Templates are the product, not a later nicety
An AI scribe without your template is a stranger summarizing a visit. An AI scribe with your template is a first-pass writer filling the headings you already sign. That difference is the whole implementation. Groups that wait for a perfect interface before they send sample notes spend six weeks in meetings and still go live on a generic SOAP. Groups that export three good notes per clinician on day one start getting recognizable drafts by the end of the week. The EHR project can proceed in parallel. Do not hold speech hostage to an interface ticket.
Collect what you already like, not a redesigned note you have never used. Preferred progress-note structure, procedure blurbs, normal exam language you actually perform, and the phrases you refuse to see. Also collect one example of a note you hate so the vendor knows the negative space. Hand those to the people configuring your AI medical scribe with the clinician's name on each file. A department average template makes everyone a little unhappy. Per-provider templates are the point. If a vendor says they cannot do per-provider, you are buying a style, not a tool.
What you actually need to collect
You need fewer artifacts than a project manager will request. For each clinician: three de-identified sample notes, the official template if it lives in the EHR, a list of macros they still want expanded, and any specialty glossary. For the clinic: document types in scope, visit types to exclude at first, and the encounter-matching rule. That packet fits in a shared folder. It does not require a steering committee. If your samples still contain PHI, send them only after the BAA is signed and only through the vendor's designated path.
Do not wait for perfect de-identification theater if your legal team has already approved a BAA-covered sample exchange. Do wait if someone is about to email identifiable notes to a personal Gmail. The delay that matters is the unsafe one. A clean packet on Thursday beats a beautiful packet next month. Name a single clinic owner who can answer 'does Dr. Chen want the assessment numbered' without a meeting. Configuration dies in group chats. It lives in one person with authority and a vendor counterpart who writes the decisions down.
A one-week loop instead of a six-week project
Day one: packet sent, users created, microphones chosen. Day two: shadow a short clinic with drafts going to a holding area, not the official chart. Day three: clinician marks five recurring errors. Day four: template and glossary updated. Day five: live on a limited visit type with mandatory review. That is a week. It is not reckless if the notes are not auto-signed and the old method still works. It is reckless if you flip every provider on Monday with no backup and a generic template. Speed without a holding area is how bad drafts become official.
The loop only works if errors are specific. 'Make it better' is not an instruction. 'Stop adding a complete ROS I did not perform' is. 'Put the device clinic settings in this heading' is. Have clinicians comment in the draft or send two-line emails, then stop commenting on style that is already correct. You are training a system, not grading an intern's personality. Sunrise can turn typical clinic drafts around in about two minutes, which makes this loop practical: review at the desk, mark the miss, move on.
EHR delivery without waiting on an interface team
Deep integration is nice. It is not a prerequisite for a useful week. Many clinics start with a secure queue, copy into the correct encounter, and add the interface later. It is how you learn whether the prose is right before you spend political capital on IT. When you do integrate, specify the encounter match, the document type, and the addendum path. EHR integrations fail when those three are assumed. They succeed when they are written. A paste ritual with a named filer is a valid phase. An undocumented paste ritual is how notes land on the wrong patient.
Keep identity hygiene from day one even if the interface is later. No shared logins. Scribe or vendor access only through the agreed role. Audio retained only for the period you named. If IT's first available sprint is in six weeks, you can still be documenting with a trained template this Friday. Tell IT the truth: you are not asking them to skip security. You are asking not to confuse security with a custom button. Custom buttons can wait. Invented exam language in a live chart cannot.
When to stop tuning and start using it
Stop when the discarded-draft rate is low on the visit types you included, and when the same two errors are not still appearing. You will never reach zero edits. Editing is clinical work. Tuning is removing the edits that are really configuration. If a clinician is still rewriting histories after three loops, look at whether they verbalize, whether the visit type belongs in ambient capture, or whether they simply prefer to type. Let them type. A forced conversion is not a training plan. It is a second burden with a kickoff slide.
Clinics should keep transcription or virtual scribes available for letters, procedures, and the holdouts. Mix is allowed. Per-line work can cover the leftovers while ambient handles the panel's routine visits. Write the mix down so billing and staff are not surprised. Then freeze the template for a month. Endless tweaking is another six-week project in disguise. Reopen it when a new clinician onboards or a specialty clinic is added. Training an AI scribe is closer to teaching a new hire your note than to installing a pacemaker. Act like it, and IT can stay in their lane.