PCCD AI Workflow — Recommended Process

The Framework

Three steps to make AI work for PCCD

PCCD's volume is hygiene; its margin is cosmetic. Each step below is built to push more of the practice's real voice, real calls, and real workflows into tools the team already uses — skewed toward the veneers and Invisalign conversations that drive margin.

Step 1 · In Motion

Custom GPT Voice Training

Train a Custom GPT on 15 PHI-scrubbed real phone calls plus recorded answers from both doctors. Result: front-desk responses and patient communications in PCCD's authentic voice — not generic AI copy.

Step 2 · Next

Batched Content Calendar

A monthly ChatGPT-batched calendar centered on veneers + Invisalign. Doubles as a DJ endorsement asset library with a UTM-tagged /dj landing page, so the radio spend finally becomes measurable.

Step 3 · Then

SOPs from Recorded Workflows

Record a workflow once (screen or voice), and Claude drafts the SOP. Builds the operations manual as a byproduct of work the team already does.

The Centerpiece

Call Library Pipeline — PeerLogic export to training pack

A six-stage automated pipeline (built, tested, committed) turns one 30-day PeerLogic opportunity export into a ready-to-load Custom GPT training pack. PHI is stripped locally before anything touches an AI model.

Export

Dakota pulls the 30-day PeerLogic opportunity export for both locations.

Dakota

Filter

Pipeline keeps only PCCD Los Altos + Mountain View rows.

Local

PHI Scrub

Regex redaction of names, phones, DOBs, fees, insurance IDs. No network, no LLM.

Local

Classify + Score

Claude classifies each scrubbed call and scores tone. Scrubbed text only.

Claude API

Select 15

Best 8 / 5 / 2 mix, 60%+ cosmetic-skewed. Anti-examples flagged for human confirmation.

Local

Outputs

15 tagged transcripts + filled worksheet + missed-opportunity report. Human eyeballs before GPT load.

Review

The free side report. The same pipeline run produces the top 5 patterns of missed opportunities across the 30-day window — which makes the front-desk coaching that follows targeted at what is actually being missed, instead of generic phone-skills training.

Training Data

The 15-call mix

8 converted calls (cosmetic-skewed) — the gold standard 5 well-handled non-converted — good process, no booking 2 anti-examples — tagged “do not replicate”

The anti-examples are the soft spot: the pipeline flags candidates, but a human must confirm them before they go in the pack. If 30 days yields fewer than 6 cosmetic-intent calls, widen the export to 60 days.

Compliance

HIPAA guardrails

Raw PeerLogic exports are treated as PHI until scrubbed

The PHI scrub runs locally — no transcript leaves the machine before redaction, and nothing raw is ever pasted into ChatGPT, Claude, or any AI tool.
Only scrubbed text reaches the classification model. One open decision, with Ron: keep the standard Anthropic API or swap to AWS Bedrock Claude (BAA-eligible). Roughly a 10-minute code change either way; it gates the first full run.
Raw exports move by secure internal share (OneDrive/SharePoint) — never unencrypted email.
Every final transcript is human-reviewed for PHI leakage before anything loads into the Custom GPT.

Execution

Who does what

Owner	Actions
Dakota PeerLogic + Neurality access	1) Send the export header row + the practice-ID, outcome, and intent values PeerLogic uses (no PHI required for this). 2) Deliver the full 30-day opportunity export for both locations via secure share. 3) Sanity-check volume: 6+ cosmetic-intent calls in 30 days, else pull 60. Neurality access comes into play in Step 2 (UTM'd scheduler links, /dj page).
Office Manager	Get the Doctor Voice Capture worksheet in front of both doctors — 10 prompts, ~30 minutes of recorded audio each. This is the slowest human step in the whole project; it can start today.
Amy	Lock the export column mapping, run scrub-only first, review scrubbed previews, then run the full pipeline and assemble the Custom GPT pack.
Ron (CIO)	Call the HIPAA posture for the classification stage: standard Anthropic API vs. AWS Bedrock (BAA-eligible).

Sequence

Run order once the export lands

Adjust the pipeline config to the real PeerLogic columns and PCCD practice IDs (from Dakota's header row).
Scrub-only run — nothing leaves the machine; eyeball the scrubbed previews for PHI leakage.
Full run after Ron's HIPAA decision → 15 transcripts + filled worksheet + missed-opportunity report.
Human review of all 15, confirm the 2 anti-examples, then load the Custom GPT.
Fold in the doctor voice recordings; move to Step 2 (content calendar + DJ asset library).