AI & Drafting
By Ian Schick, PhD, Esq.
Google’s own blind tests show AI drafts usually win. The bottleneck is no longer the technology. It is that we are still operating it by hand.
“Number, please.”
For most of the twentieth century, that is how a phone call started. A person answered and plugged a cord into a board to connect you. In 1950, about 1 in 13 working women in America was a telephone operator. Then calls just happened. The operator wasn’t a bad technology. It was a transitional one.
That is where patent practice is with AI. We have the technology. We are still operating it by hand. That was my message today at the NAPP/NCPP AI Symposium.
The buyers have moved
Two weeks ago I moderated the AIPLA Annual Meeting’s first panel of day one, “AI as a Patent Drafter: A View from the Client’s Side of the Table.”
As Law360 reported, Google patent counsel Steve Gong described how the company tests AI drafting. One person drafts an application alone, another drafts the same invention with AI, and an AI drafts it solo. A seasoned lawyer then picks the best of the three without knowing who wrote them. “AI actually usually wins,” Gong said. “Over 60% or 70% of the time, AI was the preferred version over the human version.” My own view is that the AI draft isn’t hard to spot. It’s the one with fewer mistakes.
Gong said Google is aiming to cut what it spends on outside counsel by 30 percent, which could have “serious implications for what the law firm model looks like.” Yet the reduced spending is “not really the end goal.” The goal is to “really try to rewire how things are done.”
Micheal Binns, Meta’s former associate general counsel and head of patents and trade secrets, described the pricing side: “you pay for the outcome, not for the time it took.” “The firms that leaned into enhancing their technology usage and were thinking creatively with us are the ones that we continue to give work to, and I don’t think they’ve seen a decline in the amount of work they do.” When Binns suggested that companies protecting more innovation could “remain flat on outside counsel,” Gong answered: “That’s exactly what we’re doing.” (Law360)
If you operate it, it’s transitional
The chat window. The prompt library. The sidebar you paste claims into. Each needs a person at the board, like encrypting your own email by hand before sending it. As I wrote in April, copilots are a transitional form. Prompt engineering is this decade’s typing pool.
Give the same disclosure to two desks. Desk one opens a chat window. The first draft comes fast, but the last mile runs longer than a mile: 16 to 24 attorney hours in all. The output depends on whoever wrote the prompts that day. That is prompt debt: no two drafts share a structure.
That inconsistency is not cosmetic. A reviewer flies through the third of ten applications built on one structure, but reads each of ten prompt-session drafts from scratch. Senior review bandwidth is now the scarce resource, and desk one spends it freely.
Desk two opens a queue. The system owns the workflow and brings people in where judgment is needed. The practitioner aligns on the disclosure and drafting strategy, the way a partner instructs an associate, and the inventor interview stays a human conversation. Search, drafting, figures and proofreading run on their own. In 2 to 4 attorney hours, the application is drafted, reviewed, approved and signed, with filing documents ready. (Those hour figures are estimates from my company, Paximal.)
The attorney’s judgment is the same at both desks. Desk two just stops spending it on operating the tool.
The judgment and the signature stay
None of this takes the practitioner out. As Binns put it, “I don’t think that we’re entering into a world where there will be no attorney involvement. I don’t think that’s plausible.” (Law360) A workflow moves the practitioner’s time to where judgment lives: the gates, not the keyboard.
The signature is never automated. But be clear about what it certifies. Not that I wrote every word. It certifies that the statements are believed true after an inquiry reasonable under the circumstances (37 C.F.R. § 11.18(b)). As I wrote in September, reasonable inquiry was never measured in hours. Budget, timeframe and available tools all shape what is reasonable, and that target moves as practice does.
Which desk is yours?
The median complex application, $12,000 in AIPLA’s 2024 survey, was priced on production cost: about 26 hours of a highly trained person’s time, most of it typing. We are not content generators anymore. Content is close to free. The first draft is about to stop being a billable event.
The telephone didn’t win when operators got faster. It won when calls just happened. The typing leaves. The judgment and the signature stay.
Which desk is yours?
Sources: Ryan Davis, “Tech In-House Attys Say AI Rivals Humans At Writing Patents,” Law360, Oct. 29, 2026 (link above). Operator figures: Price, Econ Focus, Federal Reserve Bank of Richmond (2019). Hours: Paximal.