AI-Powered Knowledge Curation for Enterprise Teams
How I mapped where enterprise teams draw the line on AI autonomy.
Role
Sole UX Researcher
Timeline
4 weeks (from planning through insight delivery)
Methods
Moderated concept evaluation & semi-structured interviews
Participants
11 professionals across 5 enterprise companies (1000–10,000+ employees)
Impact
Findings directly shaped product development decisions within 3 weeks of delivery
💥The Problem:
Enterprise teams use an average of 5–8 tools to create and store knowledge. This fragmentation leads to:
Context switching that disrupts flow and wastes time
Duplicated content across tools, with no clear "source of truth"
Eroding trust in documentation, causing teams to fall back on synchronous communication
💭The Proposed Solution:
The product team was exploring whether AI in Confluence could solve this by automatically surfacing and organizing third-party content into a team's existing wiki structure—effectively making Confluence a "smart hub" for all organizational knowledge.
🙋🏻♀️ My task:
Inform product strategy by evaluating this new feature with users, identifying value-proposition and areas for improvement, and exploring underlying user perspectives on AI for knowledge management.
Research Questions:
Do enterprise teams experience the "scattered knowledge" problem acutely enough to change behavior?
Will the proposed solution meaningfully increase engagement with Confluence?
What specific value propositions resonate—and what concerns must be addressed before shipping?
How do users conceptualize autonomous AI, and where do they draw the line between helpful and invasive?
Approach
Participant Recruitment
Our target participants had the following traits:
Large enterprises (1000+ users)
AI features enabled
Active use of at least one third-party content tool (cloud drives, video, messaging)
Varying levels of current platform engagement (to test both adoption and retention hypotheses)
Why?
I wanted to to ensure we tested with users whose environments were complex enough to surface real curation pain points— and varied engagement levels let us distinguish adoption barriers from retention risks.
Session Design
I structured sessions as contextual inquiry → concept walkthrough → reaction to ground each participant's feedback in their actual workflows before introducing new concepts, reducing desirability bias and surfacing unmet needs we hadn't anticipated.
Each session included:
Contextual inquiry: current tool ecosystem, pain points, and workarounds
Concept walkthrough: interactive prototype of new features
Reaction & feedback: desirability, concerns, and feature requests
Why?
Grounding each participant in their actual workflows before introducing concepts reduced desirability bias and surfaced unmet needs we hadn't anticipated.
Human-in-the-loop is non-negotiable.
The most consistent finding across all sessions. Participants universally required:
A review/approve step before anything publishes
Notifications when changes are staged
Ability to opt out entirely
Manual override for anything AI-generated
Strategic implication:
Any AI-driven feature in this space must launch with review-before-publish as the default state. Skipping this to reduce friction will backfire— enterprise buyers will block rollout at the admin level if they can't guarantee governance controls to their security team.
💡Insights
We see total buy-in for knowledge aggregation, but conditional buy-in for automation.
Embedded, editable third-party content was the single most praised element: it felt native and trustworthy. AI-driven auto-organization was valued in principle but raised concerns about transparency, messy existing structures, and lack of customization.
Strategic implication:
Ship aggregation as the lead value proposition— it's a universal win that requires no behavior change. Position automation as a power layer that users graduate into, not a default they must trust on day one.
🛠️ Impact
Organizational context determines feature value.
This concept polarized participants along a single axis: how intentionally their team manages documentation today. Teams drowning in stale content wanted automation immediately; teams with dedicated knowledge owners wanted veto power. One default won't work — the feature needs tiered controls based on governance maturity.
Strategic implication:
This feature can't ship with a single default. The product needs a segmentation strategy— either detect governance maturity signals to set appropriate defaults, or offer explicit setup modes that let teams self-select their comfort level.
Product team entered active development within 3 weeks of report delivery
Findings directly informed:
dismiss controls
feedback mechanisms
daily view limits
LLM-based content summaries
content filtering for draft/confidential material
Reflection
What went well
Speed to impact. The team entered active development within three weeks of report delivery — a direct result of structuring findings around ship/don't-ship decisions rather than abstract themes. By framing insights as product requirements (e.g., "human-in-the-loop is non-negotiable"), I gave the team something they could act on immediately.
Participant diversity paid off. Recruiting across governance maturity levels (from chaotic wikis to tightly curated knowledge bases) revealed that the same feature could delight or alarm depending on context. Without that range, we'd have shipped a single default that alienated a key segment.
What was a challenge
Recruiting decision-makers vs. end users. Several participants were platform owners or admins — great for governance insights, but their reactions skewed toward risk and control. A follow-up round with individual contributors (the people who'd interact with these features daily) would strengthen confidence in the desirability signal.
Prototype fidelity gaps. The auto-organization concept was harder to evaluate because participants couldn't see their own content being organized. Reactions were partly hypothetical. A higher-fidelity prototype seeded with realistic content— or a live dogfood— would have produced sharper signal on trust and accuracy expectations.