Toni set out four stages of AI adoption:
Curiosity (experimenting with chatbots)
Piloting point tools – individual tools to complete a specific task
Platforms with agents that can access an organization’s records (where Legora and Harvey sit)
A connected ecosystem where every tool works from the same data.
Most organizations he speaks to are in stage 2, often running several tools at once. He described law firm partners who had signed up to three drafting tools while building their own. Some work was faster, but they were getting more volume than they could handle.
The barrier to stage 4 is the existing tool stack.
IP teams typically run between 6 and 15 separate tools and databases: IPMS, document management, patent search, trademark search, watch services and brand protection. In Toni’s words: “None of them talk to each other.”
“You’ve got a bunch of people running around on a weekly basis, monthly basis, trying to pull information, put reports together for decision support,” Toni said. “That’s the world that I’ve been working in for the past 20 years. And that has not gotten any better.”
Buying more point tools that sit on separate data adds to the problem. This echoes the view of our Head of AI, Rohit Saluja, who argued in a recent blog that the key to the success of AI adoption lies with the AI architecture - as fragmented tools scatter the context AI needs.
The intellectual property Management Solution (IPMS) holds the truth: filings, due dates, documents, inventors, owners and applicants. Toni was clear that this layer isn’t going anywhere. “You do not want to have an agentic system that is non-deterministic to guess every time what the correct data is,” he said.
The IPMS also has gaps. It records decisions, and rarely the reasoning behind them, such as why a team chose to file or renew.
Toni described three layers. The system of record holds the facts. The system of intelligence reads the documents, understands the portfolio and flags risks. The system of action is where agents act, with an IP professional always in the loop.
The order matters. “Anything you build on top of this, if this is not good, you’re going to just extrapolate and amplify all the problems,” he said. If the data is poor, fix it first. If access controls and governance are unclear, fix those first too.
Toni’s view is that most of the progress in AI over the past year has come from the harness: the memory, context, tools, guardrails and evaluations that surround a model.
“I see the LLM as the engine. The harness is actually the car, everything around it.”
Memory, context and tools are becoming manageable as more products add Model Context Protocol (MCP) – a universal standard that allows AI applications to connect to external sources seamlessly - connections. Evaluations are the hard part. Toni estimates that 95% of organizations don’t deeply understand the harness or deploy it properly. No CTO has been trained on this, and the people who can build one well are rare.
As one attendee, Manish Joshi from law firm JOSHI-IP put it during the session:
“I think it [AI] creates that skeleton, really usefully, really quickly, that would otherwise have taken me a lot longer to do… I use that to put my input, my knowledge in, to then make my end product better and more specific to the client’s needs.”
“AI changes everything apart from judgment,” Toni said. “And that judgment is actually what’s going to create your moat.” Toni made a similar case in our recent e-book, A New Way of Working: Adopting AI Into IP Processes:
“AI needs IP experts more than ever to ensure its outputs are plausible and effectively managed.”
That judgment mostly lives in people’s heads. Toni has asked many patent attorneys how they draft an application or respond to an office action, and the common answer is some version of “I just do it.” Without a clear description of what good looks like, there’s nothing to build evaluations against.
He drew on his own background. After years of playing tennis for four to six hours a day, he started coaching three- and four-year-olds on Saturday mornings and found he couldn’t explain how to hit a forehand. His advice: bring people into the organization who are good at teaching, because codifying expertise is a teaching job.
Docketing shows how much work sits around the model. The USPTO has around 650 event types, and many are close to identical. One email can carry 14 attachments relating to a patent filed in 12 countries. After three and a half years of work with LLMs, Anaqua’s docketing agents now reach more than 95% accuracy. Getting there meant codifying the events of the major patent offices and building deterministic checks around the agents.
The bigger lesson came from the build itself. For seven months, they built near-exact workflows and reached 78 to 79% accuracy. Then they built an agent, gave it two tools and two skills, and stepped back. Accuracy rose to 96%.
Toni’s conclusion: the return comes when the process changes so that people review an agent’s finished work. For docketing staff, that means moving from data entry to review, with one split screen showing what arrived, what the agent extracted, and what it will create or close.
Toni was blunt about why legal is adopting AI so quickly: clients are demanding it. On one call, a global corporation told its outside counsel in several countries that any firm not using AI should leave the call. No one left. The company then asked for 30% off the price list for the following year.
Toni’s advice to in-house teams: ask outside counsel how their AI use shows up in the bill. Then tier the work, with full attorney time on the inventions and brands that matter most and AI-first handling for routine matters. He has seen very little of that so far.
On confidentiality, he was pragmatic. Most enterprise data already sits in the cloud, and zero-data-retention APIs, such as OpenAI models through Microsoft Azure, answer most concerns. He sees control as the bigger question, and recommends sandboxing, where each agent task runs in isolation with access only to the data it needs. Any provider should be able to tell you where your data sits, who controls access, and how to reverse what an agent did.
Olivier Huc, VP, Head of Sales EMEA at Anaqua, named the two biggest bottlenecks our clients face: handling incoming documents and correspondence, and getting inventions out of R&D.
For docketing teams, the job moves from searching Outlook, finding the record, uploading the email and keying in data, to checking work that’s already been done. Correspondence and patent office documents arrive in one inbox. Before anyone opens an item, agents match it to the record, fill in fields, propose tasks to create or close, and extract prior art from office actions. Anything that can’t be matched is flagged as an exception, with a confidence score.
For inventors and in-house IP teams, the problem is inventions that are duplicated, never submitted, or take so long to process that the technology is obsolete by the time they’re filed. “R&D and IP don’t always speak to each other,” Olivier said. “Sometimes they don’t even like each other.” Smart Invention adds AI capabilities to the inventor workflow. It maps the inventor’s input to the disclosure form and searches historical inventions and patent databases, and the inventor accepts each field before submitting. The first version is due before the end of 2026, with an early adopter program opening soon.
Toni also demoed a longer-term concept, built over a weekend, where each inventor has their own agent. Given an audience idea for a wheel that reads the road surface ahead, it scanned Perplexity, Google and a patent database for 10 market signals and set up a novelty analysis within seconds. While this brings up questions around governance that need answering, it’s clear where the direction of travel is going.
For in-house IP teams, the session points to a clear order of work. Fix the data and access rules in the system of record first. Write down how your best people make decisions, so there’s something to test AI output against. Then pick one high-volume process, such as docketing, and redesign it so your team reviews finished work.