Skip to main content

AI-Powered Document Automation for Specialty Pharma

How Katyella delivered a production-ready AI platform in 4 weeks with almost zero meetings

4 weeks

SOW to delivery

75%

time reduction

100%

claim coverage

<$500

monthly infra cost

Client

Evo Advertising (via Lucien Vaccaro)

Industry

Specialty Pharma / Healthcare Market Access

Engagement

Fixed-scope MVP with optional retainer

The Challenge

Evo Advertising's pharma client faced a critical bottleneck in their market access workflow: Letters of Medical Necessity (LMNs).

Every time a physician needed to prescribe a specialty gene therapy, they had to manually draft a multi-page letter to the patient's insurance payer, citing specific clinical evidence, aligning to payer-specific coverage criteria, and formatting everything to regulatory standards.

Hours per letter

Each LMN required 2-4 hours of physician time to research payer policies, pull clinical citations, and draft custom prose.

Payer fragmentation

Every payer has different coverage criteria, required documentation, and policy language. One template doesn't fit all.

Compliance risk

Manually assembled letters risked omitting required clinical claims or using non-approved language.

Provider friction

The administrative burden discouraged providers from prescribing, directly impacting patient access to treatment.

Before and after workflow comparison: 2-4 hours manual process reduced to 30 minutes with AI platform

The Solution

Katyella built a full-stack AI platform that transforms a 5-minute physician survey into a complete, payer-compliant Letter of Medical Necessity, generated in under 30 minutes.

1

Guided Survey

A structured form captures 29 clinical data points across 6 sections: coverage details, clinical profile, functional assessments, treatment history, quality-of-life impact, and clinical rationale.

2

Deterministic Template Assembly

Survey answers are routed through a logic map that selects the correct pre-approved clinical claims for the specific payer and authorization type. Every claim is deterministically selected from a curated library.

3

AI Polish Pass

Claude processes the assembled draft, improving grammar, flow, and professional tone without removing or altering any clinical claims. If the AI pass fails, the system falls back to the deterministic draft.

4

Dual-Format Output

The platform generates both an editable Word document for physician review and an archival PDF simultaneously. PHI placeholders appear in highlighted brackets for easy identification.

5

Email Delivery + Admin Dashboard

The completed letter is delivered to the physician's inbox within 60 seconds. A searchable admin dashboard tracks every submission with a complete audit trail.

System Architecture

System architecture diagram showing data flow from physician survey through logic map, template assembly, Claude AI polish, to dual document output and email delivery

Payer Intelligence

The platform supports a matrix of payer-specific claim libraries, ensuring every letter meets the exact requirements of the target payer.

Initial Authorization Reauthorization
Aetna 8 payer-specific claims 7 payer-specific claims
UnitedHealth 9 payer-specific claims 7 payer-specific claims

Plus 29 core clinical claims shared across all payers, each derived from approved clinical evidence with citations.

Payer intelligence matrix showing how survey input flows through logic map to payer-specific claim paths and converges into a complete letter

Tech Stack

Frontend

Next.js 16, React 19, TypeScript, Tailwind CSS 4

AI Engine

Claude API (Haiku 4.5) via Anthropic SDK

Database

Neon Serverless PostgreSQL + Drizzle ORM

Infrastructure

Vercel, Vercel Blob, Resend

The Delivery

4 Weeks. Minimal Meetings.

4-week project delivery timeline: Week 1 Requirements and Survey Form, Weeks 2-3 AI Pipeline and Documents, Week 4 Admin and Deploy

What the client expected

Weeks of discovery meetings, multiple design revisions, iterative feedback loops to nail the look and feel.

What actually happened

Katyella shipped full working prototypes early and often, paired with async video walkthroughs. The client could test real functionality, not review mockups.

Style and UX questions resolved themselves because the client could use the thing instead of imagining it.

Results

Speed

Before: 2-4 hours per letter (manual drafting)
After: Under 30 minutes per letter (AI-generated)

Quality

  • 100% of required clinical claims included in every letter
  • Payer-specific policy alignment baked into the logic
  • Complete audit trail for every submission

Delivery Efficiency

  • 4-week delivery from signed SOW to working platform
  • Working prototypes replaced meetings
  • Async video walkthroughs instead of status calls

Business Impact

  • Demo platform led to an expanded engagement
  • Proved viability of AI-assisted document automation
  • Reusable architecture for additional products and payers

Why It Worked

Prototypes Over Presentations

Instead of burning weeks on wireframes and design reviews, Katyella shipped functional prototypes within days. The client tested real software, not slide decks. This collapsed the feedback loop from weeks to hours.

Deterministic AI, Not Prompt Roulette

The letter generation pipeline doesn't rely on AI to "figure out" what to include. A logic map deterministically selects every required claim. AI only handles prose quality. If the AI layer fails, the letter still generates correctly from the template.

Full-Stack, One Team

No handoffs between a design agency, a frontend shop, and a backend contractor. Katyella owned the entire stack: survey UX, AI pipeline, document generation, email delivery, and admin dashboard. One team, one codebase, one deployment.

Right-Sized Architecture

Serverless Postgres, edge-deployed Next.js, blob storage, transactional email. No over-engineered microservices. No Kubernetes. Infrastructure costs under $500/month, not $5,000. The architecture matches the actual scale of the problem.

Ready to Automate a Manual Workflow?

We build production systems with AI, not proofs of concept that need to be rebuilt.