Austin Walters sells software to in-house patent teams, and he doesn’t tell them what they want to hear. He estimates they’re missing 95% of their patentable ideas. He says one IP attorney for 10,000 engineers will never catch them. And he predicts companies will end up with in-house counsel or outside counsel, but not both.
Clause 8 is brought to you by:
IP Copilot - Discover, manage, and protect your IP.
Junior - Your legal team’s documents, drafted in your team’s style.
His record gives him standing to say so. Walters spent about two years at one of the world’s largest patent filers before anyone told him the company had a patent program. Once he found out, his name went on hundreds of patents, most of them about training neural networks. Before transformers arrived, his team built what he calls a large language model (LLM), one that could synthetically replicate any data set. And around 2018, years before ChatGPT, he wrote a blog post with a two-word verdict: “it’s over.”
The episode also features a friendly debate between Walters and Clause 8 host Eli Mazour over where AI patent drafting is headed. Walters says general-purpose models are already “pretty good at making a patent,” and that an application his own system drafted in about an hour was good enough for a provisional. Clause 8 host Eli Mazour isn’t convinced. “I’ve had a somewhat different experience, at least for now,” Mazour says. If the goal is a patent that is valuable and worthwhile, he argues, “somebody still has to make that decision.”
Walters’s predictions, and when we’ll know
2027
About half of practitioners use AI to draft, up from about 20% today
Inside large companies, AI invents as much as humans
2028
AI drafting reaches about 70% of practitioners
2030
Drafting becomes one click, and a $30,000 patent costs about $200
The USPTO moves to an effectively registration-style process measured in months
Companies keep in-house counsel or outside counsel, not both
Within about 10 years
Smaller, cheaper models replace today’s giant ones
Walters and Mazour agreed to compare notes in 2030.
In the episode
Two years at a major patent filer before anyone told him there was a patent program
Building a large language model before transformers, and the 2018 blog post that said “it’s over”
The AI agent that tried to break into its own company’s website to run a prior art search
Why LLMs are “the least efficient way to solve a problem”
Why most companies miss 95% of their patentable ideas
Flipping the invention disclosure: confirm a summary instead of filling out a form
AI scoring of invention disclosures landing within 5% of human reviewers
Why building your own AI tool rarely works out
What IP teams actually care about: not looking bad to the boss, and getting home on time
🎧 Listen to the full episode above or on your favorite podcast app, and subscribe to the Clause 8 YouTube channel for bonus content.
