Home care AI experimentSeptember 30, 20267-minute read

We tried to break home care AI with a year of after-hours chaos.

71,296 workflow events. Four offices. Caregiver conversations in six languages. We deliberately pushed HeyHomeCare through difficult situations to find what needed fixing before asking an agency to trust it.

By the HeyHomeCare team. A synthetic experiment with fictional people, scripted human responses and simulated communications; these are not live agency results.

The experiment at a glance
The 365-day calendar completed with 62,384 resolved records and 8,912 unresolved handoffs retained for staff. Separate conversation and coverage tests exposed problems and led to corrections. Testing continues; the experiment did not measure labor savings.
The workload

A difficult night, repeated across a year.

A caregiver is still driving. Another can stay with a client for 30 more minutes, but cannot commit to the entire night. A family needs an update. The usual on-call contact is unavailable. Tomorrow’s schedule still has to make sense when the office opens.

For an agency with hundreds of caregivers and several offices, each situation involves people waiting on an answer. We wanted to see whether the work would stay organized when interruptions overlapped and the easy answers disappeared.

After our chaotic-month experiment, we created a fictional agency with 480 caregivers, 480 clients and four offices. Its simulated calendar covered 365 days, with extra pressure on weekends, holidays, winter illness periods and days when the primary communication route failed.

The calendar exercised missing clock-ins, missing clock-outs, visits ending, late arrivals, schedule-change requests and messages for clients or authorized contacts. Separate tests covered call-off coverage, competing responses and AI conversations.

These were deliberately selected problem cases. They were not an agency’s complete schedule, a recorded year of calls or an estimate of how often each problem occurs.

The results

A year of work. A record of what remained.

71,296Official workflow records
279,249Audit entries recorded
62,384Records resolved under scripted conditions
8,912Unresolved handoffs with open staff tasks

The calendar also produced 240,576 simulated communication requests. These were test requests, not actual phone calls, delivered texts or conversations with caregivers.

Of the schedule requests, 4,456 met the configured automatic-change limits, 2,228 received scripted office approval and 2,228 were declined without changing the schedule. Repeated requests did not apply the same edit twice.

We deliberately left message relays without timely answers. All 8,912 stayed unresolved with open office tasks. A late response did not erase work that still needed the office. That distinction matters when a morning team inherits the night.

People’s responses and office decisions were scripted. The resolved total is a test result, not a measured autonomous-resolution rate at a real agency.

The difficult cases

Small words can change the whole decision.

“I’m still driving” needs to remain a late-arrival situation. It should not start replacement coverage as though the caregiver had called off.

“I might accept if the rate changes” needs clarification. It should not become an accepted shift or a final refusal before the caregiver decides.

“I can stay another 30 minutes” has a real limit. The office needs that limit visible, with a person reached if relief has not been confirmed.

These cases exposed problems in testing and led to corrections that have been built and retested. The test results describe those selected cases. A controlled agency evaluation still needs to verify the full workflow under real conditions.

A useful evaluation should ask the vendor to demonstrate these distinctions with your agency’s rules. A general promise to “handle call-offs” leaves out the decisions your on-call team actually makes.

Six languages

The conversation tests were a separate challenge.

The AI checks used authored caregiver conversations in English, Spanish, Mandarin, Vietnamese, Tagalog and Arabic. The selected contextual checks passed in the retest. We also tested synthesized speech in clean and degraded conditions.

Noisy speech still sometimes lost words or meaning. Five adversarial conversations in a later validation run were blocked by the AI provider and were not counted as passing.

The calendar’s workflow totals do not count successful AI conversations. These separate checks do not certify real accents, native-speaker performance or every way a new call or message enters the system.

For an agency serving multilingual caregivers, a pilot should include the people, language choices and phone conditions your team encounters—and a clear path to a person when understanding is uncertain.

Your agency’s opportunity

What could your team do with more capacity?

More time for new clients. Fewer interruptions for an operations lead. More capacity to support another office. Those are outcomes to measure against your current workload, including the approvals and follow-ups people still handle.

This experiment did not measure labor hours saved, cash savings, live phone capacity or results at a customer agency. Use your own assumptions below to explore a potential opportunity, then validate them in a controlled pilot.

Start with your own workload. The example below assumes 1,000 events, six staff minutes per event and a possible 50% reduction. That reduction is an assumption for you to change; it was not measured in the experiment.

Estimated staff capacity50 hours / month
Equivalent labor value$2,000 / month

Your inputs imply 100 hours of current monthly staff work.

Planning estimate only. Labor value represents capacity, not cash saved. The calculation excludes HeyHomeCare and communication costs. Actual results depend on the workflow, approvals and human follow-up; measure them in a pilot. Your inputs stay in this page and are not submitted with the review request.

Review the opportunity at my agency
For owners and operations leaders

Make the next test your own operation.

Useful automation should keep routine work moving, preserve the agency’s decisions and make exceptions clear enough for a person to act. You need to know what happened, who was reached, what changed and what still needs attention in the morning.

  • Start with your workload: call volume, offices, languages and the situations that interrupt staff.
  • Name the decisions: who can approve a change and who owns an unanswered escalation.
  • Agree on evidence: actual delivery, verified schedule changes, human follow-up and a usable morning record.
  • Measure the outcome: staff time, coverage, response time and total operating cost.

HeyHomeCare offers a primary home care CRM for private-pay, non-medical agencies with AI capabilities optional, and an after-hours operations solution for agencies keeping their existing systems. Available connections and permitted actions are confirmed for your agency.

Request an operations review. We’ll discuss your call-off process, office structure, escalation rules and communication volume, then agree on what a controlled evaluation should measure.

FAQ

Frequently asked questions.

Q.01
Did AI handle 71,296 real calls?
No. The calendar contained 71,296 authored workflow events with fictional people and scripted responses. Its communication requests were simulated. AI conversation and speech checks were separate tests.
For large and multi-office agencies

What would change in your after hours?

Bring your workload and business rules. We’ll identify a practical starting point, the decisions your team retains and the outcomes to measure.

Your agency, your numbers

What would AI change in your after hours?

Request a 25-minute operations review. Bring your call volume, staffing process and on-call rules. We'll identify workflows to evaluate, decisions your team should retain and the results to measure in a pilot.

Reply from a real human within one business day
Discuss your offices, caregiver volume and current scheduling system
Define the approval and backup plan before discussing automation

Request your operations review.

A few details · 30 seconds
Next step: pick a time on the calendar · We reply within 1 business day · No marketing list