Use cases
Real workflows. Specific reasons they stall.
These are the healthcare AI workflows we take to production. Each one works in a demo. Each one has a concrete reason it doesn't go live - and a concrete fix.
01Fax and document intake
Health plans, TPAs and benefits administrators
The workflow
Faxes, PDFs and scanned forms arrive all day. A model classifies each one, pulls the key fields and routes it to the right queue, replacing manual sorting.
Why it stalls
- Documents come from outside parties, and any of them can contain text the model treats as instructions.
- Extracted fields feed downstream systems, so a wrong member ID or date is a real error, not a typo.
- The OCR, model and storage vendors each touch PHI, and the BAA chain has gaps.
- Fax / PDF
- OCR
- Injection filter
- Model: classify + extract
- Confidence check
- Queue or human review
What we do
- Build an evaluation set from a sample of your real documents, scored field by field
- Test prompt injection through the documents themselves, then add input filtering and output checks
- Set confidence thresholds that send uncertain documents to a person
- Map every vendor in the path and close BAA and retention gaps
What you walk away with
- Field-level accuracy on your documents, not a vendor benchmark
- Routing rules your operations lead can read and change
- Injection test results your security team can review
02Medical record summarization
Workers' comp, legal-medical review and independent review
The workflow
Reviewers read hundreds of pages of medical records per claim. A model drafts a chronology and summary, so the reviewer starts from a draft instead of a stack.
Why it stalls
- A summary that misses a prior injury or invents a date can change a claim decision.
- Reviewers won't trust a summary they can't check against the source.
- Whole records go to a model vendor, often with more PHI than the task needs.
- Medical records
- PHI minimization
- Model: chronology + summary
- Citation check
- Reviewer sign-off
What we do
- Require every statement in the summary to cite its source page
- Measure omissions and unsupported statements on records your reviewers already closed
- Strip identifiers the summary doesn't need before the model call
- Design the review step so reviewers can check any statement against its source
What you walk away with
- Measured rates for omissions and unsupported statements
- A review step your reviewers helped design
- Documentation of what PHI goes where, and why
04AI vendor security review
AI companies selling to health plans, hospitals and life sciences
The workflow
Your buyer's security team sends a long questionnaire and asks for your SOC 2 report. More of the questions are about AI every quarter, and the answers live in Slack threads and a few engineers' heads.
Why it stalls
- Questions about model providers, training on customer data and PHI in prompts have no written answers.
- There is no SOC 2 report yet, or the report doesn't cover the AI features.
- Engineers lose days to each questionnaire instead of shipping.
- Buyer questionnaire
- Answer library
- PHI data-flow doc
- Evidence binder
- Approved vendor
What we do
- Document how PHI moves through your product, end to end, including model providers
- Write a reusable answer library for the AI questions buyers ask
- Map your controls to SOC 2 and HITRUST and close the gaps
- Organize the evidence your assessor will ask for, before they ask
What you walk away with
- Questionnaire answers your sales team can send without an engineer
- A data-flow document you can share with buyers under NDA
- A readiness plan and evidence binder for your assessor
Have a pilot stuck in review?
Tell us what you built and where it's blocked. We'll tell you honestly whether we can help.