Insurance, 28 October 2026
EvenUp: AI Agents Cut Personal-Injury Case Drafting Time by 99%
EvenUp — an AI platform for personal-injury (PI) law firms, now valued at over $2 billion, used across the full case lifecycle by solo practitioners and multi-state firms alike. EvenUp says it has helped resolve more than 200,000 cases and secure over $10 billion in damages for injury victims.
Problem
A personal-injury case can mean 1,000+ pages of medical records, billing ledgers and police reports, arriving late, out of order, and inconsistent — the same hospital visit documented three different ways by three different providers. Turning that pile into a demand letter, discovery response or mediation brief used to be entirely manual, skilled work. At one Ohio firm, paralegals built medical chronologies and totaled bills by hand: 8 to 15 hours per case. At a California firm with 1,500+ active cases, all demand writing ran through a single person, creating a 45-day backlog that slowed every case behind it. Firms facing that math had to either eat the hours or turn the case away.
Result
99%
- Drafting time per case fell from 8–15 hours of skilled work to about 30 minutes. EvenUp calls this roughly a 99% drop, and its own materials headline it as “15 hours to 15 minutes” — the exact figure moves depending on which case is used as the example, but the shape holds everywhere EvenUp describes it: multi-hour skilled work compressed to well under an hour.
- Settlement offers rose 300% at customer firms presenting a fuller, cited damages picture; demand output tripled.
- The California firm cleared its 45-day backlog: demands that took a month now take about ten minutes to produce, and attorney review runs 75% faster.
- Discovery responses dropped from 5 hours to 30 minutes; medical-expense accuracy rose 14% on smaller cases; draft completeness rose 16%.
- An Oregon firm cut drafting turnaround from months to days and reported 400% revenue growth in one year.
Solution
EvenUp built a document pipeline, not a single chatbot. Each document is sorted by type, then read at the cheapest method that works — native text first, then standard OCR, then Claude’s vision only for scanned bills and handwritten checkboxes. Every fact extracted keeps a citation to its source page, and duplicates across providers are merged, so the case record is settled once, before any drafting starts. Only then does drafting begin: Claude Opus builds the case “blueprint” — the applicable legal rules, likely counterarguments, the throughline of the case — and checks the finished document for consistency; Claude Sonnet writes each section from that blueprint, following state-specific rules written out in plain English; Sonnet or Haiku handles routine extraction. EvenUp’s own rule of thumb, from principal ML engineer Emre Yamangil: spend the expensive model only where an error would be hard to catch later. A human attorney reads, checks and signs every draft before it goes out.
Could a small team do this?
Largely yes, in a narrower slice. The hard-to-replicate parts are scale (200,000+ cases), the legal domain depth needed to get every US state’s rules right, and the trust a $2 billion valuation buys with risk-averse law firms. But the actual technical pattern — sort documents, extract with the cheapest method that works and escalate to vision only when needed, cite every fact to its source page, merge duplicates into one settled record, then draft with a capable model following plain-English rules and a cheaper model for routine steps — is exactly the kind of pipeline a solo operator can build with Claude’s API in a few weeks, no special infrastructure beyond file storage and a queue. A realistic solo version isn’t “replace EvenUp”; it’s picking one narrow, recurring paperwork bottleneck — medical chronology summaries for two or three local firms, or discovery-response drafts for one practice area — and selling that as a boutique service, with a human signing off on every draft exactly as EvenUp’s attorneys do. The real risk to manage, not just the AI: an error in a legal document has real financial consequences, so the review step is not optional.