# ReviewOS — Executive Memo

**To:** MediGen leadership · **From:** AI Strategy · **Re:** Making the document corpus AI-addressable

## Recommendation

Do not start by building a document-review application. Start by making MediGen's existing ~50K-document corpus **AI-addressable**: a secure, read-only retrieval layer (API + MCP server) that returns source-grounded, cited passages and can be called from the AI tools teams already use. Prove retrieval value and adoption first, then layer ingestion, a citation agent, and a workflow application only as usage justifies it.

## The Problem

Legal, regulatory, and research teams spend a large share of their time turning a fragmented corpus into defensible, cited answers — searching repositories, re-reading long documents, hunting for the exact supporting passage, copying citations, summarizing, and reconciling sources by hand. The knowledge already exists; it just isn't operationally accessible. For a 25-person team this plausibly runs ~29,600 hours and **$4.1–4.5M** in annual fully loaded cost, concentrated in search, retrieval, citation, summarization, and synthesis.

## Why Now

The work is language-heavy, repetitive, evidence-grounded, and governed by review rubrics — an ideal fit for retrieval-augmented AI with citations. Hybrid search (semantic + exact-term) over an approved corpus can return the right passage in seconds, with full source traceability, while experts retain accountability for judgment.

## The Solution (staged)

| Stage | What it delivers |
|---|---|
| **MVP 0 — Corpus API + MCP** | Hybrid search + cited passages over the existing corpus, read-only, permission-aware |
| **MVP 1 — Ingestion** | Authorized users keep the corpus current |
| **MVP 2 — Citation agent** | Grounded multi-document synthesis, contradiction/gap detection |
| **MVP 3 — ReviewOS app** | Projects, queues, approvals, exports — built only after usage proves the workflows |

## Value Case

If 60–70% of review hours are AI-assistable and 25–40% of that is captured as productivity, the annual upside plausibly exceeds **$750K–$1.5M** — before faster cycle time, reduced rework, better consistency, and stronger auditability. The first wedge (MVP 0) targets the largest, most measurable cost: time-to-cited-answer.

## Roadmap

- **Week 1:** corpus assessment, metadata map, 30–50 question eval set
- **Weeks 2–3:** MVP 0 — searchable, citable corpus callable from existing AI tools
- **Week 4:** MVP 1 — ingestion / corpus maintenance
- **Weeks 5–6:** MVP 2 — headless citation + cross-reference agent
- **Weeks 7–12:** workflow discovery → MVP 3 application, if justified

## Trade-Offs

Cut from scope: autonomous legal/regulatory judgment, broad eDiscovery replacement, deep per-DMS integrations, model fine-tuning, automated external production. MVP 0 is deliberately **read-only** to minimize security and operational risk. Build the substrate first; productize the workflow second.

## Decision Ask

Approve a **6-week pilot** to stand up MVP 0–2 against a real corpus slice and validate retrieval accuracy (>80% top-5), citation faithfulness (>95%), and cycle-time compression before committing to the full application.