Why Architecture Will Define the Future of IP in the AI Era
By Rohit Saluja, Head of AI, Anaqua
As organizations race to adopt AI, many IP teams are finding themselves managing a growing collection of disconnected tools, copilots, point solutions and custom applications. Each may deliver value on its own. Together, they scatter the context that AI depends on, and create new problems around governance, data quality, security, user adoption and decision-making. In an industry where a single missed deadline can mean the loss of an asset, that fragmentation carries real risk.
Most of the conversation so far has been about models. But models are becoming commoditized, and raw intelligence alone does not produce reliable work in a specialist field like IP. What separates the organizations that get value from AI is architecture: where the AI gets its context from, what it is permitted to do and how its decisions are governed and improved over time. Get that right, and every new model release makes the whole system better. Get it wrong, and you are left with tools that impress in a demo and fail in production.
Why the Model Is Not the Moat
One of the biggest misconceptions in the AI market is that success comes down to choosing the smartest model. While model capabilities continue to improve, they’re also becoming increasingly similar. In the years ahead, the thing that will separate successful IP organisations from the rest is context, not just access to powerful AI tools.
Ask a general-purpose assistant about a filing deadline and it will answer confidently. It has never seen your docket. An AI that is right 95% of the time sounds impressive. Across 10,000 deadlines, that is 500 it got wrong. In IP, accuracy that would be excellent elsewhere is a liability. And 95% is generous. Without access to the portfolio, a general-purpose assistant rarely gets close to that on IP work.
That context exists across the IP lifecycle, from dockets and family structures to fee schedules, prosecution histories, inventor discussions and years of institutional knowledge. Without access to that information, even the most advanced AI model is working with an incomplete picture.
This is why the focus on model selection often misses the point. As models become more standardized, the real value lies in how effectively AI can access, understand and act on the context surrounding IP work. Organizations that can connect AI to a complete and trusted view of their portfolio will be better positioned to generate meaningful outcomes than those simply chasing the latest model release. The question has shifted from whether a model is intelligent enough to whether it has access to the context needed to make the intelligence useful.
The Exponential Cost of Fragmentation
If context is what makes AI valuable in IP, then fragmentation is one of the biggest barriers to realizing that value.
There is no fixed number of AI tools an IP team should run. But the cost of fragmentation is not linear. Every point solution introduces another source of data, another workflow, and another version of the truth. Scattering the context that AI systems depend on.
The challenge is already evident across the wider legal industry.
A 2025 study of midsized law firms revealed 83% of legal professionals lack confidence that the information they access is fully up to date and accurate.
For IP teams, where deadlines, portfolio data, prosecution history and commercial strategy all intersect, the risks of fragmentation are even greater. Each disconnected system holds a different piece of the picture, making it harder for both humans and AI to access consistent, trusted context.
Point solutions can be very effective at tackling specific tasks such as translation, prior art search or document review. The problem is that once the task is complete, the output is often handed back to a person and the process stops there. Context is lost between systems, meaning follow on actions, workflows and decisions still need to be managed manually. As organizations adopt more AI tools, this fragmentation can create operational risks, with different systems working from different datasets, formats and versions of the truth.
The real value of AI comes when context and memory are carried across the entire workflow, enabling tasks, decisions and processes to be connected end-to-end. That's why organizations need to ensure every AI tool operates from the same data, context and governance framework.
IP Data Quality as the Foundation
The least glamorous part of any AI strategy is often the most important. Before organizations think about agents, automation or advanced reasoning, they need to think about the quality of the data those systems depend on. Unless the system is designed to detect gaps in the data, the model fills them, and in IP a confident fill is worse than a blank.
When presented with incomplete, inconsistent or inaccurate data, it will confidently produce a wrong answer.
As AI evolves from assisting users to taking actions on their behalf, the consequences of poor data quality become even greater. An inaccurate summary is one thing; an automated decision based on inaccurate portfolio data is another.
In an IP environment, where decisions depend on the accuracy of deadlines, family relationships, prosecution history and portfolio, that creates significant risk. The failure modes are familiar to anyone who has run a portfolio. Family structures that do not reconcile across jurisdictions, status codes that mean different things at different patent offices, deadlines that drift from the official registers, duplicates left behind by legacy migrations, and law firm records that disagree with the client’s own. This is why data quality is a requirement; AI can only be as reliable as the information it has access to. Without it, even the most sophisticated AI architecture will struggle to deliver consistent results.
Data quality remains one of the greatest barriers to AI success. Gartner predicted 60% of AI projects would be abandoned by 2026 due to a lack of AI-ready data.
For IP organizations, this statistic stresses the importance of creating an architecture that ensures every model, copilot and agent can access complete, trusted information.
Architecture determines how context flows through the system, while data quality determines whether that context can be trusted.
The Four Layers of AI Architecture in IP
To really understand why architecture matters, we can break AI down into four core layers. Trusted data is the foundation they all sit on. The layers are what turn that data into context an AI can act on, and keep it trustworthy over time:
- Ontology: The first layer is Ontology. It provides a shared meaning to IP data, defining how concepts such a patent families, inventors, matters and deadlines relate to one another. It is also where data quality is enforced, so the same family, status and deadline mean the same thing in every system. It ensures that humans and AI agents are working from the same understanding of the portfolio, reducing ambiguity and improving decision-making.
- Orchestration: If Ontology defines what things mean, orchestration defines what happens next. This layer embeds business rules, workflows and handoffs into the AI architecture, ensuring that insights are translated into actions. In an IP environment, this could mean routing work to the appropriate attorney, triggering deadline-related tasks, escalating exception or coordinating activity across multiple systems and teams.
- Governance: The more AI is trusted to act, the more important it becomes to define where its authority begins and ends. Governance and permissioning establish the rules that determine what an agent can read, what actions it can take and when human intervention is required. As agents gain the ability to act across enterprise systems, these controls become essential for managing risk and ensuring accountability.
- Feedback loop: The fourth and final layer is the feedback loop, which ensures the system improves over time. When a human corrects an AI-generated recommendation, updates a classification or overrides a decision, that knowledge should be captured and used to improve future outputs. Every human correction is also a data correction, so the loop is how the foundation improves rather than decays. Just as importantly, users need to see that their corrections are making a difference.
Without a feedback loop, mistakes are repeated. With one, every correction becomes an opportunity to improve both the system and the user’s confidence in it.
IP Governance in an Agentic World
As IP teams move from copilots to agentic workflows, governance stops being a policy document and becomes part of the system's architecture. It's no longer enough to define acceptable use principles. Organizations must engineer controls that determine what an agent can access, what actions it can perform, when humans must be involved, and how every action is recorded.
Historically, organizations have created identities for employees, applications and systems. Agentic AI introduces a new category of actor. If an agent can access portfolio information, generate instructions or trigger workflows, it requires an identity of its own, complete with permissions, ownership, and accountability.
With that identity comes authority, and that authority must be carefully defined. An IP agent may need access to portfolio and prosecution data, but that doesn't automatically grant permission to alter records or initiate actions. Reading information, updating data and approving decisions represent different levels of authority and should be governed accordingly.
Authority also needs clear escalation boundaries. Not every decision should be automated. Inventorship disputes, ownership changes, portfolio abandonment recommendations and spending decisions may all require human review, regardless of an agent's confidence level. In IP, the goal isn’t to eliminate human oversight, but to ensure it’s applied in the right moments.
But defining authority is only part of the challenge. As agents take on a greater role in IP workflows, organizations need to be able to explain now just what decision was made, but how it was made. Which documents were reviewed? What information was used? What actions were taken and where was the human judgement applied?
Every recommendation, approval and exception should be traceable, creating a complete audit trail that allows decisions to be reconstructed and defended long after they've been made.
This is where architecture becomes vital. Identity, permissions, escalations paths and auditability all depend on a shared foundation that spans systems, workflows, and data. When data, workflows and governance are fragmented across disconnected tools, agents lose the context needed to operate effectively and accountability becomes harder to maintain. The most successful organizations will be those that combine trusted data, connected workflows and clear governance into a single operational framework.
Without it, governance becomes fragmented across disconnected tools, making accountability increasingly difficult as AI takes more responsibility.
Turning IP Intelligence into Judgment
Intelligence will be fairly commoditized by 2030; trust in that intelligence will be what creates value. That trust is built on clean, unified and well-governed data, as well as end-to-end processes where AI is embedded rather than bolted on. The organizations that thrive will be those that reinvent their processes in an AI-native way, keeping the humans in control where autonomy cannot yet deliver.
This shift is already influencing how organizations evaluate AI. As Christof Wolpert, Vice President Global Legan Innovation for adidas, recently observed: "We get approached by lots of people saying, 'I have a tool for this, I have a tool for that.' But they're island solutions. We prefer a one-stop shop."
Christof Wolpert’s point captures a broader challenge for IP facing teams. The future will not be defined by how many AI tools an organization deploys. But by how effectively the tools operated in a shared architecture. As intelligence becomes more available, the organizations that stand out will be the ones that can transform it into informed judgement. Judgement, in this sense, is intelligence combined with the right context, clear limits on what it is permitted to do, a record of how each decision was reached, and a loop that learns every time a human corrects it.
Further reading: