AI File Search: Redefining How HubSpot Customers Find Assets
Helping customers find files by description, not just by name, led to a 22.5% increase in successful File insertion through Search.
Role
Lead Product Designer
Timeline
3 months
Team
Product, Engineering lead
Before and After
The Situation
The File Picker (used across multiple tools like CMS, CRM, and Email) relied on literal name-based search. The workaround users did was simply to upload the same file multiple times instead of trying to find the original. This behavior created tens of thousands of duplicates, fragmented storage, and wasted time across teams.
The opportunity was clear: make finding assets feel intelligent, not procedural.
File picker audit
What I discovered
After auditing data, reviewing competitors, and mapping the flow, three signals stood out:
Search wasn’t failing users; the system was. They typed what they remembered (“blue logo”), but the picker required exact strings.
AI could bridge the gap between human memory and file metadata.
Duplicates were diagnostic, not just messy. They exposed missing logic in how files were indexed, named, and surfaced.
This work wasn’t about fixing the UI: it was about rethinking how HubSpot understands and retrieves media.
Strategy & Decisions
I partnered with Product and Engineering leads to balance quick impact with systematic improvements.
1. Introduced AI-Generated Metadata
→ Automatically created file descriptions and tags at upload, enabling search by description or keyword instead of exact name.
2. Added Editable Descriptions
→ Allowed users to modify AI tags, giving them control as models matured.
→ Also gave lower tiers without AI features, manual tagging tools for parity and inclusivity.
3. Removed Flow Dead Ends
→ Redesigned file picker flow to prevent dead ends and overcomplicated rabbit holes.
4. Deferred Heavy Redesigns
→ Focused on scalable improvements rather than one-off UI polish, knowing a global redesign was imminent.
Results
Customer: Users can now find files by concept or description, saving time and frustration.
Platform: Created the metadata layer that future AI tools (e.g., Image Gen, Copilot) can leverage.
Business: Reduced duplicate uploads and storage redundancy, a hidden but significant operational cost.
Key Learnings
Education is part of the UX: Teaching customers how to search by description was as impactful as the feature itself. A new capability requires behavioral onboarding.
Design shapes AI outcomes: Prompting logic influenced how images were described and tagged -> directly impacting user experience.
Investigate before you fix: Duplicate uploads revealed systemic issues (e.g., auto-upload logic in Quotes). Solving those unlocked value beyond the surface layer.
Solve for future cases: Allow customers to find duplicates as soon as they upload one.
