When we released the Guidewire Documentation Assistant in March 2026, our goal was straightforward: help customers and partners get direct, useful answers from complex product documentation without having to navigate across multiple pages.
Read about how we built that experience in From Search to Flying Car: The Story Behind the Guidewire Documentation Assistant.
Since then, the assistant has answered well over 160,000 questions. In user testing, complex questions that previously took several minutes to answer were resolved in under a minute, and more than 90% of surveyed users said finding information felt very easy. Documentation Assistant users also report higher satisfaction with our documentation.
But operating the assistant at scale taught us something bigger. We had built it to answer questions. Those questions started teaching us how to improve the knowledge behind the answers.
And that led us to a broader realization: the assistant is not the destination. It is one visible experience built on a larger capability—the ability to turn product knowledge into a governed, reusable foundation for AI.
From Behavior to Intent
We have always tried to be data-driven. Before the assistant, search analytics and usage data showed us where users went, while surveys told us how they felt. But neither could fully tell us what users were trying to accomplish.
The Documentation Assistant changed that. Through the questions users asked, we could see not only what information they were looking for, but the goals and context behind those searches.
Instead of seeing only behavior (clicked here, searched for that), we could now see intent. For example, a question like “How do I set up this integration?” reveals more than a topic. It can also reveal a goal, an assumed level of knowledge, and implicit expectations about prerequisites or user roles. Taken together, the questions users ask create a remarkably rich picture of the information experience.
That also changed the kinds of questions we asked ourselves:
- What was the user really trying to accomplish?
- Was the answer missing, difficult to retrieve, or difficult to understand?
- Did we have a canonical source for the information?
- Did the retrieved content contain enough context to stand on its own?
Documentation Assistant as a Diagnostic Instrument
Once we could see user questions at scale, the assistant became more than a delivery channel. It became a diagnostic instrument for the content and knowledge system behind it.
Those questions, combined with analytics, user feedback, and continuous evaluation, helped us identify patterns, distinguish systemic issues from isolated ones, and understand where problems were actually occurring.
A weak response might result from several causes: a misinterpreted intent, a retrieval failure, an incomplete source, a ranking problem, or a response-generation issue. To the user, those failures may look similar, but they require very different fixes.
Over time, a feedback loop emerged: questions revealed intent; intent exposed gaps in content, retrieval, information architecture, metadata, and experience; those gaps became priorities for improvement; and we evaluated each change against real questions and answer quality.
This has changed how we think about the relationship between content and users. Traditionally, much of the feedback loop begins after we publish. With conversational experiences, users become a more direct part of how the knowledge system evolves.
Content Quality Is AI Quality
One of the clearest lessons from that diagnosis was how often a weak answer was really a content problem.
Traditional documentation is often designed in the context of a page, a section, a navigation hierarchy, or a larger information set. Readers have access to those structures and visual cues and accumulate context as they move through the experience. AI systems work with chunks of information and don't necessarily consume information the same way.
AI can expose and amplify weaknesses in the underlying content. A missing detail, an implied assumption, or an unclear title can derail an AI response in ways a human reader might overlook. When information is fragmented, sources conflict, or crucial context is missing, the system has to do more inference. That increases the chance of incomplete or misleading responses.
For a grounded assistant, high-quality, context-rich knowledge is a prerequisite for reliable answers.
Knowledge needs to be self-contained, explicit, well structured, semantically rich, and grounded in authoritative sources.
These qualities make content better not only for AI retrieval, but also for search, maintenance, and, most importantly, people.
Turning Insights into Action
Those insights began shaping what we improved next—across content, technology, and experience.
On the content side, we’ve restructured major information sets and eliminated more than 30,000 duplicate or unnecessary topics. On the technology side, we’ve improved retrieval, filtering, and grounding. Increasingly, both are driven by the same signals: the questions users ask and the quality of the answers they receive.
That also broadens the role of content teams—from producing content to helping design and steward the knowledge system itself, defining structure, evaluating answers, and using evidence to guide priorities.
Knowledge as Infrastructure
Once we began looking at the Documentation Assistant as part of a larger learning system, a broader question emerged: does every new AI experience need its own knowledge foundation, or could parts of that foundation be shared?
That question leads us to think about knowledge less as something tied to a single channel or application, and more as infrastructure that can support many different experiences.
If every assistant, application, and workflow builds its own representation of the truth, we risk creating a new generation of knowledge silos, only this time with AI interfaces sitting on top of them.
Ideally we’d build a shared knowledge foundation that draws from multiple authoritative sources—documentation, learning, support, product, and others—while allowing the teams closest to that knowledge to retain ownership. Shared standards, metadata, relationships, and AI services could then make that knowledge easier to discover and use across experiences.
Those experiences need not look the same. A developer working in an IDE has very different needs from someone using a documentation site, getting guidance inside a product, or responding to an RFP. What they could share is the same underlying infrastructure: trusted knowledge, common standards, and capabilities that improve how that knowledge is retrieved and used.
The key benefit is compounding improvement. If a source is corrected, a relationship is clarified, or retrieval improves, that improvement does not have to remain confined to a single interface. It can benefit every experience that depends on the common underlying knowledge foundation.
The Next Experience Is Not Another Bot
It is tempting to measure AI progress by the number of assistants launched. But the more interesting question is: how many useful experiences can we enable with the same trusted knowledge?
Those experiences will continue to evolve. Some will be conversational. Some will be embedded directly into products or development environments. Others may be largely invisible, helping an intelligent system understand context or complete a task without presenting an explicit interface at all.
The common element is not the bot. It is the trusted knowledge underneath.
That changes how we think about the future of information experience. Instead of designing around a single destination where users go to find information, we need to think about how knowledge can reach people—and increasingly intelligent systems—in the context where it is needed.
Better models will certainly matter. But the next generation of AI-powered experiences will also depend on strong knowledge foundations, and on teams that know how to make that knowledge trustworthy, contextual, and ready for whatever experience comes next.