By Sagar Shankaran, Founder of CallSphere
Faxed prescriptions, monitoring program reports and recorded calls cannot leave the premises. On-premises AI in 2026 reads them in your own back office.
Key takeaways
That is the sentence, almost word for word, that independent owners say when someone brings up AI reading their fax tray. The pitch made sense — sixty faxed prescriptions a day from prescriber offices, half of them handwritten or sent through a machine with a dying drum, every one typed into the dispensing system by hand. Then the attorney read the terms, worked out that every page would leave the building and sit on somebody else's servers, checked the long-term care facility's contract, and said no.
The objection was correct in 2024. It is no longer the right answer in 2026, because the thing that changed is not the software's ability to read a fax. It is where the reading happens.
On-premises AI means the model runs on a computer you own, sitting inside your building, so the pages and recordings it reads never travel to anyone else's server. That is the whole idea, and it is the reason the conversation with your attorney goes differently now.
Every independent has data that is not merely sensitive but contractually or statutorily fenced.
Prescription monitoring program queries. When a pharmacist checks the state database before dispensing a controlled substance, the report that comes back is governed by a state statute and a user agreement that almost always restricts re-disclosure. Sending those pages to a third party for processing is a conversation you do not want to have with your board.
Long-term care and facility contracts. The business associate agreement with a fifty-bed skilled nursing facility, or with a group home operator, frequently contains language about subcontractors, offshore processing, and prior written consent before any protected health information reaches a new party. Most owners have never read that clause. The facility's compliance officer has.
Hear it before you finish reading
Talk to a live CallSphere AI voice agent for healthcare in your browser — 60 seconds, no signup.
Recorded phone calls. Every call to a pharmacy line contains a patient name and usually a drug name. If you record for quality or training, you are sitting on a large pile of protected health information in audio form — and audio is precisely the thing people casually upload to whatever tool is handy.
Camera footage over the C-II safe and the will-call bins. Your controlled-substance security arrangements exist to satisfy federal requirements and your insurer. Footage of the vault area, of who opened it and when, is not something you hand to a vendor because you wanted automatic alerts.
Two things moved in the last eighteen months. Local processors matured to the point where a small machine — Qualcomm's Dragonwing-class parts are the clearest example — can run capable models on site rather than shipping everything out. And large organisations started choosing on-premises deliberately rather than as a compromise: Cisco's rollout of a personal AI agent to roughly 90,000 employees puts explicit emphasis on on-premises handling for control and data protection. When a company that size keeps sensitive work inside the building, the small operator's version of that decision stops looking eccentric.
The economics moved too. Running high-volume work on your own hardware costs roughly 90% less than sending the same volume to the cloud, and a pharmacy fax tray plus a day of phone calls is exactly the shape of "high volume, repetitive, every single day" where that gap compounds. But price is not why you do this. You do it because the answer to "where does the page go?" becomes "into the back office, and nowhere else," and that answer fits in one sentence of a facility contract.
flowchart TD
A["Fax arrives from the prescriber's office"] --> B["Page lands on the machine in the back office"]
B --> C["Reads prescriber, drug, strength, directions, quantity"]
C --> D["Matches patient against the profile in the dispensing system"]
D --> E{"Anything unclear or a controlled substance?"}
E -->|Yes| F["Held for pharmacist review with the page on screen"]
E -->|No| G["Pre-filled data entry queued for a technician to verify"]
G --> H["Nothing left the building"]
F --> H
Today: the machine spits out thirty-one pages between 8:00 and 11:00. A technician picks them up in batches, works out which patient each belongs to, types the prescriber, drug, strength, directions and quantity into the dispensing system, and puts the page in the scanning queue. She reads three of them twice because the directions are handwritten. She calls one office because the strength is ambiguous and she is not guessing on a paediatric dose.
With the box in the back office: the same thirty-one pages arrive already read. Twenty-three of them are sitting in the data entry queue pre-filled, with the original image on the right-hand side of the screen so the technician verifies rather than types — the difference between forty seconds and four minutes per script. Six are flagged for the pharmacist because the directions are genuinely unclear or the strength does not match anything the patient has had before. Two are controlled substances and go straight to a pharmacist by rule, no exceptions. The pages never left the room; the same machine also transcribes the day's recorded calls overnight so the pharmacist-in-charge can search "who called about the amoxicillin shortage" without listening to four hours of audio.
Assumptions, illustrative: sixty inbound faxes a day averaging 1.4 pages, 140 phone calls a day averaging 2.1 minutes, 305 working days a year, cloud processing at commodity 2026 rates.
| Line | Working | Amount |
|---|---|---|
| Faxed pages read per year | 60 × 1.4 × 305 | 25,620 pages |
| Recorded call minutes per year | 140 × 2.1 × 305 | 89,670 minutes |
| Cloud processing, per year | illustrative commodity rates | ~$3,100 |
| On-premises machine, one time | small local-processing box | $2,400 |
| Power, backup and maintenance | per year | $180 |
| Year one on premises | $2,400 + $180 | $2,580 |
| Year two onwards | $180 |
The money is real but it is second. The first-order effect is that the facility contract question, the monitoring-program re-disclosure question and the recorded-call question all get the same short answer, and the project survives its meeting with your attorney. The labour figure is bigger than either column anyway: at roughly three minutes saved per faxed prescription across 18,300 faxes a year, you are talking about hundreds of technician hours that go back to the fill counter.
Still reading? Stop comparing — try CallSphere live.
See the healthcare AI agent handle a real call — complete, industry-specific, and live in your browser. No signup.
Be clear about the boundary or you will oversell it to yourself. Claims still go out to the switch and the plan — that is how you get paid. Electronic prescriptions still arrive through the national network. Your dispensing software vendor may already hold your data in their cloud, and putting a local box in the back office does not undo that. The honest claim is narrower and still worth a lot: the new processing you are adding does not create a new party holding your patients' information.
On the clinical side, nothing about local processing changes who is allowed to make a judgement. A machine reading a fax produces a proposal, not a prescription. Verification stays a technician's and then a pharmacist's job under your state's rules, controlled substances go to a pharmacist by rule and not by confidence level, and any ambiguity in a paediatric dose is a phone call to the prescriber's office exactly as it is today. The machine's contribution is that the technician is checking a filled-in screen against an image instead of typing from scratch — which is where the errors were coming from in the first place.
And there is a genuine downside to owning the hardware: it is yours when it breaks. You need a plan for the Wednesday the box is dead, and that plan is "we type them in like we did in 2024" — which is fine, as long as everyone knows it before it happens rather than during.
Not automatically — a neglected computer in an office nobody locks is worse than a well-run cloud. What changes is the legal and contractual picture: no new party is holding the data, no new business associate agreement is needed for that processing, and the answer to the facility compliance officer's question is short. You still owe the same physical and access controls you owe on your dispensing terminals.
Faxes queue and your staff work them by hand, exactly as they do now. Design it so the local machine is an accelerator rather than a dependency: if a failure stops you dispensing, you have built it wrong. Put it on the same backup power as your dispensing terminals and keep the paper workflow documented for new technicians.
For the local processing itself, generally not, because no outside party is receiving the information — but you almost certainly still need them with your software vendor, your delivery service, your billing help and anyone who touches the data remotely. Have your attorney confirm the specific arrangement rather than taking a vendor's word for it, and keep the executed agreements where an auditor can find them.
Better than the tools of two years ago, and far better on the printed prescriptions that make up most of the tray. Handwriting on a poor fax is still where it will fail, and the correct behaviour is that it says so and hands the page to a human rather than guessing. Set the threshold conservatively for the first month and watch what it flags.
Do not start with a purchase order. Start by counting: for one week, tally how many faxed prescriptions arrive, how many minutes your technicians spend on data entry, and how many of those pages are handwritten. Then read the subcontractor clause in your largest facility contract. Those two pieces of paper tell you both whether the saving is worth it and whether the local version is the only version you are allowed to buy — and most owners find the answer to the second question is yes.
The same privacy logic applies to the phone. CallSphere builds voice and chat agents that answer the pharmacy line, take refill requests, tell a caller whether an order is ready and hand anything clinical to a pharmacist — and the questions to ask any voice vendor are the same ones you would ask about the fax tray: who hears the call, where the recording sits, and what you can produce if a facility's compliance officer asks.

Written by
Sagar Shankaran· Founder, CallSphere
LinkedInSagar Shankaran is the founder of CallSphere, where he builds production AI voice and chat agents deployed across healthcare, hospitality, real estate, and home services. He writes about agentic AI, LLM engineering, and shipping voice agents that handle real calls in production.
See how AI voice agents work for your industry. Live demo available -- no signup required.
Supplement brands get adverse events reported by ticket, not by form. On-premises AI reads all 3,400 a month without that text ever leaving the building.
Concealed-damage claims cost a warehouse manager half a Friday. On-site AI searches your own footage and scan trail without any data leaving the property.
Casino surveillance footage cannot leave the property. On-premises AI in 2026 cuts a 90-minute disputed handpay review to nine, with no video sent to a vendor.
The EU AI Act's August 2 date, Texas TRAIGA and California SB 53 all landed. What a US rehab clinic must document, disclose and log, and what it can skip.
Custody orders, health plans, subsidy files and classroom video can stay in the building. What on-premises AI unlocks for a child care center in 2026.
A 250-claim PBM desk audit eats three weeks of a technician's time. Four agents splitting the pile turn it into an afternoon and a 19-claim exceptions list.
© 2026 CallSphere Inc. All rights reserved.
Made within San Francisco
Watch how CallSphere handles real customer calls, schedules appointments, and processes payments — live.
Try Live DemoBook a DemoCalculate Your ROI