Fields, thresholds and measures live in config files, so adding one is a one-line change.
- Publishable output. Aggregate reports are separated from PHI worklists, and only the aggregates are ever published.
Platform tooling
The scripts that keep a data platform healthy and documented, so the agents have something reliable to stand on.
- Pipeline controlMonitors Azure Data Factory runs, surfaces failures, extracts the SQL behind each pipeline and re-runs on demand.
- Cloud cost reportingBreaks Azure spend down by resource and month for showback and right-sizing.
- Data dictionary generatorDocuments every table, column and relationship straight from the database.
- Flow documentationExports Power Automate and RPA flows, maps their dependencies, and writes readable documentation.
- Power BI administrationTriggers dataset refreshes, audits users and access, and loads the activity log into Fabric.
- Report publisherPublishes agent-generated reports to an identity-protected static web app on a daily schedule.
How I keep it safe
Patient trust and compliance come first. Every engagement follows these rules.
- BAA-covered services onlyPHI only goes to AI services covered by a Business Associate Agreement. Never to consumer chat tools.
- Your cloud, your dataSolutions run in your own Azure tenant where possible, so data never leaves your control.
- People approveAgents draft, sort and flag. Clinical, coding and billing decisions stay with your staff.
- Minimum necessaryEach agent sees only the fields it needs to do its job.
- Full audit trailEvery AI action is logged with its inputs, outputs and who approved it.
- Measured before scaledEach pilot is compared to a baseline. If it doesn't beat the old way, we stop.
Domain knowledge, written down
AI only helps when it's built on the right definitions. These explainers show how I think about the numbers agencies run on.
Census, ADC and patient days: getting the basics right
Why an end-of-week census and an average daily census tell different stories, and how each service line counts.
PDGM from intake to cash
The data captured at the door, how a 30-day period is paid, and how cash is posted and recognized.
Not sure where to start?
An AI readiness assessment maps your workflows, data and systems, then ranks use cases by value and effort. You leave with a roadmap and a first pilot ready to build.
You get
- Workflow maps for intake, scheduling, documentation and billing
- A ranked list of use cases with estimated hours and dollars saved
- A data and security review of your current systems
- A pilot plan with success measures
Find the first AI win in your agency.
A free 30-minute call. Bring your messiest workflow.