> ## Documentation Index
> Fetch the complete documentation index at: https://docs.corti.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# AI Scribing

> Learn about best practices for audio capture and user behavior in AI scribing workflows

This guide covers how to get reliable, high-accuracy audio into an AI scribe built on the Corti speech to text API capabilities. It's split into two parts: guidance for **IT and implementation teams** rolling out and operating the scribe, and guidance for **clinicians** using it day to day.

* Audio quality is the single biggest driver of transcription accuracy: many note-quality issues trace back to microphone choice, placement, or how the encounter was verbalized.
* User behaviors are the single biggest driver of note completeness: when generated document details are derived from captured audio, clinicians must steer the conversation and information capture proactively.

<Tip>
  For the underlying device and audio format specifications, see audio [format](/stt/audio) and [configuration](/stt/best-practices-audio), [recording best practices](/stt/best-practices-recording), and [recommended microphone devices](/stt/microphones).

  The sections below focus on turning those specs into a working rollout.
</Tip>

***

## Part 1: for IT and implementation teams

These are practical setup and rollout tips for teams standing up an AI scribe on Corti: device provisioning, network, and workflow decisions that affect audio quality before a single encounter is recorded.

<Accordion title="Device and connection">
  * **Standardize on a short list of approved devices** rather than letting clinicians bring their own. Pick a near-field mic or headset for dictation workflows and a beamforming array (or array-equipped webcam) for ambient, in-room workflows, and issue the right one for how each clinician actually works. iPhones also work great when best practices are followed.
  * **Default to wired USB devices.** They're more reliable at scale than Bluetooth, which introduces battery- and range-related failure modes that are harder to diagnose remotely.
  * **Use a high-speed, stable network connection** for every recording device or workstation. A weak or congested Wi-Fi signal is one of the most common causes of dropped audio and incomplete notes. Fail back to offline audio recording and file transcription when reliable signal is not available.
  * **Don't let other applications compete for the microphone.** Close or block software that can silently grab the mic (softphones, meeting apps, screen recorders) on clinical workstations. Be sure to review hardware device defaults on the workstations.
</Accordion>

<Accordion title="Room and workflow setup">
  * **Position devices centrally** for ambient/room capture and **close to the speaker** for dictation. Don't leave devices in pockets, bags, or drawers as the microphone can be obstructed.
  * **Reduce avoidable background noise** where you have control over the physical space: close doors, move away from shared printers or hallway traffic, and prefer carpeted rooms or add sound dampers to the wall over hard, reflective surfaces.
  * **Build patient consent into the workflow, not just the policy.** A checkbox, verbal-consent prompt, or EHR flag at the start of each encounter ensures consent isn't left to memory.
  * **Design for one patient per recording.** Typically AI scribes link recordings and generated facts or documents to specific patient encounters. The application is not built to dynamically separate multiple unrelated patients from a single recording. Therefore, the workflow should make starting a new recording per patient the default action.
</Accordion>

<Accordion title="Ongoing support and monitoring">
  * **Turn on audio quality alerts** in the app so clinicians get real-time feedback (for example, a quiet indicator) when background noise or a dropped connection is degrading capture, instead of finding out after the note is generated.
  * **Give clinicians a clear, fast way to report issues**, and route those reports to whoever manages devices and accounts. Most "the AI got it wrong" tickets turn out to be a mic placement or connectivity issue rather than a model issue.
  * **Pilot before a full rollout.** Test in real clinical rooms with real background noise, not just a quiet office, and include every workflow you plan to support (dictation, ambient, telehealth) since device needs differ across them.
  * **Never leave devices logged out or uninstalled as a first troubleshooting step.** Check known issues or contact support first, since logging out can compound recording and upload problems rather than fix them.
</Accordion>

***

## Part 2: for clinicians

<Accordion title="Set up the room">
  * Place the microphone or device between you and the patient with a clear, unobstructed line of sight to both speakers.
  * Keep background noise low: close doors, mute notification sounds, and avoid recording near HVAC vents or hallways when possible.
  * If using a phone, place it on a stable surface within 1 to 3 m of the conversation; don't put it in a pocket or bag.
</Accordion>

<Accordion title="Get consent, then start clean">
  * Always obtain and document patient consent to record before starting, per your organization's policy.
  * Record one patient encounter at a time. The AI scribe is not designed to separate multiple unrelated patients from a single recording.
</Accordion>

<Accordion title="Talk naturally, but verbalize everything" defaultOpen="true">
  The AI scribe can only capture what's spoken aloud. Non-verbal cues, gestures, and unspoken observations won't make it into the note unless you say them.

  * **Converse, don't dictate.** Speak to the patient as you normally would; there's no need for robotic phrasing or command-style speech.
  * **Introduce the patient visit context.** At the start of the recording, consider explaining what your primary focus area during the visit will be to prime the AI scribe for pertinent details.
  * **Verbalize the physical exam.** Say what you're doing and what you observe as you do it:
    * Tie findings to a specific body location and laterality: "Both ears are clear of blockages" rather than "all clear over here"; "I'm ordering an x-ray for your left shoulder" rather than "let's get an X-ray for this."
    * Use sensory and action language: "On neck palpitation I feel a swollen lymph node" rather than "there's swelling here."
    * Cue the transition into the exam: "Let's take a look at your throat now." A brief verbal signal helps the AI recognize the exam has started.
    * Say normal findings explicitly ("Heart exam is normal"). The system reflects what's said; it won't infer or expand a vague statement into a templated finding.
  * **Speak your reasoning and plan out loud**, including why you're ordering a test or changing a medication. This context is what turns a transcript into a clinically useful note.
  * **State orders with specifics:** drug name, dose, and frequency ("I'm ordering Lasix 40 mg twice a day") rather than "I'll put you on a diuretic."
</Accordion>

<Accordion title="Handle interruptions and pauses deliberately">
  * Pause and resume recording as needed: for phone calls, sensitive discussions you don't want captured, or between patients.
  * If you're interrupted mid-visit, resume recording by bringing the app back to the foreground when you're ready to continue.
  * Use the time immediately before or after the visit to add verbal context or clarify terminology if something was rushed during the encounter.
</Accordion>

<Accordion title="Before you finalize">
  * Always review the AI-generated note before signing. It's a strong first draft, not a final clinical judgment.
  * Check that exam findings, orders, and plan details match what you intended to convey, especially anything with laterality or specific dosing.
  * If the note is consistently missing or incorrectly attributing details, then revisit how findings are being verbalized and \[submit feedback] for your IT and Corti to review.
</Accordion>

***

## Quick reference checklist

* [ ] Correct microphone type for the workflow (directional for dictation, beamforming array for ambient)
* [ ] Wired connection where possible; Bluetooth used only with adequate battery and range
* [ ] Audio format configured per [Corti's audio configuration reference](https://docs.corti.ai/stt/audio) (16-bit, 16 kHz, mono PCM recommended)
* [ ] Audio events enabled and surfaced to the user
* [ ] Room background noise minimized; mic unobstructed and correctly positioned
* [ ] Patient consent obtained and encounter identified at the start of recording
* [ ] Key context, findings, physical exam results, orders, plan and reasoning verbalized throughout the recording
* [ ] Note reviewed and confirmed before finalizing

<br />

<Note>Need further guidance and support setting up your AI scribe? Please [contact us](mailto:help@corti.ai) to discuss.</Note>