What Happens When Both Sides of a Case Use AI?

Summary

• AI could make previously uneconomic plaintiff cases worth taking. Eve CEO Jay Madheswaran says lawyers generally need at least $5,000 in fees to accept a case, while 96% of labor-and-employment cases get rejected. Automating intake, qualification, and legal work could unlock that demand; Alex Rampell’s framing is “revenue expansion,” because firms can represent more people.
• One Los Angeles firm increased capacity from about 100 to 500 clients per attorney. Madheswaran credits intensive adoption across intake, handoffs, and settlement preparation. Rampell calls this a “shadow labor market”: the opportunity extends beyond historical legal-software spending into work that previously required additional employees.
• The defensibility argument rests on workflow context and legal information absent from public datasets. Madheswaran says fewer than 1% of cases produce opinions, leaving settlement outcomes and effective tactics largely inaccessible to general models. Even Albert Einstein dropped into a firm “will not be productive for quite some time”; Eve supplies context while, he says, never training on privileged client-case data.
• Small firms could become much larger businesses through specialization and consolidation. On a $25,000 settlement, Madheswaran estimates roughly $8,000 in attorney fees, with half consumed by labor and 30% by marketing. He expects proprietary datasets to help firms look “more like software companies,” expanding narrow expertise nationally or integrating acquisitions faster.
• Outcome pricing fits contingency law, but broader adoption requires predictable delivery costs. Eve shifted away from its initial “tall, grande, venti” subscription approach after learning that customers valued additional clients and outcomes. Madheswaran expects savings to fund more compute and legal work, while preserving a hedge on hourly billing: “I don’t think it’s going to die completely.”
• Stronger plaintiff firms could change defense tactics and reshape how disputes are resolved. One customer answered holiday-timed discovery requests the same day using an Eve agent, defeating a workload-pressure tactic. Madheswaran expects routine exchanges through “agents talking to agents” and, in some areas, automated dispute resolution that assists human judgment.
• Fully autonomous case resolution remains a development target. Asked whether Eve has completed an entire case without intervention, Madheswaran answers, “Not yet, but we’re close.” Cases involve “probably 10,000 back and forths”; engineers are the constraint, and his claim of hundreds of billions of dollars in vehicle-accident value depends on delivering that end-to-end capability.

Deep dive

1. Labor costs exclude clients even when lawyers charge on success

• Madheswaran’s starting point is “legal AI for real people”: contingency firms collect only when clients win, making representation possible without an open-ended $500–$600 hourly bill. More legal work can improve outcomes, but someone must finance that work. Eve aims to cover the path from the first call to settlement, “giving David a slingshot to take down Goliath.”
• Rampell’s market-failure example is a client with $1,000 of damages: a lawyer taking one-third cannot cover the necessary paralegal and attorney work. Contingency removes the client’s upfront bill without making the case economical. As Rampell puts it, “even if I win, I lose,” because carrying costs can exceed the potential return to the firm.
• Madheswaran says lawyers do not take cases yielding less than $5,000 in fees; even at that level, many lose money, with a few big wins carrying the portfolio. His example is a fender bender below the deductible: even if that deductible is $3,000–$4,000, “no lawyer in America will touch that case.” For an individual, he notes, that can mean a lost paycheck.
• In employment law, he describes a 30–45-minute intake call followed by roughly 10 hours researching whether alleged misconduct fits a cause of action. Only then might counsel discover the case cannot produce $5,000; he says 96% of labor-and-employment cases are rejected. Eve answers calls, qualifies prospects against the firm’s criteria, and can begin a demand or preliminary work immediately, changing that screening equation.

2. Revenue expands when the whole workflow gains capacity

• Rampell’s central distinction: “if I use Eve, I make a lot more money for my firm,” because it can accept more cases. “It’s actually revenue expansion.” Madheswaran describes bottlenecks moving from a solo attorney returning calls to an intake team overwhelmed by successful marketing; automating one task is insufficient if work simply piles up at the next handoff.
• One Los Angeles firm increased its load from about 100 to 500 clients per attorney, Madheswaran says. He credits the firm with working hard and using Eve “up to the max”: intake, qualification, agent-managed handoffs, and mediation preparation, including settlement outcomes and previously successful tactics. He pictures the lawyer as “almost a data scientist and engineer,” improving a continuously advancing process.
• Historically, he says, a plaintiff practice needed little more than a cellphone and email; some solos avoided even Microsoft Word. Software had not solved the labor constraint, while institutional defense buyers attracted development budgets. Rampell calls the opportunity a “shadow labor market”: outcomes run into hundreds of billions of dollars, so historical software spending understates the addressable work. His 12-cent-an-hour paralegal is a hypothetical illustrating labor AI could supply.

3. Private case knowledge makes general intelligence useful

• Rampell’s application-layer challenge is direct: why won’t ChatGPT, Claude, or their developers absorb this business? Madheswaran welcomes better models, but “if you drop Albert Einstein into a law firm today, he will not be productive for quite some time.” Intelligence still needs the firm’s working context and the information required to act effectively.
• Madheswaran says less than 1% of cases actually settle and get turned into case opinions; most settle before litigation, leaving essential information outside scrapable public data. Rampell makes it concrete: knowing GEICO actually paid $29,000 on an injury claim matters. An insurer dealing with 100,000 firms sees patterns fragmented plaintiff practices cannot. Rampell points to insurers’ Colossus software as an existing information advantage.
• For Madheswaran, the missing knowledge is practical: an effective demand against State Farm for soft-tissue injuries, the local courts, and opposing counsel’s behavior. His organizing question is “what is the best thing I can do at this particular moment?” That decision repeats throughout a case; Eve is meant to make the context usable at every step.
• Rampell presses on sharing data across separate firms. Madheswaran distinguishes privileged cases—“we don’t train on that data ever”—from tactics attorneys already exchange through continuing legal education. Eve hires attorneys and tracks case success; its Auditor reviews every page of every case overnight for issues an ideal attorney would catch. He says “almost every single case has missed information,” explaining the value of comprehensive review.

4. Stronger plaintiffs change tactics and expose court bottlenecks

• Rampell’s pushback is that insurers and defense firms will acquire AI “superpowers” too. Madheswaran accepts a world where 99%, maybe 99.9%, of representation work is AI, while retaining a human difference in client understanding, local relationships, and advocacy. Institutions still have money and time as weapons. He thinks routine exchanges will become “agents talking to agents”; some Eve customers already spend more time preparing for trial.
• One observed change is tactical: a defense firm repeatedly sent discovery just before Thanksgiving or another major holiday, expecting the plaintiff team to rush or sacrifice vacation. The plaintiff firm created an Eve agent that answered the same day. Madheswaran’s reading of the defense reaction: “this is what happens when the other side has resources.” He says defense still largely bills hourly and has not matched plaintiffs’ pace.
• Rampell’s court-capacity concern comes from his own Wells Fargo dispute: after four hours of frustration, he told Codex/ChatGPT, “Sue Wells Fargo.” He says four hours later it had filed the suit and found a process server, but the court date was in February. Lowering the cost of initiating legal work leaves adjudication as a bottleneck.
• Madheswaran thinks it is “inevitable” that more disputes move outside courts, extending existing channels such as veterans’ committees, the EEOC, mediation, and arbitration. In “some pockets,” he expects “ADR, automated dispute resolution,” using AI-assisted neutral parties to assemble historical facts and accelerate decisions. He does not think judgment itself will be replaced; people must still address biases that might exist in those comparisons.

5. Software economics could support specialists and rollups

• Rampell asks whether a firm handling 100 cases per lawyer, then 500, perhaps eventually 5,000, could become a billion-dollar business. Madheswaran thinks firms will look “more like software companies.” On a $25,000 settlement, he estimates roughly $8,000 for the attorney: about half goes to labor, 30% to marketing, and the remainder may be profit. That structure makes reinvestment difficult when improvement means hiring a better attorney.
• The prospective change is a proprietary dataset that improves results in a narrow field. Madheswaran expects specialization to deepen from employment or intellectual property into soft-tissue injuries with particular fact patterns. Targeted advertising might then beat Morgan & Morgan’s broad billboards. Rampell draws out the expansion path: get good locally, then take that competence nationwide; Madheswaran agrees.
• The opposite path is consolidation. Madheswaran says customers on the scale of Morgan & Morgan are buying other firms because Eve can shorten integration around the client’s desired legal outcome. By handling operational and workflow changes, the platform reduces painful reintegration and can accelerate the point at which an acquisition becomes worthwhile.

6. Outcome pricing requires knowing what delivery will cost

• Rampell notes that contingency firms already sell the outcomes SaaS vendors are now trying to price. Madheswaran admits Eve initially copied the “tall, grande, venti” subscription model: “we didn’t know any better.” Customers cared about serving more clients, forcing an early shift toward outcome-based pricing, roughly a year and a half before the conversation.
• Rampell’s defense of hourly billing is information asymmetry: clients may not articulate their needs, and scope can change. Madheswaran agrees that a new case might require 10 or 100 hours before anyone knows. His broader requirement is to infer, from the facts, how much compute an outcome needs—through better data and better execution—so legal work can be priced and delivered more predictably.
• Madheswaran’s hedge matters: “I don’t think it’s going to die completely,” he says of the billable hour. Some legal work reveals its value only years later. Elsewhere, predictable outcomes permit amortized costs, and labor savings can fund more tokens and more useful work. The intended cycle is better outcomes and more revenue, with savings reinvested in serving clients.

7. End-to-end autonomy remains an engineering problem

• Asked what limits growth, Madheswaran says benchmarks grading whether one task went “mostly okay” miss the problem. Eve needs reliable outcomes across long-horizon workflows, built separately for each case type. There is still “a decent amount of human touch needed per case,” although he says that is falling month over month. The main constraint is engineers.
• Rampell asks the essential test: has Eve autonomously answered the phone, onboarded a client, and obtained the insurer’s check end to end? “Not yet, but we’re close.” Madheswaran estimates a single case involves “probably 10,000 back and forths.” Earlier, he describes matters lasting six months to a year, making reliability over extended sequences central.
• His concrete example is medical-record collection: outbound calls to insurers and healthcare providers take two weeks today, he says. Eve’s voice agents make those calls, sometimes encountering other agents and learning to get past them. Record collection, document creation, and opposing-party conversations are being automated, but some human intervention remains; early versions of the broader workflow are live with customers.
• Madheswaran says completing vehicle-accident cases end to end would unlock value in the “hundreds of billions of dollars,” conditional on delivering that capability. He says Eve has helped “a couple hundred thousand clients” so far, against an aspiration to reach hundreds of millions.