AI integrations
Live todayAI that reads faxes, drafts messages, and writes quiz questions — inside HIPAA aware ABA software
Google Gemini powers OCR, compose assist, and MCQ generation today. An MCP server for Claude is next. AI access is scoped to operational aggregates. PHI never leaves your controlled environment.
Pediatric Partners
Carolina Dev. Clinic
Unknown · 919-555-0142
AI extraction
Gemini OCR · 98% confidence- Referral type
- ABA assessment
- Referring provider
- Dr. ██████, MD
- Client
- ██████ ██████
- DOB
- ██/██/20██
- Payer
- BCBS NC
- Diagnosis
- F84.0
What are TargetFlo AI integrations?
TargetFlo AI integrations are narrowly scoped uses of Google Gemini inside HIPAA aware ABA software: OCR and field extraction on inbound faxes in the Fax Inbox, compose assist for SMS and email broadcasts, and multiple-choice question generation in the course builder. A planned MCP server lets Claude answer operational questions from non-PHI aggregates. Every feature has a documented input, a human confirmation step, and an audit record.
Four AI features
Three live today, one on the roadmap
Each feature does one job with one kind of input.
Gemini OCR on faxes
Every inbound fax page is read and text-extracted, with suggested client, guardian, payer, and referring-provider fields for a coordinator to confirm.
Compose assist
Draft SMS and email broadcasts faster. The assistant proposes wording from your prompt; the PHI guard still checks every outbound text.
MCQ generation for courses
Turn a chapter's video transcript or PDF into multiple-choice questions for the course builder. Trainers edit and approve before publishing.
MCP server for Claude
Ask Claude operational questions — pipeline counts, task summaries, training completion — against non-PHI aggregates only.
Gemini OCR
Live todayFrom a scanned referral to pre-filled intake fields
Referral faxes arrive as images. Gemini OCR reads each page, extracts the text so the document is searchable, and proposes the fields an intake coordinator would otherwise type: client name, DOB, guardian, referring provider, payer, and diagnosis codes.
- Multi-page documents handled as one referral
- Suggested fields are highlighted and confirmed by a coordinator — never auto-written
- Extracted data stored in the isolated PHI schema alongside the document
- Linked to a client or a New Intake card in one action
Pediatric Partners
Carolina Dev. Clinic
Unknown · 919-555-0142
AI extraction
Gemini OCR · 98% confidence- Referral type
- ABA assessment
- Referring provider
- Dr. ██████, MD
- Client
- ██████ ██████
- DOB
- ██/██/20██
- Payer
- BCBS NC
- Diagnosis
- F84.0
Compose assist & MCQ generation
Live todayFaster messages and faster training content, with a person in the loop
Two smaller uses of the same model. Compose assist helps front office staff draft a clear broadcast to families or staff. MCQ generation turns a training chapter into a draft quiz so trainers spend their time reviewing, not writing.
- Compose assist drafts from a plain-language prompt; the PHI guard checks SMS before send
- MCQ generation reads chapter transcripts or PDFs and drafts questions with answer keys
- Trainers approve, edit, or regenerate before a test is published
- Both features can be disabled per organization
1. Which CPT code covers direct ABA therapy delivered by an RBT?
ABCD2. When must a session note be completed relative to the session?
ABCD3. Who signs the Plan of Care before authorization is submitted?
ABCD
What each feature sees
Data scope, feature by feature
The right question about AI in healthcare software is not 'is it smart' but 'what does it see'. Here is the answer for each feature.
| Feature | Input it receives | What it produces | PHI involved? |
|---|---|---|---|
| Gemini OCR | The fax pages a coordinator is triaging | Extracted text and suggested fields, shown for human confirmation | Yes — fax content is PHI; processed inside the PHI boundary, stored in the PHI schema |
| Compose assist | The draft message and the prompt the sender types | Suggested wording; nothing is sent until a human clicks send | Only what the sender types; the PHI guard blocks identifiers in SMS |
| MCQ generation | Course chapter text, video transcript, or PDF | Draft questions and answer keys for trainer review | No — course material is training content, not client data |
| MCP server (roadmap) | Server-computed aggregates: counts, rates, queue sizes | Answers to operational questions in Claude | No — never client names, DOB, diagnosis, or insurance IDs |
Safety
AI access is scoped to operational aggregates. PHI never leaves your controlled environment.
Scoped by feature
Each AI feature is wired to one input: fax pages, a draft message, a chapter, or an aggregate query. There is no general-purpose model with access to the database.
Human confirms
OCR fields, compose suggestions, and generated questions are proposals. A person confirms before anything is written to a record or sent.
PHI stays in the PHI schema
Extracted fax data is written to the isolated PHI schema. Operational AI — including the planned MCP server — reads only from ops-level aggregates.
Logged like any other action
OCR runs, compose uses, and MCQ generations are recorded in the audit trail with actor and timestamp.
No training on your data
TargetFlo uses Gemini through enterprise API terms; your documents and messages are not used to train foundation models.
Aggregates only for assistants
The MCP server for Claude will answer with counts and rates computed server-side. It cannot look up an individual client.
MCP server for Claude
Coming soonAsk Claude about your operation — never about a client
TargetFlo is building a Model Context Protocol server that exposes non-PHI operational data to Claude: pipeline stage counts by location, fax queue status, task summaries, and course completion rates. Aggregates are computed server-side; the model never receives an identifier.
- “How many referrals entered the Insurance stage this month?”
- “Which location has the oldest unassigned faxes?”
- “What percentage of RBTs completed the HIPAA refresher?”
- Scoped tokens, server-side aggregation, full audit log
How many referrals are in Pending Authorization by location?
Raleigh 14 · Charlotte 9 · Durham 6. Raleigh's median age in stage is 11 days, up from 8 last month.
Counts only. No names, DOB, diagnosis, or insurance IDs are available to the model.
FAQ
AI integration FAQs
See Gemini OCR read one of your real referral faxes
Book a demo and bring a redacted sample referral. We run OCR, show the suggested fields, and walk through exactly where the extracted data is stored.