How a New York medical practice can summarize referral paperwork with AI without breaking patient privacy

The morning starts the same way at a small clinic. A thick pile of faxed referrals and outside records waits to be read, condensed and routed, and the office manager is the only one who can do it. An AI summary workflow, with a redaction rule the front desk actually follows, can shrink that pile without exposing patient details.

The idea is simple. A general AI assistant such as ChatGPT or Claude can condense a long referral into a few usable lines: the reason for the referral, the urgent items and what has to happen next. What it cannot do is forget what you paste into it. So the practice is not about the tool. It is about the rule that decides what never leaves the page.

The rule itself is short enough to say in one sentence. Before any referral text goes into a chatbot, remove the names and identifiers, and paste only the clinical content that the summary actually needs. Everything that identifies the patient stays in the practice's own system, where it already lives. If a summary comes back, it is matched to the right record by a person, inside the practice, not by the AI tool.

What the workflow looks like

The workflow runs the same way on every referral. Someone opens the faxed or scanned document, copies the parts that matter, strips the identifying details, pastes the clean text into the chat assistant, and asks for a short summary in a set format. The summary then goes into the practice's own chart or routing system, where the patient details already live. Nothing about the patient needs to travel through the AI tool at all.

The set format matters more than the tool. A prompt that always asks for the same three things, the reason for the referral, the findings that matter, and the next steps, produces summaries a front desk can skim and act on. When the format changes from person to person, the reader has to relearn each summary, and the time saving disappears.

The check step is not decoration. AI assistants can misread a scanned page or invent a detail that was never there, and a summary that is wrong is worse than no summary. Whoever runs the workflow reads the summary against the original before it goes anywhere. That check takes seconds and it is the difference between a routine you can trust and one you cannot.

Who this is for, and who it is not for

This suits a small practice where one or two people absorb the reading load from outside records, and where the summaries feed a routing or scheduling decision. It does not suit a practice that wants the AI to store patient charts, match records to patients automatically, or make clinical calls. Those jobs need systems built for them, and this workflow is not that.

The rule also depends on every person following it, not one. A single staff member pasting a full referral with the name attached breaks the whole exercise, which is why the rule has to be trained and repeated, not posted once on a corkboard.

Getting the whole front desk on the same rule

Team habits are the hard part, and this is where hands-on training earns its place. A session built on the team's own approved tasks, with everyone practicing during the session rather than watching slides, is what turns a rule into a routine. People leave with take-home routines, safe-use ground rules and a follow-up check-in, so the redaction habit survives past the first week.

New York AI Lab runs AI Training for Business this way. Sessions run 90 minutes and up on Zoom, for one owner, a small group or the whole front desk, and the agenda is built on two or three of the team's approved real tasks rather than generic examples. For a clinic, those tasks could be the referral summary itself, drafting replies to common requests, and the redaction rule that governs both. The first step for anything with New York AI Lab is a free 25-minute fit call, which looks at two tasks that repeat in the week and picks a first result worth having. Most people need one AI subscription such as ChatGPT or Claude, paid directly, and any cost is named before anything is bought.

A practice can also build the rule in-house without any outside help. Write the redaction rule down, write the summary prompt once, and run ten referrals through the workflow as a test. If the summaries hold up and the rule holds, the practice has the routine. If the team drifts, that is when training is worth the time.

Questions before you try it