SVTR AI Venture Database · AI Venture Salon #005 | AI + Legal: What AI can swallow today is the 70% boilerplate of a patent spec; the moat is software plus lawyer service · December 22, 2024
At this SVTR members' roundtable on December 22, 2024, Silicon Valley founder Luo Tao put a set of prices on the table: filing a single U.S. patent costs $20,000 to $25,000 on average, while the U.S. Patent and Trademark Office (USPTO) charges only a few hundred dollars in official fees, as little as $300 for a micro entity. Nearly all of that $20,000 gap is legal fees. Luo previously built a machine-learning platform at Uber and, before that, spent about a decade on chip placement-and-routing optimization in semiconductors. His co-founder is a patent agent with twenty years of experience who has personally handled more than two thousand filings. What the two are after is whether AI can erase most of that $20,000 spread.
What can be standardized is the 70% spec, not the 30% claims
Luo splits a patent into two parts. Per his account on-site, the bulk of the page count is the detailed description: invention background, industry context, embodiments, figure captions. This runs roughly 70% of the document, is highly formulaic (he called it "eight-legged essay" boilerplate), follows fixed patterns, is barely read by examiners, yet eats up a lawyer's billable hours. The hard part is the claims: only a couple dozen of them, but they define the scope of protection. Draft them too broad and they collide with prior art and get rejected; too narrow and a competitor tweaks one detail to design around them. Per his account, 90% of U.S. patents are rejected on the first pass, and the judgment in that back-and-forth is exactly where a lawyer's experience earns its keep.
His call is blunt: what this generation of AI can already swallow is that 70% boilerplate spec, cutting its cost in half; the 30% that is claims, like the hardest part of writing code, cannot be replaced near-term and still needs a lawyer. The demo followed that logic: auto-expanding a one-line invention description into a full spec, integrating prior-art search, feeding retrieved patents into a retrieval-augmented database to assist drafting, and generating first drafts for non-professional inventors. He noted they are rebuilding the lawyer-facing piece into something like Cursor: lawyers told him they do not want "fully automatic generation," only guidance alongside their own writing, with the tone and style close enough to their own that, in his words, "no one can tell it wasn't written by them."
Pure software sold to lawyers is a narrow road
By the end of 2024 the field was no longer empty. Per his account, YC had backed two players last year, Edge in Silicon Valley and Solve Intelligence out of the UK, plus a French company, DeepIP, advertising heavily. Luo's read is that these firms are almost all subscription software sold to lawyers, without prior-art search, figures, or full-lifecycle patent management. He deliberately took a different road: not just software, but a layer of human service through his co-founder's partner patent firm, with the primary customer being not lawyers but price-sensitive engineers and startups that genuinely need patent protection. Self-serve drafting via the software runs a hundred or two hundred dollars; if only the claims need a lawyer's eye, there is no need to hand over the whole filing for $20,000, just book an hour at around a thousand dollars to fix that one section.
The real resistance in the legal profession, and where it is loosening
What investors cared about most on-site was whether this is genuine demand or something that has to be coaxed along by education. Luo's answer carried a chill from the front line: legal AI has had a rough couple of years. Even a leader like Harvey, when raising, could feel capital's appetite for the legal vertical slowing; retention was a problem, with some users cancelling after a while. The profession itself is deeply conservative. Some small firms still manage a patent's twenty-year lifecycle in Excel; some keep their computers off the internet, wary of errors, wary of leaks, and unwilling to learn new tools.
But the loosening is real. Per his account, more lawyers are trying ChatGPT and feeling threatened by it. The more direct push comes from the client side: patent firms are being pressed by their own clients to cut prices, "AI can write this much now, so why are you still charging me $20,000?" That one line says more about the temperature of this field than any funding headline.
Who this session drew in was itself a signal. Attendees included patent-industry practitioners and a seasoned Silicon Valley early-stage angel, an organizer who has run more than three hundred inventor events across China, and a batch of engineers who described themselves as "having a decent idea but unable to afford the $20,000 in legal fees." The people who gathered around a topic as narrow as patent drafting are precisely the truest demand side of this business.
Same problem, a stiffer answer from China's patent law
The session also surfaced the gap between the two patent systems. Per his account, the U.S. is more permissive when amending claims after a rejection: you can pull the gist from the spec into the claims without matching the wording exactly, and you can merge or trim claims. China requires the original text be carried over verbatim, with no free interpretation. On timeline, a U.S. patent often takes one to three years; China is faster and cheaper, but more than 90% of the examination and response flow is common to both, often just a matter of translating between Chinese and English. The Chinese inventor-community organizer picked it up on the spot: if that 70% spec can be spared, you save 70% of the cost, and handing the remaining 30% to firms on both sides is a split where the software vendor, the firm, and the client all win.
Seen through the SVTR AI Venture Database, the judgment sharpens. The database counts more than 40 funded AI + legal companies worldwide and tracks about 108 legal-sector companies; but across the 17 AI application sectors, legal ranks 13th by number of rounds (42), and 14th by both total scale ($840M) and average check (about $20.03M). It is a mid-to-lower-resourced sector with modest checks. The thin resources are exactly what forces Luo's approach: chew through the standardizable 70% first, backfill the 30% the model cannot handle with human service, rather than opening with a promise to replace lawyers.
Looking back at this late-2024 roundtable about nineteen months on, several threads can be checked. Legal AI overall did not stay cold the way the room feared: Harvey raised at a very high valuation, and DeepIP and Solve Intelligence took on new rounds, with the sector warming back up, a layer that diverges from Luo's cautious read at the time. But the bet he really made holds up better: what can be swallowed at scale is still the templated spec, while claim work of this intellectual weight still needs a human lawyer to backstop it. "Software plus service," not "pure software replacing lawyers," remains the more defensible way to play this patent niche. That is the point of SVTR keeping a first-hand field record: preserve a sector call before it is validated, so it can be checked line by line later.
The full recording of this session is archived in the SVTR members' area. For more first-hand field judgment, deal information, and direct access to founders, join SVTR membership.