EvenUp cuts document drafting from 15 hours to 15 minutes with Claude

Weeks of skilled work inside every document

A personal injury case runs on documents: demands, negotiation sheets, discovery responses, medical summaries. Each one has to be formatted correctly, factually airtight, and toned to match the firm. Before EvenUp, getting there was entirely manual. "A paralegal with a highlighter worked through stacks of records that were weeks late and out of order, said Emre Yamangil, Principal Machine Learning Engineer at EvenUp. “The same visit could be documented three different ways by three different providers." 

One California firm juggled 1,500+ active cases, 250 of them in litigation at any given time, and ran its entire demand-writing function through a single person.  A 45-day backlog slowed every case behind it. Paralegals at an Ohio-based firm built medical chronologies and totaled bills by hand, taking 8 to 15 hours for each case. Firms faced a choice: take the case and eat the hours, or refer it out.

Selecting Claude for long-context performance

EvenUp selected Claude, with Opus 4.8 handling the deepest reasoning. The choice came down to two capabilities: consistency and long-context performance. "Other models can write one good paragraph, but Claude stands apart,” said Yamangil. “Keeping a single consistent story across thirty pages is the real test." Claude supplies that consistency; the routing and verification EvenUp built around it turn that into documents a firm can afford and an adjuster can't discount.

“PI law presents some of the toughest challenges for AI,” said Rami Karabibar, CEO at EvenUp. “Claude Opus 4.8 delivers a new level of reasoning, reliability, and long-context performance. EvenUp adds PI data, domain expertise, and purpose-built workflows.” 

Which Claude model gets a task depends on where a mistake could hide. Everyday extraction runs on Sonnet or Haiku. Opus takes the jobs where polish could hide an error: making sure the document's argument holds together. "If a model gets a date wrong, that's easy to catch," added Yamangil. "Attorneys and staff can check the source page. If it builds an argument that's subtly wrong, that mistake looks just as confident and polished as a correct one." 

Settle the facts once, then write

A case often arrives as 1,000 pages or more, and includes records from a dozen providers, billing ledgers, and police reports. EvenUp's drafting product called AI Drafts is supported by Claude and reads records on a ladder, climbing only as far as each page requires. 

The system reads each page the most affordable way it can, and only climbs higher when it has to: real text, then standard OCR, then Claude's vision for the scanned bills and checkbox forms neither one can handle. Vision costs the most to run, and it earns that cost, because a misread number hurts the client directly: the demand asks for too little, or the adjuster catches the error and the firm's credibility goes with it. Every document also gets sorted first, a billing statement, a clinical note, an intake form, so the system only asks each page what it can actually answer. Every fact keeps a citation to its source page, duplicates get merged, and the whole record is settled once, before drafting starts. Nothing downstream would catch two parts of the system disagreeing about the same fact, and one contradiction is enough for an adjuster to discount the entire demand.

Only then does drafting begin. Opus sets the blueprint: what legal rules apply, what the other side will attack, what the throughline of the case is. Then it checks those answers for consistency across the whole document, just once. Sonnet writes each section from that blueprint, following state-by-state rules and conventions written in plain English instead of code. "That only works if the model obeys the fortieth instruction, buried deep in a long document, as faithfully as the first," said Yamangil. "Claude did." Every draft still ends with a person: an attorney reads, checks, and signs it.

Another feature that totals what a case is worth on paper works off records the pipeline has already extracted and validated. On smaller cases, under 150 pages, a single Opus pass beat a longer chain, lifting accuracy on past medical expense totals by 14 points.

"Getting one case right is easy," Yamangil said. "Getting the hundred-thousandth case right at a cost a firm can actually afford is the hard part." EvenUp tracks the cost of every task, so a spike traces to its source within a day. “Cost is something you engineer,” he added. 

A month of drafting comes back in minutes

Drafting that once took 8 to 15 hours of skilled work now returns as finished drafts in about 30 minutes, a 99% decrease. Settlement offers across EvenUp’s customer firms are up 300%. At a national personal injury firm that uses EvenUp, a lead case manager said they’re "seeing demand output that is 3x over what we've produced internally."

The California firm that ran all of its demand writing through one person adopted EvenUp and cleared its 45-day backlog. Demands that took a month now take about ten minutes, and attorney review moves 75% faster. The firm says that speed hasn’t come at the cost of quality: and says it has traced its strongest settlement offers back to having detailed explanations from drafting. After filing suit, the firm now sharpens arguments into mediation briefs with AI Drafts, which helped settle a traumatic brain injury case well into six figures. In one early slip-and-fall case, the first offer came back 10x higher than expected. An Oregon firm's drafting turnaround fell from months to days, and their team reported that revenue has grown 400% in one year.

The same shift runs inside EvenUp itself, where 200+ employees use Claude Enterprise and Claude Code is spreading through engineering. Yamangil's team used Claude to build an AI cost-and-forecasting system that now steers the company's model decisions, and a set of agents whose job is debugging the other agents. “Neither project would have ever been staffed on its own,” he said. “Both exist now.”

The change reads differently from inside a firm. "I've been doing medical billing spreadsheets for over 25 years,” said a paralegal at an Ohio personal injury firm on EvenUp. “At first, I was hesitant to give up control, but with the way our firm is growing, I just don't have the time I used to. EvenUp has been a huge stress relief. It helps us move cases faster, benefits the client, and supports the whole firm."

That relief is what EvenUp wants to build on. Yamangil's advice to anyone building AI for a specialized industry hasn't changed: "Spend your expensive model where you can't verify cheaply," he  said. "The case doesn't end when the demand goes out, and neither does the opportunity to save a firm time on it.”