The after-hours cliff is real,
and here is its size.
We called the same 100 Southern California senior living communities three times in one week. On a Wednesday morning, 77.3% of calls reached a live person. On Tuesday night, 60.8% did. On Sunday night, 53.6% did.
77.3%
Reached a live person, Wednesday 10 AM to noon
60.8%
Reached a live person, Tuesday 8 to 10 PM
53.6%
Reached a live person, Sunday 8 to 10 PM
23.7 pts
Gap, the same communities, Wednesday morning against Sunday night
Conflict of interest
Elevyr sells Sloane, an AI admissions agent that answers phone calls for senior living communities around the clock. A poor after-hours number helps us sell it. That is why every rule for this study was registered in public before the first call, why the headline was written in advance for every possible result and why the call-level data is below for anyone to check.
How often does a senior living community answer the phone after hours?
In this sample, about half the time. On a Sunday night, 52 of 97 working numbers reached a live person, or 53.6%. The same communities reached a person on 75 of 97 calls on a weekday morning, or 77.3%.
The drop comes from the hour, not the day of the week. Tuesday night looked a lot like Sunday night. Both looked worse than Wednesday morning.
A human answering service counted as a live person. So did a receptionist who could not answer a single question. A voicemail, a phone menu that never reached a person and an AI voice all counted as no. The rules were fixed before anyone dialed, and the registration is public on OSF.
What share of calls reached a live person each night?
Share of calls to working numbers that reached a live human. The range is a 95% Wilson interval with a finite population correction, as registered. Numbers that were disconnected or out of service are left out and counted below.
| Session | Reached a person | Calls | Share | 95% range |
|---|---|---|---|---|
| Wednesday, 10 AM to noon | 75 | 97 | 77.3% | 69.1% to 83.5% |
| Tuesday, 8 to 10 PM | 59 | 97 | 60.8% | 52.1% to 68.8% |
| Sunday, 8 to 10 PM | 52 | 97 | 53.6% | 44.9% to 62.0% |
The study is powered to detect a large gap. It is not powered to estimate a precise rate. Each night’s share carries a range several points wide, printed next to it. The comparison below is the stronger result.
Did the same communities answer less on Sunday night?
Yes. This is the main test, and it compares each community with itself. 30 communities reached a person on Wednesday morning and not on Sunday night. 7 went the other way. Everything else was the same both times.
That is a gap of 23.7 percentage points, with a 95% range of 12.1 to 35.1 points. An exact McNemar test gives p = 0.00019. A community that was reachable by day was often unreachable on a Sunday night, and the reverse was rare.
| Same community | Sunday: person | Sunday: no person |
|---|---|---|
| Wednesday: person | 45 | 30 |
| Wednesday: no person | 7 | 15 |
The two bold cells are the whole test. Anyone can recompute it from them by hand. Pairs where either number was out of service are left out: 3 communities.
Is it the hour or the day of the week?
The hour. Tuesday night exists in this study to answer exactly this. Wednesday morning against Tuesday night changes only the hour. Tuesday night against Sunday night changes only the day.
| Comparison | Person first time, not second | The reverse | Gap (points) | 95% range | p |
|---|---|---|---|---|---|
| Wednesday morning vs. Tuesday night (the hour) | 22 | 6 | 16.5 | 6.2 to 27.0 | 0.0037 |
| Tuesday night vs. Sunday night (the day) | 19 | 12 | 7.2 | -4.2 to 18.6 | 0.28 |
Moving from morning to night costs a clear share of calls. Moving from Tuesday night to Sunday night may cost a little more, but the range crosses zero, so this study cannot tell a Sunday night from any other weeknight. Both of these are registered secondary comparisons.
When a person answered, could they help?
The caller asked one question of every person who answered: are you taking new residents? At night, close to half of the people who picked up could not say.
| Session | People who answered | Could say whether taking residents | Share | 95% range |
|---|---|---|---|---|
| Wednesday, 10 AM to noon | 75 | 63 | 84.0% | 75.1% to 89.6% |
| Tuesday, 8 to 10 PM | 59 | 31 | 52.5% | 41.6% to 63.2% |
| Sunday, 8 to 10 PM | 52 | 29 | 55.8% | 44.0% to 66.8% |
This is a softer measure than the main result. It depends more on what the caller heard and how the caller judged it, and the people who designed the study have a stake in it. Read it second.
Where the person was
We asked this only when the person who answered could not tell us whether the community was taking new residents. Almost all of them said they were at the community, and none said they worked for an answering service. Read that with caution. The answer is whatever the person told us, answering services often pick up in the community’s own name and anyone who could answer the residents question was never asked where they were. This study cannot tell you how many communities use an answering service after hours. It does not change the main result, because a live person counts as a yes whether they work on site or for a service.
| Session | Could not help, said at the community | Could not help, not sure | Said answering service |
|---|---|---|---|
| Wednesday, 10 AM to noon | 11 (14.7%) | 1 (1.3%) | 0 (0.0%) |
| Tuesday, 8 to 10 PM | 26 (44.1%) | 2 (3.4%) | 0 (0.0%) |
| Sunday, 8 to 10 PM | 22 (42.3%) | 1 (1.9%) | 0 (0.0%) |
Shares are of the people who answered that session.
Did any community answer with an AI voice agent?
No. On every night, 0.0% of calls were answered by a conversational AI. The upper end of the 95% range is 3.6%, so AI answering is at most rare in this sample. Because nothing answered with AI, the registered companion number, calls handled by anyone human or AI, is the same as the main result: 77.3% on Wednesday, 60.8% on Tuesday and 53.6% on Sunday.
Did size, price or our own prospect list make a difference?
The three subgroups registered in advance, and only those. Splitting a sample of 100 leaves small groups with wide ranges, so these are descriptive only, with no significance test.
Licensed beds
| Group | Wednesday: share (range, calls) | Tuesday: share (range, calls) | Sunday: share (range, calls) |
|---|---|---|---|
| 36 to 80 beds | 70.6% (55.6% to 81.4%, 34) | 52.9% (38.7% to 66.6%, 34) | 52.9% (38.7% to 66.6%, 34) |
| 81 to 150 beds | 81.0% (70.8% to 87.6%, 63) | 65.1% (54.2% to 74.3%, 63) | 54.0% (43.3% to 64.2%, 63) |
Published starting rate, from the community’s own website
| Group | Wednesday: share (range, calls) | Tuesday: share (range, calls) | Sunday: share (range, calls) |
|---|---|---|---|
| $6,000 a month or more | 71.4% (39.3% to 88.3%, 7) | 42.9% (19.5% to 71.3%, 7) | 42.9% (19.5% to 71.3%, 7) |
| Under $6,000 a month | 80.0% (62.7% to 89.3%, 25) | 72.0% (54.5% to 83.7%, 25) | 56.0% (39.3% to 71.1%, 25) |
| No rate published | 76.9% (66.6% to 84.2%, 65) | 58.5% (47.8% to 68.2%, 65) | 53.8% (43.3% to 64.0%, 65) |
Operator appears in Elevyr’s own prospect list (a check on us, not a finding)
| Group | Wednesday: share (range, calls) | Tuesday: share (range, calls) | Sunday: share (range, calls) |
|---|---|---|---|
| On the list | 83.3% (70.8% to 90.3%, 42) | 64.3% (50.9% to 75.3%, 42) | 64.3% (50.9% to 75.3%, 42) |
| Not on the list | 72.7% (61.2% to 81.3%, 55) | 58.2% (46.6% to 68.7%, 55) | 45.5% (34.6% to 56.9%, 55) |
One flag, reported as a flag. Communities whose operator appears in Elevyr’s prospect list reached a person more often on Sunday night than the rest. The prospect list leans toward higher-priced operators, which is why this check was registered. The groups are small and their ranges overlap, so it points at the sample, not at a finding.
What else the calls showed
Numbers that no longer work
3 of 100 listed numbers were out of service in every session, after a second-phone check. They are the only calls left out of the results, which is the one exclusion the protocol allows.
Who ended the unanswered calls
A skeptic’s first question is whether the caller hung up too soon. The rule was six rings or 30 seconds. On Sunday night, 27 of the 45 unanswered calls ended because the community’s own phone system sent the call to voicemail or a menu before the rule applied. The caller ended 16. On Tuesday night the split was 22 and 16, and on Wednesday morning 11 and 10. The rest are the corrected calls from the second-phone check, which have no ending recorded.
How long an answered call took
The median answered call ran 44 seconds on Wednesday, 50 on Tuesday and 55 on Sunday. That is measured from the start of the call to the moment the result was logged, so it includes ringing and overstates the time anyone at a community spent on the phone.
Did the checks change the answer?
Two sensitivity analyses were committed to in advance. Dropping the calls corrected after the second-phone check gives the figures from before the correction. Dropping the second call to each of two phone lines shared by two licenses leaves every conclusion in place: the Wednesday against Sunday gap stays clear, and no night moves by a meaningful amount. We describe that one in words because its exact figures would reveal how those two specific lines were answered. Both are in the deviation log below.
What this study does not show
It covers Southern California only. Whether the rest of the state or the country looks the same is unknown, and nothing here claims it does.
Licensed beds are not luxury. The frame is every licensed residential care community in eight Southern California counties with 36 to 150 beds. Size says nothing about price or private-pay mix. Price appears only as a subgroup, read off each community’s own website.
The number we dialed may not be the number a family would dial. The state roster lists each license’s phone number of record, which can differ from the number a family finds by searching. We kept the roster number, because swapping in a search result would put our own judgment back into a mechanical sample.
The roster is dated May 25, 2025, the newest file the state publishes. Some communities have since closed or changed hands. Dead numbers are counted and reported, not hidden.
One caller placed every call and was paid by Elevyr. The caller was never told what we expected to find, and the main outcome, whether a person spoke, leaves little room for judgment. There are no recordings to audit, by design, because California requires two-party consent to record.
Every community heard from the same number three times in four days, in the same order: Sunday, Tuesday, then Wednesday. That number was new, and Google’s spam filter labeled it on some phones. A label could lower pickup, but it was the same on every night, so it cannot explain the gap between nights. If anything, communities that learned to ignore the number would hurt Wednesday, the last session, which shrinks the gap rather than inflating it.
Sunday’s calls were placed in the right window but followed Wednesday’s dial order (DEV-012). The order within a session only spreads calls across the evening, so this is disclosed rather than treated as fatal.
Everything rests on one field week. A single unusual week would move every number together. The week was chosen to avoid holidays.
Every change from the registered plan
A study with no logged deviations is less believable than one that shows its work. Each line below is summarized. The full log is published word for word, with community names, numbers and the caller’s name withheld. Read the full deviation log.
| Entry | Date | What changed |
|---|---|---|
| DEV-001 | 2026-07-31 | The caller changed from a blind contractor to the study’s designer, after two contractors went unresponsive. Superseded by DEV-005 before any call. |
| DEV-002 | 2026-07-31 | Field week moved for lack of a caller. Same seed, same sample, same dial order. |
| DEV-003 | 2026-07-31 | Calling line set to a free Google Voice number with recording off. Superseded by DEV-007. |
| DEV-004 | 2026-08-02 | Field week moved again because the caller could not field the session. Same seed, same sample. |
| DEV-005 | 2026-09-20 | The caller changed to a paid research assistant who is not the designer and was never told the hypothesis. |
| DEV-006 | 2026-09-23 | Field week moved to the week of September 27. Same seed, same sample, covariate file unchanged. |
| DEV-007 | 2026-09-24 | Calling line moved to a paid Quo number with recording off. Google’s filter labeled the new number as spam on some phones. |
| DEV-008 | 2026-09-24 | Coding rule added before any call: an automated call screener is not an answer. |
| DEV-009 | 2026-09-24 | Two pairs of sampled communities share one phone line. Kept as drawn, with a pre-committed sensitivity check. |
| DEV-010 | 2026-09-24 | Price covariates coded from each community’s own website and frozen before any call, with five reading rules written down. |
| DEV-011 | 2026-09-25 | Coding rule, a post-field check and two sensitivity analyses added before any call, so a blocked call could not be miscounted as a dead number. |
| DEV-011 check | 2026-09-30 | The post-field check, run from a second phone. Three lines were confirmed dead. Two returned a busy signal and were corrected to no-answers. |
| DEV-012 | 2026-09-28 | Sunday’s session was run on schedule but logged under Wednesday’s label and dial order. Relabeled in a copy; the original is kept untouched. |
| DEV-013 | 2026-09-30 | The call logger never recorded ring counts or phone menu depth, and its call length covers the whole call. No main result uses those fields. |
| DEV-014 | 2026-09-30 | Tuesday’s export was not committed to version control at the end of its session. It matches its own automatic backup and the schedule exactly. |
| DEV-015 | 2026-09-30 | When a count is zero, the lower end of its range is shown as 0.0% instead of the value the registered formula gives. |
| DEV-016 | 2026-09-30 | A busy signal on the post-field check is read as a connected line, and correction rows replace the rows they correct in the analysis input. |
| DEV-017 | 2026-09-30 | The public data file withholds three registered columns, because with them a reader could trace some communities to their own results. |
Check our work
Everything needed to rerun the draw and the analysis is here. No email, no form.
- The call data (CSV). One row per call. Rows are shuffled and community numbers are reassigned, so no row can be tied to a named community. Three registered columns are withheld for the same reason (DEV-017), so the price and prospect subgroups above cannot be rerun from this file. Every other number on this page can.
- analyze.js, the analysis script registered before the first call. Every number on this page comes from its output.
- The analysis output (JSON), exactly as analyze.js wrote it.
- 00-draw.js, which draws the sample from the state’s public roster. The registered seed is 0203102737, the five white balls of the California SuperLotto Plus draw held after the study was registered, so nobody could pick the sample in advance.
- The full deviation log.
How the study was designed, and every rule that was fixed before the first call, is on the method page and in the OSF registration. For more original data on the same industry, see California senior-care bed loss by county.
Corrections
If you find an error, write to ed@elevyr.com. Corrections are made on this page with a dated note, not quietly.
Cite this study
These results may be reproduced with attribution and a link. No permission form, no email required. No individual community is named anywhere in this study, and none will be.
Plain text
Elevyr, "The After-Hours Answer Study: Results," September 2026, https://www.elevyr.com/research/after-hours-results
HTML
<a href="https://www.elevyr.com/research/after-hours-results">Elevyr, "The After-Hours Answer Study: Results," September 2026</a>
Press and data requests go to ed@elevyr.com.
Sloane, our AI Admissions Agent for senior living, answers every call at any hour. That is the product this study could have been built to sell, which is exactly why it was built the hard way.