AI Went From $0 to a $1,900 Line Item Across Three PT Clinics. Here's How to Cap It Without Breaking January.

"We tried this in 2024 and the bill was a rounding error"

That objection was fair two years ago. One clinic, one documentation assistant, two therapists using it — the invoice was smaller than the coffee service and nobody asked a second question. Then it worked, so it spread. The phone agent went live on the second location's line. Billing started using an assistant to chase aged claims. Every therapist got the note-drafting tool because the two who had it stopped charting at 8 p.m. Then somebody built a spreadsheet and the AI line across three sites came to $1,900 a month, up from nothing, and nobody could say which part earned its keep.

That is not a horror story. It is what success looks like in outpatient rehab in 2026, and it arrives without a decision ever being made. What makes it feel alarming is that it landed in the general and administrative column without a budget line, an owner or a review.

Where the money actually goes in a rehab practice

Before you cap anything, know the shape of the spend. It is not evenly distributed, and the instinct to cut the biggest number is usually wrong.

The cost sorts into four buckets. Documentation support is largest by volume — every visit generates a note, so it tracks visit count, and 4,200 visits means 4,200 small charges. The phone and web agent follows call volume, a different curve entirely; it spikes in the first two weeks of January and after any Saturday you are closed. Billing and follow-up — appeals, claim status, denial letters — is lumpy and lands in the last week of the month. Everything else is what clinic directors and marketing do ad hoc, and it is where the runaway hides.

The single most useful move an owner can make is to stop budgeting AI as a monthly subscription and start budgeting it per patient visit, because a rehab practice's volume swings by a third between January and October.

What the July 2 governance update changed

On 2 July 2026, Claude Enterprise added the controls that make this manageable: a cost and usage dashboard, spend limits at both the organization and individual user level, automatic alerts at 75 percent and 90 percent of a cap, and the ability to set which model people get by default and what each person may use. Your bookkeeper's tools can pull the usage numbers automatically instead of someone screenshotting a dashboard monthly.

The translation for an owner: AI stopped being a mystery charge and became a line item you budget, cap and hold someone accountable for — like clearinghouse fees, modality supplies and Google Ads spend. The alerts matter more than the caps, because a hard cap tripping at 4 p.m. on the 27th turns off the phone agent during the busiest week of the month.

flowchart TD
  A["Set the cap from last quarter's visit count, not last month's bill"] --> B["Split it four ways: notes, phone agent, billing, everything else"]
  B --> C{"Alert fires at 75 percent before the 20th?"}
  C -->|No| D["Leave it alone until the quarterly review"]
  C -->|Yes| E["Check visit and call volume in the EMR first"]
  E --> F{"Is volume up too?"}
  F -->|Yes| G["Raise the cap. It is working, not leaking."]
  F -->|No| H["Find the one user or one workflow doing it"]
  H --> B
  G --> B

January will wreck a flat monthly cap

Outpatient rehab has a seasonal shape no vendor's default settings account for. The first two weeks of January are the worst combination in the calendar: deductibles reset, so every established patient calls to ask what a visit costs now, while new-year referrals arrive and half your existing plans of care quietly stall because patients are suddenly paying full freight. Call volume spikes while visit volume falls. The pattern flips in Q4, when patients who have met their deductibles come in for everything they postponed and volume runs 20 to 30 percent above the summer trough.

A flat $1,900 cap set in September trips in the second week of January on phone volume alone — exactly when answering benefit questions fast is what keeps plans of care alive. Set the cap per visit and per call, then multiply by your forecast: a January cap higher on the phone bucket and lower on documentation, and an October cap that is the reverse.

The same applies to the sports seasons that drive your referral mix. If you hold high school contracts, the fall injury wave from August through November raises evaluation volume, and an evaluation generates roughly three times the documentation and phone work of a follow-up visit. Budget the ramp before it hits.

Entitlements on a rehab org chart: who gets the expensive model

Model defaults and per-person entitlements sound like an IT concept until you map them onto the roles on a clinic org chart.

• Rehab techs and aides — fast, cheap default, access limited to scheduling, reminders and waitlist work. No tech needs the most capable model.
• Front desk and patient access — fast model by default, escalation path for benefit questions. Cap per person; volume is highest here.
• PTs and PTAs — capable model for documentation drafting, per-user cap set from visits per week. A therapist seeing 12 patients a day generates roughly twice the drafting cost of one seeing 6, and the cap should say so.
• Billing and revenue cycle — capable model, deliberately generous cap. This is the one seat where more spend returns cash; an appeal that recovers a denied block of visits pays for a year of that person's usage.
• Clinic directors and the owner — capable model, no per-task cap, monthly review. Rationing it is false economy.

A worked budget for 4,200 visits a month

Illustrative numbers for a three-site practice with 18 clinicians. Substitute your own visit count and vendor pricing; the structure is the point.

That is $0.44 per patient visit. Next to the numbers you already know — roughly $95 net collection per visit — AI is running under half a percent of revenue, a fraction of one part-time front desk person, and about what you spend on kinesiology tape and consumables. The point of the table is not the total. It is that each row has a driver, so when a row moves you can tell in ninety seconds whether volume moved with it. A note-drafting line that jumps 30 percent on flat visits is a problem. The same jump on 30 percent more visits is a good month.

What not to cap, even when the alert fires

The instinct when the 90 percent alert arrives on the 22nd is to throttle everything. Resist it in three places.

Do not cap the billing seat mid-appeal cycle. Timely-filing windows and appeal deadlines do not care about your budget month. Cutting off the person working denials in the last week to save $80 can cost a five-figure write-off.

Do not cap the phone agent in the first three weeks of January. Every unanswered benefit question in that window is a patient who quietly stops coming, and a stalled plan of care costs you six remaining visits — several hundred dollars — against a few dollars of saved usage.

Do not put a hard stop on documentation. If the drafting tool goes dark at 3 p.m. Thursday, your therapists finish notes at home, which is the exact problem you bought it to solve. Use an alert plus an approval step, not a wall.

One more thing stays human: the review itself. The right monthly ritual is the office manager putting the AI usage report next to the visit count, collections and the cancellation rate in one fifteen-minute meeting. Anyone reading the AI bill in isolation makes the wrong call every time, because the bill going up is not information by itself.

Frequently asked questions

How do I know whether the AI spend is actually earning anything?

Pick one measurable per bucket and track it monthly against the spend. For documentation: notes completed same-day as a percentage, and therapist hours logged after 6 p.m. For the phone agent: calls answered versus abandoned, and new evaluations booked outside business hours. For billing: denied dollars recovered. If a bucket cannot produce a number, it should not have a cap — it should have a decision about whether to keep it.

Should each clinic site get its own budget, or should it be pooled?

Pool the total, cap by site, and set the site caps from visit count rather than splitting evenly. Sites differ more than owners expect: a location with a heavy workers' comp mix generates far more documentation and correspondence per visit than a cash-pay wellness-heavy site. Even splits guarantee one site is throttled while another has room it does not use.

What is a sensible starting cap if I have no history at all?

Take last quarter's total visits, multiply by $0.50, and set that as the organization cap for the first ninety days, with alerts on. It will almost certainly be too high, which is what you want at the start — you are buying information about the real shape of usage, and you can tighten in month four with actual numbers instead of a guess.

My therapists will hate a per-person limit. How do I introduce it?

Set the limits well above what anyone is currently using, tell staff the number, and explain that the alerts exist so nothing gets switched off unexpectedly. The resentment comes from surprise shutoffs, not from limits. In practice most clinicians never come close to their cap, and the ones who do are usually doing something genuinely useful that you want to find out about.

Monday's version of this is one number: total AI spend for June divided by June visit count. Under a dollar a visit and you do not have a cost problem, you have a visibility problem — setting alerts at 75 and 90 percent takes ten minutes. Over two dollars a visit, find out which bucket owns it before changing anything.

CallSphere builds AI voice and chat agents that answer clinic phone lines and web chat, book appointments and capture patient inquiries 24/7. If you are budgeting this the way described above, the phone bucket is the one with the cleanest return to measure — calls answered after hours and evaluations booked outside business hours are countable, and worth asking any voice vendor to report to you monthly rather than making you go looking for it.