AI for the practice. The patient record never leaves it.
Practices, clinics, and wellness businesses: places where the front desk carries the whole operation and the paperwork grows faster than the patient list.
Where does AI actually pay in healthcare?
Three places, in our experience. These are the clearest wins we see in practices like yours, not results we promise. None of them touch clinical judgment, and all of them start with the question every practice should ask first: where does the data go? We answer that plainly in Where Your Data Goes.
The calendar fills itself instead of the phone line.
Phone tag, no-shows, and gaps the front desk discovers too late to fill. An automation drafts confirmations and reminders, offers open slots when someone cancels, and keeps the day full. The front desk approves what goes out.
The paperwork is done before the patient sits down.
An intake agent collects forms and history ahead of the visit, chases what is missing, and hands staff a clean, chart-ready summary. Your clinicians spend the visit on the patient, and every clinical call stays theirs.
The claim is drafted from the documentation, not from memory.
Coding and claims work piles up at the end of the day and the denials pile up after that. An agent drafts the codes from the documentation and flags the uncertain ones for a person to review, so the biller stays accountable for every claim that goes out.
For most practices the right first move is the AI Readiness Audit: two to four weeks mapping where the hours actually go, and where patient data can and cannot go, before anyone builds anything. Pricing is published, in plain numbers: audits run $3,500–$8,500; sprints $18,000–$60,000, fixed quote in writing. The cloud-versus-on-prem decision, in plain English: where your AI data actually goes.
Nowhere. That is architecture, not a promise.
The question every practice asks first, answered in the only way worth answering it: by where the machine physically sits.
The models run on hardware in your building, bought in your name. A patient record read by that machine has not left your environment, because there is no outside service to leave to. This is the same architecture we run for a regulated lender whose examiners ask exactly where every borrower file went — the full build is public on the MARCUS teardown and every control is itemized on our security page.
Before any model reads a document, a Presidio-class privacy filter strips the identifiers — names, dates of birth, member and account numbers, addresses. The model works on the clinical or administrative shape of the record, not on the person. There is no route around the filter, which is why the answer is a design fact rather than a policy hope.
Everything stored is encrypted at rest. Every action the system takes writes to an append-only, tamper-evident log: what was read, what was drafted, who approved it, and when. When a compliance officer asks what happened to a record, the answer is a lookup rather than an investigation.
On the business associate agreement. HIPAA requires a signed BAA before any vendor handles PHI on a practice’s behalf — that is true of us and of every other vendor on your list. Settle it before scoping, not after. Raise it in the first thirty minutes of the free assessment; a vendor who treats it as paperwork to sort out later has told you something useful.
We hold no HIPAA certification and claim none — no such certification exists for a vendor to hold. What we can show you is the architecture above, running, with the log to prove it.
Two days a practice already knows. Drawn before and after.
Illustrative vignettes, not client claims. In both, the machine prepares and the person who knows the patient decides.
The new-patient packet is complete before anyone sits down.
Before. Monday brings eleven new patients. Each arrives with a packet filled in on a clipboard, half of it illegible, a third of it blank. The front desk re-keys it into the practice management system between phone calls, chases three of them for an insurance card, and finds at 4pm that two histories are missing the medication list the provider needed at 9am. Nobody did anything wrong. The paper simply moves slower than the schedule.
After. The packet arrives digitally before the visit. An intake agent reads it, flags what is missing against the fields your providers actually need, and sends one plain-language follow-up asking for exactly those. It reconciles the insurance details against what is already on file and drafts the summary the provider reads before walking in. The record never leaves your environment; the identifiers are stripped before any model reads a line. A person reviews every packet before it posts.
The win: the front desk stops re-keying and starts handling exceptions.The backlog gets prepared overnight, not worked down on Fridays.
Before. Encounters pile up waiting on documentation. The biller works the queue oldest-first, pulling charts to find the one line that decides the code, and the ones that need a provider question sit until Friday. Claims go out late; a share come back denied for something a second read would have caught; the appeal costs more than the visit earned.
After. A coding agent reads each encounter overnight and stages it: the supporting documentation pulled and cited, the likely code proposed with the text it rests on, and the specific question flagged where the note does not support the level. The biller opens a prepared queue instead of a pile. Nothing is submitted by the machine. A certified coder approves, edits, or sends it back — every time, on every claim.
The win: the same coder clears more, and the denials that were avoidable get caught before they are filed.Illustrative builds, not client claims. What we would actually scope is whatever your free assessment shows is worth automating first. The one build we can name is MARCUS, a private AI back office for a regulated lender — the same architecture, a different set of nouns.
Nothing clinical. Not now, not later.
The line is bright on purpose, and it is the same line we hold for a lender: the machine prepares the work, a licensed person decides it.
We do not build systems that diagnose, triage, interpret an image or a lab result, adjust a dose, or decide medical necessity. Not as a launch limitation to be relaxed later — as a design rule. In the lender build the equivalent rule is that MARCUS never decides eligibility, credit, or price; those numbers come from systems of record, never from a model’s memory. In a practice, clinical judgment is the thing that is not ours to automate.
What is left is the paperwork around the care: intake, eligibility and benefits checks, prior-authorization packets assembled from documents you already hold, recalls and reminders, referral letters drafted from the chart, coding prepared for a certified coder to approve. Administrative work, carried. Nothing sends, files, posts, or bills until a person approves it, and every one of those approvals is in the log.
AI for medical and dental practices.
Want the numbers?
The full price list is published — audits, sprints, managed services, all on one page.
Read the price list →01What can AI do for a healthcare practice?+
The administrative work around the care: intake packets completed and checked before the visit, eligibility and benefits verified, prior-authorization packets assembled from documents you already hold, recalls drafted, and coding prepared for a certified coder to approve. Never the clinical decision.
02Where does our patient data go?+
Nowhere. The models run on hardware in your building, bought in your name, so a record read by that machine has not left your environment. A Presidio-class privacy filter strips identifiers before any model reads a document, everything stored is encrypted at rest, and every action writes to an append-only, tamper-evident log.
03What about HIPAA and a business associate agreement?+
HIPAA requires a signed business associate agreement before any vendor handles PHI on your behalf — true of us and of every vendor you are considering. Settle it before scoping, not after; raise it in the first thirty minutes of the free assessment. We hold no HIPAA certification and claim none, because no such certification exists for a vendor to hold. What we can show you is the on-premise architecture and the audit log behind it.
04Will it touch clinical decisions?+
No. We do not build systems that diagnose, triage, interpret an image or a lab result, adjust a dose, or decide medical necessity. That is a design rule, not a launch limitation. Nothing sends, files, posts, or bills until a licensed person approves it.
05How much does it cost for a practice?+
The AI Readiness Audit runs $3,500–$8,500; an AI Implementation Sprint runs $18,000–$60,000, quoted fixed in writing before work begins and guaranteed live within 90 days. For a practice, the first build is usually intake or the coding backlog — whichever your assessment shows is costing more.
06How long does it take?+
Plan on 2 to 4 weeks for the audit and 4 to 12 for the first build, scheduled around clinic hours rather than against them. The guarantee holds either way: live within 90 days, or we keep building at no charge until it is.
07Do you work remotely or on-site?+
Practices near Denver, Colorado or Phoenix, Arizona can have us on-site; elsewhere in the US we work remotely. On-premise hardware ships to your building and gets installed either way, and the fixed prices do not change with the zip code.
See the wins in healthcare.
30 minutes with a senior advisor who walks your scheduling, intake, and coding workflows and tells you what is worth automating, and what is not.
A fixed price, quoted in writing before we start, and a straight answer within 24 hours.
— Christopher Myers, Founder