---
name: medical-mode-claims
description: The source of truth for how to describe AssemblyAI's Medical Mode in any content — blog, web, product pages, ads, email, social, sales replies, decks, docs, and FAQ. ALWAYS use this skill whenever you are writing, drafting, editing, or reviewing AssemblyAI content that mentions or could mention Medical Mode, medical or clinical transcription, healthcare speech-to-text, medical entity accuracy, clinical documentation, ambient scribes, EHR/EMR, ICD/CPT terms, or the medical-v1 domain — even if the user does not say "Medical Mode" explicitly. Use it to pull the exact approved accuracy, competitive, activation, language, pricing, and compliance claims, and to catch any Medical Mode claim that has drifted from canon. Pairs with /hipaa-language-check for the HIPAA/BAA wording.
---

# Medical Mode claims

This skill is the canonical, approved way to describe **Medical Mode** — AssemblyAI's clinical-accuracy add-on for Universal-3.5 Pro and Universal-3.5 Pro Realtime. Marketing, sales, and docs all draw from the same numbers and language so the product reads consistently everywhere. When you write or review any Medical Mode content, use these exact claims. If existing copy says something different, treat this sheet as the source of truth and flag the divergence — don't invent a new phrasing or silently match the wrong one.

Why it matters: Medical Mode sells into healthcare, where over-claiming on accuracy or compliance creates real legal and credibility risk. Precise, developer-credible claims beat impressive-sounding ones.

## Approved claims — use these exact versions

| Claim | Canonical version | How to use |
| --- | --- | --- |
| **Accuracy** | 3.2% Missed Entity Rate (MER). ~20% fewer missed medical entities vs. Universal-3.5 Pro alone. | Lead with 3.2% MER as the absolute figure; use ~20% as the improvement over base Universal-3.5 Pro. See the 87% note below — it's a separate, also-correct stat. Don't round up or merge the numbers. |
| **Competitive** | Lowest Missed Entity Rate across benchmarked providers — Deepgram, Speechmatics, AWS, Google. | Always link to /benchmarks. "Across benchmarked providers" is the most defensible form; the on-page benchmark table also supports "lowest medical entity error rate in the industry." |
| **Activation** | One parameter: `domain: "medical-v1"`. Works on Universal-3.5 Pro and Universal-3.5 Pro Realtime. | This is the developer hook — lead with the single-parameter simplicity and that it runs on both async and streaming. |
| **Languages** | English, Spanish, German, French — pre-recorded and streaming. | Name all four languages and both modes; don't imply broader language coverage. |
| **Pricing** | $0.15/hr add-on. | Frame as an add-on on top of Universal-3.5 Pro, not a standalone product price. |
| **Compliance** | PHI redaction across audio and transcripts. BAA available for customers processing PHI. SOC 2 Type 2. | Supporting feature, never a headline. See "Compliance language" below and run /hipaa-language-check. |

## The "87% fewer entity errors" stat

Medical Mode "reduces medical entity errors by ~87% versus the base Universal-3.5 Pro model" is **confirmed correct** and is the headline stat used across the site. It is a *different metric* from the accuracy row above, not a contradiction:

- **87%** = reduction in entity **errors** vs. the base model.
- **~20%** = reduction in **missed** entities (MER) vs. base Universal-3.5 Pro; **3.2%** = the absolute MER.

Both are valid because they measure different things. Use whichever fits the context. Do **not** "correct" 87% down to 20%, and don't flag 87% as unsourced — it's approved.

## Compliance language (HIPAA / BAA)

Compliance is a **supporting feature, never a headline or badge**. AssemblyAI is a *business associate* under HIPAA, not a covered entity, so we never claim to *be* HIPAA-anything. Run **/hipaa-language-check** on any healthcare content; the essentials:

**Say:** "BAA available for customers processing PHI" · "AssemblyAI is a business associate under HIPAA and signs a standard BAA" · "BAA-eligible infrastructure" · "PHI redaction across audio and transcripts" · "SOC 2 Type 2." For a badge or eyebrow, use **"BAA Available."**

**Never:** "HIPAA-compliant," "HIPAA-ready," "HIPAA-certified," "comply with HIPAA," "Full HIPAA compliance," or "HIPAA-eligible." Never use HIPAA as a headline or a one-word badge.

This is the approved language per the AssemblyAI design system. Don't source compliance wording from Slack. For public-facing references, point to the BAA FAQ (`assemblyai.com/docs/faq/can-you-sign-a-baa`) and the BAA legal page (`assemblyai.com/legal/business-associate-agreement`).

## How to apply

1. Pull the exact claim from the table above — don't paraphrase the numbers.
2. For accuracy, pick the metric that fits the context: 3.2% MER (absolute), ~20% fewer missed entities (vs. base), or 87% fewer entity errors (vs. base). All three are approved.
3. For anything touching HIPAA, PHI, or BAA, run /hipaa-language-check before finalizing.
4. Keep compliance a supporting point, never the headline.
5. If you find live copy that contradicts these claims, flag it for correction rather than matching the wrong version.

## Quick reference

- **Accuracy:** 3.2% MER · ~20% fewer missed entities vs. U3 Pro · 87% fewer entity errors vs. base model (all approved)
- **Competitive:** lowest MER across benchmarked providers (Deepgram, Speechmatics, AWS, Google) → link /benchmarks
- **Activate:** `domain: "medical-v1"` on Universal-3.5 Pro + Universal-3.5 Pro Realtime
- **Languages:** EN, ES, DE, FR — pre-recorded + streaming
- **Pricing:** $0.15/hr add-on
- **Compliance:** PHI redaction · BAA available · SOC 2 Type 2 — supporting feature only; run /hipaa-language-check