Winston Weinberg

Harvey CEO Winston Weinberg’s plateau call is consumer-only: “I think that we’re seeing a plateau in performance for consumer use cases” — and it’s a misnomer anyway, because “four was good, like we’re done”; consumers need context (calendar, app connections), not better reasoning. Enterprise keeps improving, and in code gen “the slope will only increase” — no plateau, “much better really, really fast” over the next 12 months.

Offered Anthropic at 350 and OpenAI at 800, he’d “buy them both at double” — though not quite OpenAI at 1.6 until he sees it triple down on consumer, where its brand outside tech is “so unbelievably powerful.” Enterprise will have multiple winners; consumer is OpenAI versus Google. The capability overhang is “higher than anyone is even talking about”: both labs “could stop developing things right now” and AI saturation of the economy “would still skyrocket.”

Enterprise adoption is 3-5 years from massive productivity gains — the capabilities “were there two years ago,” but the average workflow pulls data from 17 systems (“17 might be on the low end. Sometimes it’s like 50”), and the long tail of agents finishing tasks start-to-finish is brutal.

Harvey’s ARR went 7 → 55 → 190, and Weinberg’s goal for this year is “much much higher” than the ~400-500 Harry Stebbings pencils. The $8B valuation math: 20-25x projected end-of-year revenue “feels fine,” 100x “feels iffy.” He benchmarks the team against Anthropic (~$3B revenue), not legal AI — success is market pull, not execution, and “the winners and losers are going to be decided in the next couple of years.”

The GRR reckoning: investors watch net-new ARR and wave off churn, while AI app companies hire ~90% front-end engineers on vibe-coded demos without infrastructure behind them — “that’s going to be like a huge reckoning for folks once they get past a hundred million ARR.” Harvey’s counter: ~40% of its EPD org is now senior infrastructure engineers, processing ~half a billion documents last year.

“The value of B2B SaaS is about to become astronomical” — not hostages (Alex Rampell’s line) but Palantir-style ROI alignment: law firms paying ~$1M/year have won $20M M&A mandates with custom Harvey builds. Budget is already shifting out of professional-services spend “in the billions a year,” not tech budgets — and a customer paying $1M today could plausibly pay $100M.

Two deal-making rules: listen more than you speak (“a lot of people in deals, they think that movement is action… all deal making is just people reading”), and know when to not negotiate — when you understand the value of something more than everyone else, throw principled negotiation aside and get the one thing. Fundraising corollary: optimize partner, not price — small early checks with information rights, hit stated milestones, and the raise “can happen in 12 hours.”

Kingmaking is mostly myth: the vast majority of Harvey’s customers don’t know who Sequoia or a16z are, and EQT carries more brand with lawyers than Silicon Valley. Capital doesn’t win (“a hundred billion… into the wrong things, that still goes to zero”); recruiting is the one real channel, but it attracts logo-chasers who “usually don’t care that much about the mission.”

🔗 Original source & video: Harvey CEO Winston Weinberg: How to Make Mega Deals | Lessons from Rabois, Halligan & Grady

Harvey was born from a deceptively strong GPT-3 result: 86 of 100 landlord-tenant answers passed a blind “send without edits” test by three attorneys. The team had brute-forced context and early chain-of-thought prompts, then cold-emailed OpenAI’s general counsel, whose reply was, “I had no idea the models were this good at legal.” Weinberg’s real bet was the slope—better models plus better context, evaluation, and application engineering—not that GPT-3 could already one-shot complex law.

Commercially, Harvey says it has more than 250 clients and $50 million in ARR after raising more than $500 million, but its distribution insight is more important than the headline scale. Instead of starting with startups or mid-market users, it pursued A&O Sherman, PwC, and other demanding institutions, then partnered with Lexis, which was involved in the round, to combine its industry goodwill, trust, and products with Harvey’s AI. “The best way to do that is to actually go after the hardest people first”: elite design partners help define workflows, establish legitimacy, and open the rest of a conservative market.

The product strategy is to expand into narrow, high-quality workflows and then collapse them into a simple interface. Harvey is building 30–50 reusable “AI patterns”—case-law research is one example—across broad productivity tools and start-to-finish specialists, then using orchestration so an uploaded share purchase agreement can trigger seven relevant actions instead of exposing “10,000 workflows.” The aspiration is an end-to-end S-4 filing; another example assesses antitrust requirements across 72 countries.

Reliability is not one bar: general tools win by making imperfect work cheap to verify, while specialists need much higher minimum quality but are easier to evaluate recursively. For the first, Harvey shows its work in a junior-to-partner review pyramid with explanations and inline citations—“show your work.” For the second, every scoped step can be tested; this is why Weinberg calls generic benchmarks “completely useless for us” and hires experienced lawyers to design tasks and judge outputs.

Weinberg expects AI to displace legal tasks, compress apprenticeship, and split law-firm economics—not simply erase lawyers or the billable hour. Repetitive work that delays strategic exposure for five to ten years can become fixed-fee, lawyer-in-the-loop output, while scarce senior advice will stay hourly and may become more expensive—he floated, with uncertainty, perhaps 10× a junior’s rate rather than 3×. Firms can also encode their practice into Harvey, sell it as software, and profitably offer work that currently serves as a discounted or loss-leading path to major transactions.

Reasoning gains push Harvey’s workflow frontier outward, while falling inference prices let it increase quality across every user base rather than optimize for cost. Reasoning models can extend an antitrust analysis from deciding where to file toward preparing the filings; lower prices let Harvey spend more compute on quality. Because legal and tax diligence both “apply all of these rules to these documents,” the same patterns can move into tax, audit, deals, and other professional services.

The broader application-layer thesis is to target work where the “price per token” is high, then learn the real workflow from practitioners rather than ideating inside tech. Weinberg urges founders to observe industries outside Silicon Valley and predicts specialized work-completion breakthroughs in coding and medicine that feel like a new ChatGPT moment to experts. His adoption read is equally important: professionals reject abstract “Skynet” replacement, but “when folks can see and they actually use these tools, they want this.”

🔗 Original source & video: How Harvey AI is Changing the Legal Industry with Winston Weinberg