KOGSYEvidenceThe scan

Are there any measured results from AI in non-medical home care?

Search run 28 July 2026. Last updated 28 July 2026. We re-run this and change the date, including if the answer changes.

01 · The answer

The short answer

We looked for a single case in US non-medical home care with a named organisation, a stated measurement method, and a figure a third party could check. As of 28 July 2026 we could not find one. The closest was a named private-pay agency whose vendor reports biweekly revenue rising from about $96,000 to $130,000 within six months of deploying AI; it does not clear the bar because the published case study does not say who pulled the revenue figures, over exactly which periods, or how the comparison was made. That is a statement about what we could find, not a claim that nothing exists: absence of evidence in a search is not proof of absence. It also applies to us. We have not published a measured result either, and the assumption inside our own calculator is unvalidated.

02 · Our interest

Our interest in this question

03 · The bar

What counted as a qualifying result

The four criteria below were fixed before the search ran and were not adjusted afterwards. That order matters: a rule written after the results can always be accused of having been fitted to them. A case needed all four, and any one missing was enough to exclude it.

• 1. A named organisation. A real agency, network or operator, identified. Not "a leading provider", not "a 200-caregiver agency in Ohio".
• 2. A stated measurement method. What was measured, over what period, against what baseline, and by whom. A before-and-after with no stated baseline does not qualify. Neither does a figure footnoted "internal data".
• 3. A checkable figure. A number someone outside the vendor and the customer could in principle verify or challenge, in a public artefact. A figure in a PDF counts. A figure that exists only in a webinar recording does not.
• 4. US non-medical home care. Private-pay or Medicaid-waiver personal care, companion care, homemaker services. Caregivers, not clinicians. Shift notes, not OASIS.

Nothing about peer review, sample size or statistical significance. Any one honest, documented deployment would have cleared it.

The fourth criterion did most of the work, and it is the one that is easiest to get wrong. Medicare-certified home health is a different business: different staff, different documentation, a different regulator. So are hospice, hospital-at-home, assisted living and nursing homes. So are ambulatory clinics, where documentation-burden research is abundant and reads as though it answers this question. Where a source covered skilled home health and non-medical home care together without separating them, it failed criterion 4 rather than counting.

Also excluded regardless of the four: unattributed "up to X%" ranges, surveys of opinion or satisfaction or perceived time saved, named-customer case studies with no method, vendor ROI calculators and modelled projections, analyst estimates, partnership and pilot announcements with no result, and anything measuring adoption rather than outcome.

04 · The method

How we searched

This was one pass. It is not a systematic review and should not be described as one.

Searches ran on 28 July 2026, between roughly 08:45 and 09:35 Central European Summer Time (UTC+02:00); a timestamp captured mid-search reads 09:27:51 CEST. The start time is approximate because no automated timestamp was taken before the first query. That is a defect in the method and we are recording it rather than presenting a tidier version.

No publication-year cutoff was applied. Material from database inception to 28 July 2026 was eligible and nothing was excluded for age alone, though in practice the results clustered between 2021 and 2026.

What was searched

• A general web search index, covering indexed pages, PDFs, press releases, vendor customer stories, trade press and scholarly results.
• PubMed, through the NCBI E-utilities API, with one primary query and four journal-scoped queries.
• ClinicalTrials.gov, through its public API.
• Vendor and organisation sites directly: SimiTree, ybot, Zingage, WellSky, HHAeXchange, AlayaCare, Axxess, AxisCare, Alora, CareSmartz360, Smartcare Software, CareVoyant, Sandata, CareAcademy, Sensi.AI, Roger Healthcare, Nightingales Care at Home, BCG, and the public file host of the MIT NANDA report.
• The Wayback Machine CDX endpoint, attempted for one missing PDF.

Not searched: Google and Bing directly, Scopus, Web of Science, CINAHL, ProQuest, and paywalled analyst databases. Some publisher content may have surfaced through the web index, but that is not the same as searching those databases and we are not going to describe it as though it were.

Screening depth

• 76 web queries, each screened to a cap of the first 10 ranked results. Where fewer than 10 were exposed, all were read. The tool does not report a stable per-query result count when several queries run together, so the defensible statement is "up to 10 examined per query", not a total-results census.
• The primary PubMed query returned 39 records. All 39 titles were screened.
• Journal-scoped PubMed queries returned 9 records in JAMDA, 2 in Home Health Care Services Quarterly, 3 in the Journal of Applied Gerontology and 9 across JAMIA and JAMIA Open. All titles screened; some records overlapped.
• ClinicalTrials.gov returned 14 records. All 14 titles, statuses and locations were screened. The only US records concerned cancer coaching and augmentative communication, neither of which is a deployment in non-medical home care.

The database queries, verbatim

PubMed, primary:

(("artificial intelligence"[Title/Abstract] OR "machine learning"[Title/Abstract]
  OR "natural language processing"[Title/Abstract] OR "voice documentation"[Title/Abstract]
  OR "ambient documentation"[Title/Abstract])
 AND ("home care aides"[Title/Abstract] OR "home health aides"[Title/Abstract]
  OR "personal care aides"[Title/Abstract] OR "non-medical home care"[Title/Abstract]
  OR "home care services"[MeSH Terms])
 AND (documentation[Title/Abstract] OR notes[Title/Abstract]
  OR workflow[Title/Abstract] OR deployment[Title/Abstract]))

PubMed, journal-scoped. The same inner clause was run against four journal filters: "J Am Med Dir Assoc"[Journal], "Home Health Care Serv Q"[Journal], "J Appl Gerontol"[Journal], and "J Am Med Inform Assoc"[Journal] OR "JAMIA Open"[Journal]:

AND (("artificial intelligence"[Title/Abstract] OR "machine learning"[Title/Abstract]
  OR "natural language processing"[Title/Abstract] OR "voice"[Title/Abstract]
  OR "ambient"[Title/Abstract])
 AND ("home care"[Title/Abstract] OR "home health"[Title/Abstract]
  OR "home care services"[MeSH Terms]))

ClinicalTrials.gov:

("artificial intelligence" OR "voice documentation" OR "ambient documentation")
 AND ("home care" OR "home care aide" OR "personal care aide")

All 76 web queries, verbatim

Published so the search can be repeated and disagreed with. Each was screened to the first 10 ranked results.

Show the full query list

SimiTree, ybot and Pinnacle Home Care (1–12)

1. SimiTree ybot Pinnacle Home Care 87.63 minutes per patient white paper PDF
2. site:simione.com OR site:simitreehc.com ybot Pinnacle "87.63"
3. site:ybot.ai Pinnacle Home Care documentation white paper
4. "Pinnacle Home Care" "87.63 minutes"
5. "The AI Advantage: Empowering Home Healthcare Teams" "Patient A"
6. "Patient B" ybot "CMS Suggested Timing"
7. "81.00" "56.40" "48.60" ybot
8. "Pinnacle Home Care" ybot methodology sample data
9. webcache "2024-05__SIM_Pinnacle-HC_AI-Advantage_WP_IA-5-2.pdf"
10. "SIM_Pinnacle-HC_AI-Advantage_WP" mirror
11. web.archive.org simitreehc.com Pinnacle-HC AI Advantage PDF
12. site:ybot.ai "Pinnacle Home Care" white paper

Roger Healthcare (13–20)

13. "Roger Healthcare" AI home care nine agencies case study
14. "Roger" healthcare AI documentation home care agencies
15. site:rogerhealth.com home care agency AI documentation results
16. site:rogerhealthcare.com home care AI agency
17. Roger Healthcare "nine" agencies AI charting
18. Roger Healthcare customers Palmeira Central Coast Capitol New Day nine agencies
19. site:rogerhealthcare.com/customer-story
20. site:rogerhealthcare.com "customer story" "documentation time"

Major home care platforms (21–36)

21. site:wellsky.com "Personal Care" AI documentation case study time saved home care
22. site:hhaexchange.com AI documentation case study home care time saved
23. site:alayacare.com AI documentation case study home care "time"
24. site:axxess.com AI documentation case study home care "time saved"
25. "Nightingales Care at Home" John Huston location
26. site:nightingaleshomecare.com John Huston owner
27. "WellSky Summarize" "Nightingales Care at Home" method
28. "WellSky Summarize" "7 hours per month"
29. site:axiscare.com AI documentation home care case study time saved
30. site:alorahealth.com AI documentation home care case study time saved
31. site:smartcaresoftware.com AI documentation home care case study time saved
32. site:caresmartz360.com AI documentation home care case study time saved
33. site:sandata.com AI home care case study measured result
34. site:careacademy.com AI home care case study measured result
35. site:smartcaresoftware.com artificial intelligence home care results customer
36. site:carevoyant.com AI home care case study documentation

Broader non-medical home care and additional vendors (37–56)

37. AI non-medical home care agency case study measured results caregiver named agency United States
38. AI home care agency case study caregiver documentation "time saved" non-medical
39. Sensi.AI case study home care agency measured results named customer
40. Homewatch CareGivers AI case study falls result measured
41. Zingage "82% reduction" "after-hours labor costs" agency
42. site:zingage.com case study Medicaid agency 82% after-hours labor
43. Zingage home care customer case study named agency results
44. "shift coverage maintained at 98%" Zingage
45. site:zingage.com/case-studies "How Sunny Days Eliminated After-hours Burnout and Cut Costs 82%"
46. site:zingage.com "Sunny Days" "82%" "John Bennett"
47. site:zingage.com/case-studies/sunny-days
48. Zingage Sunny Days case study baseline after-hours cost period
49. "How Senior Helpers Cameron Park Grew Revenue 35%"
50. "Desiree Trunzo" Zingage "35%" revenue
51. "$96K" "$130K" "Senior Helpers" Zingage
52. site:zingage.com/case-studies "Cameron Park"
53. site:sensi.ai "Visiting Angels" "88%" "85%" "50%" period
54. site:sensi.ai "Tim Welbaum" Sensi client base billable hours revenue growth
55. site:sensi.ai case study "reduced hospitalizations" method home care agency
56. site:sensi.ai case study "22%" hospitalizations named agency

Scholarly literature and trials (57–64)

57. site:pubmed.ncbi.nlm.nih.gov ("artificial intelligence" OR "ambient documentation" OR "voice documentation") ("home care" OR "personal care") caregiver aide documentation
58. site:pubmed.ncbi.nlm.nih.gov AI "home care aides" documentation United States
59. site:clinicaltrials.gov AI voice documentation home care caregiver aide
60. site:clinicaltrials.gov artificial intelligence non-medical home care personal care aide
61. site:jamda.com OR site:pubmed.ncbi.nlm.nih.gov JAMDA artificial intelligence "home care" aide documentation
62. site:tandfonline.com "Home Health Care Services Quarterly" artificial intelligence documentation home care aide
63. site:journals.sagepub.com "Journal of Applied Gerontology" artificial intelligence home care aide documentation
64. site:academic.oup.com/jamia artificial intelligence home care aide documentation

Enterprise and sector survey evidence (65–76)

65. MIT NANDA August 2025 The GenAI Divide 95% organizations no return interview survey counts primary report PDF
66. site:mit.edu "The GenAI Divide" 95% 150 interviews 350 employees 2025
67. BCG September 2025 1,250 executives 60% laggards minimal revenue cost gains AI primary source
68. site:bcg.com September 2025 "1,250" "minimal revenue" "cost gains" AI laggards
69. "The GenAI Divide" filetype:pdf NANDA Challapally Pease Raskar Chari
70. "State of AI in Business 2025" "52 organizations" "153 senior leaders" PDF
71. site:nanda.mit.edu "GenAI Divide"
72. site:mit.edu "State of AI in Business 2025" NANDA PDF
73. site:axiscare.com/wp-content/uploads/2026/01/wp_rd6.pdf "400" "AI" home care providers survey methodology
74. site:axiscare.com "State of AI for Home Care Agencies" 400 survey methodology
75. "48%" "AI has simplified administrative tasks" home care AxisCare
76. "State of AI for Home Care Agencies" AxisCare PDF

What we could not reach

This belongs in the method, not in a footnote, because it is where a qualifying result is most likely to be hiding.

• The SimiTree white paper PDF. Requested directly on 28 July 2026 and returned 404; opening the indexed copy returned 502. The search index still exposed substantial extracted text including the methodology and the tables. The current gated landing page was reachable.
• The Wayback Machine. CDX requests for both the simitreehc.com and www.simitreehc.com forms of that PDF timed out after 30 seconds. No archive copy was retrieved.
• Google cache and mirrors. No separate downloadable copy surfaced. Google Cache is no longer available as a product.
• Home Health Care News. The relevant article returned 403 when opened, though the index exposed the quotation and figures.
• The MIT NANDA report's cited host. The commonly cited mlq.ai URL returned no content; the report itself was readable from other public hosts.
• Everything private. Internal dashboards, accounting records, customer data exports, unpublished pilots, results held under commercial confidence, private conference decks, and talks with no public artefact. If a measured result exists inside one of those, this search could not have found it. That is a real limitation, not a rhetorical one.

Why it stopped

The pass ended because every named workstream had been searched, every listed platform queried, the literature and trial registries screened, and repeated queries were returning the same candidates failing on the same criterion. That is saturation within a one-pass cap, which is a much weaker claim than an exhaustive census, and it is the only one the work supports.

05 · The near misses

What came closest, and what each one was missing

This is the most useful section on the page. A conclusion is easy to assert and hard to trust; a list of what nearly counted can be checked case by case.

Read the next paragraph before the list. Every figure below is the organisation's or vendor's own published claim, quoted as published and linked to source. We have not measured any of these products and we are not suggesting any of these figures is false. The absence of a published method is not evidence that a number is wrong. It means an outsider cannot check it, which is a different and much more boring problem.

• 1. Senior Helpers Cameron Park, with Zingage. The vendor's case study reports biweekly revenue rising from about $96,000 to $130,000 within six months of implementation, a 35% increase, plus one scheduling position removed described as a 13% cost reduction. Named agency, exact figures, in scope, and an approximate baseline and period. Missing: criterion 2. The page does not identify who extracted the data, the exact comparison dates, what counts as revenue, or what else changed in those six months — acquisitions, price changes, a sales push, new territory. A second write-up repeats the same figures but attributes them to the vendor rather than measuring independently.
• 2. Nightingales Care at Home, with WellSky Summarize. WellSky's flyer quotes the owner saying pre-visit documentation went from 30 minutes or more to at most 15, saving more than seven hours a month. The agency is a named Colorado provider of personal care, homemaking and companionship, so it is squarely in scope. Missing: criterion 2. No sample, observation period, task count, timer source or outside analyst. It is also self-reported time rather than logged time, which the rule excluded from the start.
• 3. Home Matters Caregiving and a PACE programme, with Sensi.AI. The vendor's pilot write-up describes a year, 33 seniors, an 18% hospitalisation rate during the programme, and PACE estimates of $350,000 in avoided hospital stays and $100,000 in avoided emergency visits. Missing: criteria 1 and 2. The agency is named but the PACE organisation used for the comparison is not, and there is no baseline rate, comparator, denominator, attribution method or analyst. The savings are explicitly estimates, which the rule excluded.
• 4. Sunny Days In-Home Care, with Zingage. The case study reports an 82% reduction in after-hours cost with 98% shift coverage maintained, against a stated baseline of five schedulers at $20 an hour covering 127 after-hours hours a week, roughly $300,000 a year. That is more baseline than most. Missing: criterion 2. No post-deployment cost denominator, no observation period for the 82%, no analyst. The service also combines human agents with software, so the result cannot be attributed to automation alone from this artefact.
• 5. Visiting Angels South Central Michigan, with Sensi.AI. A testimonial reports clients rising from 75 to 141, weekly billable hours from 1,189 to 2,201, gross revenue up nearly 50%, and inquiry-to-assessment conversion from 33% to 85%. Missing: criterion 2. No measurement period, source systems, analyst, comparison protocol, or account of other growth activity in the same window.
• 6. Pinnacle Home Care, ybot and SimiTree. This was the lead we most expected to qualify, and it is now closed. The white paper reports 87.63 minutes and $73.03 saved per patient across start of care, recertification and discharge, and its method text does refer to timesheets, comparison against CMS expectations, employee surveys and random sampling. Failed on criterion 4, decisively. The paper is about OASIS, start of care, recertification, discharge, RNs and clinicians. Pinnacle describes itself as a skilled provider of nursing and therapy, and CMS lists Pinnacle entities on its home health agency list. It is skilled home health, not non-medical home care. Only a separately reported personal-care or companion-care cohort would change that.
• 7. Roger Healthcare's named customers. The vendor's customer stories name several providers and report documentation-time reductions of 75% to 90%, OASIS completion under 15 minutes, and revenue growth. Failed on criteria 2 and 4. No reproducible method is stated, and the product and the cases concern OASIS, nurses and therapists — Medicare-style home health documentation again.
• 8. AlayaCare's workflow figures. The announcement cites up to 50% of care-plan documentation time, up to 80% of manual scheduling cost, more than 80% of visit-verification errors resolved, and vacant visits down 42%. Failed on criteria 1 and 2. No customer is attached to the figures and no method is published. "Up to" ranges were excluded from the start.

The rest of the platforms, and how each was dispositioned

This table records what this search found and how the rule applied to it. It is not a claim that any of these companies has never measured anything.

The peer-reviewed literature

The primary PubMed search returned 39 records and the dominant pattern was model development or retrospective prediction using skilled home health nursing notes, not an evaluated deployment in non-medical home care. Representative examples: a 2024 JAMDA paper using natural language processing to flag skilled home health patients at risk of incapacity without directives; a 2025 JAMIA Open study comparing machine classification against expert nurses on de-identified home health notes; a 2022 JAMIA Open pilot comparing recorded patient-nurse discussion against the record and finding information gaps; a 2025 CHI paper on equity and governance in home care AI that reports no agency outcome; and a Home Health Care Services Quarterly paper evaluating documentation software qualitatively in Austrian 24-hour care, which is non-US and qualitative.

None of them is a qualifying result, and none of them claims to be. Searching CINAHL, Scopus and Web of Science with the same protocol is the obvious next step and has not been done.

06 · The findings

What we found instead

The public evidence fell into five shapes.

• Named customer, exact figure, thin method. More common than the sector gets credit for. The customer is named and the number is specific; what is absent is the analyst, the data source, the exact comparison dates, the sampling rule, and what else changed at the same time.
• "Up to" ranges with no customer. A percentage attached to no named organisation and no method.
• Self-reported time. Someone describing how long documentation took before and after. That is self-reported active effort, not logged labour time.
• Completion latency presented as time saved. How quickly a note reaches the system after a visit. That is not the worker's active effort and cannot be turned into a labour cost without measuring labour separately. This distinction is the reason we changed our own sprint definition.
• Measured evidence in an adjacent setting. The strongest methods were in Medicare-certified home health, nursing, OASIS documentation, clinics and residential care. Useful for product design; it does not answer this question.

So the practical position for a buyer is not that vendor numbers are lies. It is that a real twenty per cent and a marketing twenty per cent are indistinguishable from outside, which is an argument for measuring your own operation rather than for picking a side.

07 · Survey evidence

The survey evidence that does exist

Two studies are worth knowing about. Both are survey evidence about enterprise AI in general, not measurements of home care, and neither should be quoted as settled fact.

• MIT NANDA, The GenAI Divide: State of AI in Business 2025. Dated July 2025 and widely discussed from August 2025. It reports that roughly 95% of organisations were seeing no measurable return from generative AI pilots, with about 5% of custom enterprise tools reaching production with measurable outcomes. Its own methods page states structured interviews with representatives of 52 organisations, survey responses from 153 senior leaders across four industry conferences, a review of more than 300 publicly disclosed AI initiatives, and a research period of January to June 2025.
• BCG, The Widening AI Value Gap, 30 September 2025. A survey of 1,250 senior executives and AI decision-makers across nine industries and more than 25 sectors, assessing maturity across 41 capabilities. BCG classed 60% as laggards reporting minimal revenue and cost gains and lacking the capability to scale.

Two corrections we have had to make to our own earlier reading of these, both in the direction of weaker claims:

• The NANDA sample. Secondary coverage widely reports 150 interviews and 350 employees. Those are not the numbers on the report's own methodology page, which says 52 organisations and 153 senior leaders. We had repeated the secondary figures and they were wrong.
• The BCG wording. The 60% figure is "reporting minimal revenue and cost gains", not "no material value". We have seen it repeated as the stronger claim, we had drafted it that way ourselves, and it is worth not doing.

There is also a sector-specific survey: AxisCare reports surveying more than 400 home care leaders, with 94% saying AI had benefited their agency and 48% reporting administrative use. That measures adoption and opinion, not outcome, so it does not qualify here — but it is the closest home-care-specific survey this pass found.

08 · Our own position

Our own position

We are in the same position as everyone else on this page, and it would be dishonest to write it without saying so.

We have not published a measured result from our own work. The calculator on this site models a reduction in documentation time, and the high end of that model assumes voice capture removes 80% of documentation time. That number is our working assumption. It is not measured, by us or by anyone whose work we could find, and the calculator states this on screen and lets you move it. That is the whole reason the sprint exists: to replace the assumption with a figure, for one agency at a time.

09 · Falsifiability

What would change this page

This page is meant to be falsifiable, and we would rather be corrected than right.

The most useful next step is not another broad search. It is three specific evidence requests, and any one of them arriving would change what this page says:

• Zingage and Senior Helpers Cameron Park: the six-month revenue comparison protocol, the exact periods, and the source records.
• WellSky and Nightingales Care at Home: timestamped pre-visit review logs from before and after deployment.
• Sensi.AI, Home Matters Caregiving and its PACE partner: the pilot protocol, baseline rate, event counts and the cost model.

More generally, any of these would do it: a named home care organisation publishing a documentation-time or billing-accuracy figure with the method attached; any vendor publishing a method alongside a figure; a peer-reviewed study of a deployment in this setting rather than an adjacent one; a trade association or academic group running a multi-agency measurement; or our own sprints producing enough measured baselines to report in aggregate, with the sample size and method stated.

One open item is now closed, and we are recording it because the site previously listed it as the strongest outstanding lead. The SimiTree white paper on Pinnacle Home Care was pursued directly, through the vendor's site, through mirrors and through the Wayback Machine, most recently on 28 July 2026. The PDF still returns 404, but enough of its text was recoverable from the search index to settle the question: it is skilled home health documentation, so it fails the scope criterion regardless of how good its method is. It is no longer an open lead.

If you have something that clears the bar, send it to henrik@gokogsy.com and we will add it and say where it came from.

Work out your own figure