150 Badges, 80% CNA Turnover, a Per-Seat AI Bill: When a Nursing Home Group Should Run Its Own Model

You priced this in 2024, got a quote of thirty dollars a seat a month, multiplied it by a headcount of 1,700 across twelve buildings, and walked out of the meeting. Fair. That math never worked in this industry and the vendors never had a good answer for why it should.

Something did change in 2026, though, and it is worth ten minutes. Moonshot AI released Kimi K3, a sparse mixture-of-experts model — meaning only a small slice of it wakes up for any one question — and it is now the largest open model in the world. More to the point, the whole open tier closed most of the gap with the commercial products. For a nursing home group that means the question stopped being “can we afford the good one” and became “do we buy seats or run our own.”

The headcount shape that breaks per-seat pricing

Look at who is actually on your payroll. A 120-bed skilled nursing facility runs somewhere around 150 employees. Roughly 60 to 70 of them are CNAs. Add dietary aides, housekeeping, laundry and maintenance and you are at 100 people who touch a computer for maybe six minutes a shift — punching in on the time clock, charting ADLs at the Point of Care kiosk in the hallway, acknowledging an in-service in the training system.

Now add the turnover. CNA turnover in this industry routinely runs from 70% to well over 100% a year. Per-seat licensing assumes a stable knowledge worker who logs in daily for three years. You have a workforce that, by design of the labor market, replaces itself annually. Every seat you provision is a seat you are half-wasting, and every offboarding is a reclamation task nobody in your business office has time to do.

The people who would genuinely use an assistant all day are a short list: the Administrator, the Director of Nursing, the ADON, the MDS coordinator, the Business Office Manager, the Staffing Coordinator, Social Services, the Marketing Director, and at corporate the regional nurse consultants, the billing team and HR. Call it 18 to 25 people per building plus a corporate office. That is a real number and per-seat pricing is fine for it.

The work that is high-volume and low-drama — that's the tell

The reason per-seat pricing breaks is not the seats. It is that the highest-value AI work in a nursing home is not a person sitting at a keyboard. It is a sweep that runs at 2 a.m. against every chart in the building, every night, and never asks anyone for anything.

Concretely: unsigned physician orders past 30 days. Missed eMAR signatures from the 3-to-11. Weekly skin assessments that did not get done. Wound measurements missing from a treatment record. Care plans past their review date. Assessment reference dates coming due inside the MDS window. Incident and fall reports where the neuro checks stopped early. Residents with a Medicaid application pending past 60 days and no follow-up note. Right now that is your ADON and your unit managers with CASPER reports and canned reports on a Monday, catching maybe two-thirds of it.

flowchart TD
  A["Nightly sweep across 118 charts"] --> B{"Per-seat licence or hosted open model?"}
  B -->|Per seat| C["Pay for 150 badges, most turn over yearly"]
  B -->|Hosted open model| D["Pay for the machine hours the sweep uses"]
  C --> E["Cost tracks headcount"]
  D --> F["Cost tracks volume"]
  E --> G["Compare to the ADON hours and citations it offsets"]
  F --> G
  G --> H["Owner decides per building, not per company"]

That workload has nothing to do with seats. It is 118 charts times 365 nights, and priced per seat it is free but impossible; priced per unit of work it is cheap and obvious. Which is exactly why the buy-versus-run question changed.

Buy per seat versus run your own, with numbers filled in

Illustrative for a twelve-building group, 1,700 total employees, corporate office of 22. Cloud prices for capable commercial models fell roughly tenfold from 2025, and running high-volume work on your own hardware costs on the order of 90% less than sending the same work to the cloud.

Against that, price what it offsets. If the sweep saves each building's ADON six hours a week of chart auditing — conservative — that is 12 buildings × 6 hrs × 50 weeks × $52 loaded, or $187,200 of nursing leadership time returned to the floor. And that is before you count a single avoided citation on a plan of correction.

What “running your own” actually costs in people

This is where owners get sold a story. You do not hire a research team. But you do need someone to own it, and in senior living that person almost never exists in-house.

Realistically it looks like this. Your regional managed IT provider — the same one who handles your PointClickCare workstations, your wireless in the resident hallways and your nurse call system — hosts the model on a server in your corporate closet or in their data center under your business associate agreement. You are buying roughly a quarter to a half of one engineer's attention: patching, monitoring, restarting the thing when it hangs, and confirming nothing left your network. Budget $40,000 to $70,000 a year for that relationship on top of the hardware, and put the responsibility for the clinical accuracy of the output where it belongs — with your Director of Clinical Services, not with the IT vendor.

The genuine upside beyond price is that resident health information never leaves your building. For a group that has been through an Office for Civil Rights inquiry, that argument alone often carries the decision.

When the answer is still buy per seat

If you run one building, buy the seats. Full stop. A single 90-bed facility does not have the volume to justify a server, and it does not have the person to babysit it. The break-even I would look for is somewhere around five to eight buildings under common ownership with an actual corporate office and an actual IT contract — below that, the per-seat bill is smaller than the aggravation.

Also buy per seat for anything where the vendor is doing more than running a model: the ones wired directly into your electronic health record, your training system, or your billing clearinghouse. You are paying for the connection and the maintenance of it, not the intelligence, and rebuilding that yourself is a project no nursing home group should take on.

Where a person stays in the loop either way

Nothing here changes who signs. The nightly sweep produces a list; the ADON decides what is a real gap and what is a documentation timing artifact. No self-hosted model writes into the clinical record, transmits an assessment or touches a claim. And be blunt with yourself about the failure mode of a cheap model on high volume: it will flag things that are fine. If your ADON spends six hours a week clearing false flags, you have moved the work, not removed it. Measure that in the first month and be willing to kill it.

Frequently asked questions

Is an open model safe to use with resident health information?

Running it on hardware you control is generally the stronger privacy position, because the data never leaves your network and there is no third party to sign an agreement with. But you inherit the whole job: access control, logging, backups, patching, and proving all of it to a surveyor or an auditor. Safe is not a property of the model; it is a property of how you run it.

How much hardware are we actually talking about?

For nightly document and chart review across a dozen buildings, two well-specified servers is the shape of it — not a data center. The models that matter here run on far less iron than they did two years ago. Have your IT vendor size it against the actual nightly volume, not against a vendor's demo.

Our software vendor says their built-in AI is included. Why pay for anything else?

Use it — it is already in the workflow your nurses know, which is worth a lot. Just check what it covers. Most built-in features handle what happens inside their own product and stop there, and much of the work in a nursing home crosses systems: the clinical record, the timekeeping system, the training system, the supply portal and the email inbox. The gap between them is where the money is.

What is the smallest sensible first step?

Pick one nightly sweep — unsigned physician orders past 30 days is the usual choice because it is unambiguous and everyone agrees it matters — and run it for 60 days in one building. Count what it found, count what was noise, and let that decide whether you scale it or drop it.

Whatever you decide about hosting, the phone is a separate problem and a simpler one. CallSphere builds AI voice and chat agents that answer the main line and web chat around the clock, route family and hospital referral calls to the right person, and book tours and appointments — the kind of steady, high-volume work that never fit neatly into a per-seat licence in the first place.