--- title: "Paperwork that never leaves the building | T10" url: "https://t10.net/blog/private-ai-practice-paperwork/" description: "Intake summaries, prior auth letters and referral notes are exactly what AI is good at. For an Upper Valley practice, the question is where that AI runs." lastmod: "2026-09-23" --- ![](https://t10.net/media/blog/2026-private-ai-practice-paperwork-hero.webp) [Back to Insights](https://t10.net/blog/) # Paperwork that never leaves the building T10 Consulting January 27, 2026 5 min read It's twenty to seven on a Tuesday in late January, the parking lot is dark, and you're still at the desk. The last patient left an hour ago. What's left is two prior-authorization letters, a referral summary for a specialist in Lebanon, and a stack of new-patient intake forms that somebody has to read before Thursday. **None of that work needs your clinical judgment until the very end. All of it currently needs your evening.** ## The paperwork problem, plainly Picture a small independent practice in the Upper Valley: a family medicine office in Enfield, a two-chair dental practice in Hanover, a therapy group in West Lebanon. Different specialties, the same pattern. The clinical work happens during the day. The writing about the clinical work happens after it. Most of that writing is assembly, not thought: - **Intake forms.** Pages of history, medications and reasons for the visit, handwritten or typed, that need to become a short summary the clinician can read in a minute. - **Prior-authorization letters.** The same argument, reshaped for each payer: diagnosis, what's been tried, why this treatment now. - **Referral summaries.** The relevant history pulled out of the chart and written up for another practice. If that assembly takes eight hours a week across the practice, that's a working day spent retyping what's already in front of you. ## Why "just use ChatGPT" is the wrong answer here A language model is genuinely good at this. Give it a completed intake form and it will produce a tidy summary. Give it the chart notes and the payer's criteria and it will draft a reasonable prior-auth letter. The problem is where the records go when you do that. **Pasting patient information into a consumer AI tool sends it to someone else's servers**, under terms you didn't negotiate and probably haven't read. Some cloud AI vendors will sign a business associate agreement for their business products, and for some practices that's the right route. But it's a legal and compliance decision, it depends on the specific product and contract, and we'd want your compliance adviser, not us, to make that call. We're not lawyers and this isn't legal advice. What we can offer is a different option that takes one question off the table. ## The alternative: the AI runs in your building A language model doesn't have to live in a data center. Capable open models now run on a single machine you can put on a shelf in your own office, or in a rack you control. The practice's documents go to that machine, the draft comes back, and **nothing is sent to an outside AI service at all.** That's what we mean by [private AI](https://t10.net/services/private-gpu-hosting/). In practice it looks like this: 1. A new-patient form is scanned or arrives through your portal, as it does now. 2. The local assistant reads it and drafts a summary in your template. 3. The clinician reads the draft, corrects it and signs it. 4. The signed version goes into the chart, the same place it always has. ![A four-step flow inside the practice's building, from a form arriving to a local AI draft, clinician review and a signed letter, with no outside AI service involved.](https://t10.net/media/blog/2026-private-ai-practice-paperwork-infographic.webp) Running it in-house doesn't make a practice compliant by itself. You still need access controls, audit logging, backups, encryption and a written policy, and your existing HIPAA obligations don't change. **But you can walk over and point at where the records went, and that's a much easier conversation to have with your compliance adviser.** ## What the assistant is good at Here's where we'd expect a private assistant to earn its place, and where it wouldn't: Job What the assistant does What stays with a person Intake summary Reads the form, pulls out meds, allergies, history and chief complaint Anything clinical: what it means, what to do Prior-auth letter Drafts from chart notes and the payer's stated criteria Reading it, correcting it, signing it Referral summary Pulls the relevant history into a one-page letter Deciding what the specialist needs to know Missing information Flags blank fields and inconsistencies Calling the patient to ask The pattern is the same all the way down. **The assistant drafts; a clinician decides and signs.** It never sends anything to a payer, a patient or another practice on its own. ## Where a person must stay in charge A few lines we'd draw before building anything: ### Every clinical statement is reviewed A model can misread handwriting, merge two medications or state something confidently that isn't in the chart. The draft is a starting point. **Nothing leaves the practice without a clinician reading it in full and putting their name on it.** ### The assistant doesn't diagnose or recommend It summarises what's written. It doesn't suggest a diagnosis, a dose or a treatment. We'd set it up to decline those questions rather than guess. ### Therapy notes need extra care Psychotherapy notes carry their own sensitivities, and many therapists will want them excluded entirely. That's a reasonable starting position. You can begin with intake paperwork and scheduling letters and never point the assistant at session notes at all. ### Access is narrow on purpose The assistant sees only the documents for the job in front of it. It doesn't get the whole EHR because it would be convenient. We'd agree what it can read, who can use it, and where its logs go, and write that down before it goes live. ## What your staff need, besides the machine The hardware is the easy part. The harder part is habits: knowing what to paste where, how to check a draft quickly, and when to throw one away and start again. **A private assistant nobody trusts is just a box in a closet.** A short, hands-on [training session](https://t10.net/services/ai-cowork-training/) built around your own forms and letters usually does more for adoption than any feature. We'd also map the paperwork before touching it. On a discovery call we walk through each document from arrival to signature, mark which steps a person must own, and write one plain sentence for every approval. That map, done in [Workmap](https://t10.net/tools/workmap/), is what the build follows. ## A narrow first job to try For an independent practice, we'd start with **new-patient intake summaries**, and only those. They're high-volume, they follow a template you already use, and the clinician reads every one anyway before the visit. Run the assistant alongside your current process for a few weeks. Compare its summaries with what your staff would have written. Keep a note of every correction. If the drafts are consistently right, add prior-auth letters for your two or three most common requests. Referral summaries come after that. Therapy notes, if ever, come last. If your evenings look like the one at the top of this post, [book a planning call](https://t10.net/#book) and we'll talk through what could stay in your building. [Back to Insights](https://t10.net/blog/) ## About this post Author T10 Consulting Published Jan 27, 2026 Reading time 5 min ## Tags - new hampshire - healthcare - private ai - medical practices - upper valley ## Related posts - [AI in the Small Law Office, Done Carefully](https://t10.net/blog/ai-small-law-office-new-hampshire/)Jul 21, 2026 - [Christmas Tree Farms: Answering the Same Questions in Five Weekends](https://t10.net/blog/christmas-tree-farm-weekend-questions-ai-nh/)Oct 14, 2026 - [Quality Paperwork at a Small NH Machine Shop, With AI](https://t10.net/blog/machine-shop-quality-paperwork-ai-nh/)Oct 13, 2026 ## Want to see where AI fits in your week? A free 30-minute planning call. Tell us how the week runs, and we'll say plainly which parts AI is good at and which it isn't. [Book a planning call](https://t10.net/#book)[More insights](https://t10.net/blog/)