In tangible chunks: a clearer path to patent intelligence automation

In the last months, I've spoken at various industry events about patent intelligence automation. After some self-reflection, I've learned that I need to recalibrate how I talk about the topic and patent AI in general.

At an IP law firm event, with almost 100 patent attorneys and corporate IP managers present, I asked how many of them had used ChatGPT Deep Research or other similar agentic generative AI products. I saw only one or two hands raised in the audience, even though such products had been there for six months or so at that point. Discussions at drinks revealed that many were not using even simpler AI products.

Six months feels like an eternity when you are deep in the AI space yourself and when your job is to adopt new technologies fast in the product you are building, and in your own work, too. Make no mistake, I don't blame the audience for not having used them. I just realized that I'm in a bubble and need to take that better into account.

On the other hand, at a few other events, even those focused on IP information, I've come to realize that there is a lot of confusion about AI. It varies from healthy skepticism to full swing overuse of Large Language Models (LLMs) for jobs that are not within their capabilities. Even many software vendors stir it either with overpromises or problem-centric approaches, sometimes even false information. I can understand someone new to it may be confused.

Little first-hand experience combined with incoherent information does not help in building trust in and adopting new technologies. 

At the end of the day, getting the job done is what matters. To get there, there needs to be both education on AI fundamentals, as well as transparent solutions whose value and readiness level in action can be easily tested.

Two archetypes of AI for patent intelligence

For advanced AI-based patent searching and analysis, one needs two types of AIs: document retrieval AI and content analysis AI. The former finds relevant sources of information, the latter processes it into digestible form. In automated systems, these are done linearly or in a more iterative or agentic manner, but both parts are still present.

The dominant technology for retrieval AI are so-called embeddings. Basically converting texts, images or sound into long lists of numbers, i.e. vectors using an AI model trained for this purpose. Similar content results in close-by vectors. Vectors are mathematically easy to compare. Embeddings are in a way fingerprints of the original long documents. Good embeddings tailored for the patent domain lead to accurate patent search results. At IPRally, we have a whole research team focused on this.

LLMs have shown their power in content analysis, and the progress in the last year has been rapid. When used for information extraction from individual documents, the risk of hallucinations is also low. But each model has its own character and performance, and complex and nuanced IP tasks, like analyzing infringement, pose a significant challenge even for the best models. Also high-level system prompts and specific task prompts matter a lot. Patent domain specific guardrails and evaluations are necessary in the long term to ensure and improve quality – also an area we are exploring.

Thus, the remaining key questions are: What information does the retrieval AI have access to? What is its hit rate and noise rate? How reliable is the LLM for the task in question? 

While there is still a lot of room for improvement in all areas, at IPRally, we see excellent progress over time with retrieval AI, meaning that the risk of missing relevant documents is going down. Regarding content analysis, we see over 1 million patents being analyzed in-depth each month, with our Ask AI and Smart filters features. The users keep coming back, which is the best proof of value. 

Why a solution-focused approach matters

While almost no one disputes that AI will play a big role in IP in the future, one still hears a lot of talk about the dangers and risks of AI, without showing solutions on how to overcome those. When the above split between retrieval and analysis is understood, it is easier to address risks, envision solutions and form an idea of whether a solution would be fit for a purpose.

At IPRally, we've decided to offer new AI in digestible chunks, providing so-called AI Assistants around our retrieval AI, but not mixing the technologies in a confusing way. This allows users to test not only the search part, but also the AI Assistants, in isolation and build trust in future full automation that leverages the same technologies. As an example, a search flow can look like this:

1. Use IPRally's Refine function to instantly transform diverse invention materials into claim language that is optimal for the retrieval AI to digest, like in this example. 
2. With this refined input, carry out highly accurate prior art searches with built-in technical understanding.
3. Create meaningful prior art review prompts, dynamically adapted to the specific search and use case, with IPRally's Prompt assistant.
4. Use Ask AI analysis and Smart filters to narrow down the results to just a few of the highest relevance, and use conversational AI reasoning to double-check the analysis.

This way, the paradigm shifts from a vague "is AI good enough for X?" to evaluate each part separately, and more tangible evaluations "by improving this part of the pipeline, I would be able to do X with a confidence level Y". This is important both for us as developers, and also our users taking their steps in the AI journey.

As the pipeline gets more mature, proving the value of AI gets easier. Complexity gives way to clarity, and trust grows, moving us closer to the reality of patent intelligence automation.

More from Sakari Arvela

On the path to automated patent workflows
‍
Two years Into the Generative AI boom: how patent work is evolving
‍
Patent work automation – just a dream? 

‍