Dobby: Enterprise AI Knowledge Agent

Dobby: Enterprise AI Knowledge Agent

Dobby: Enterprise AI Knowledge Agent

Private, cited answers on-prem. No code. Governed publishing. First sourced answer in under 10 minutes (target).

TL;DR

TL;DR

“I led end-to-end design for Dobby’s on-prem AI knowledge agent. Non-technical owners stalled at setup, so I made presets the default and moved approvals into a focused lane with audit + rollback. Targets: 60% no-code build completion, 70% publish in 30 days, and a first cited answer in <10 minutes.” 
“I led end-to-end design for Dobby’s on-prem AI knowledge agent. Non-technical owners stalled at setup, so I made presets the default and moved approvals into a focused lane with audit + rollback. Targets: 60% no-code build completion, 70% publish in 30 days, and a first cited answer in <10 minutes.” 

Intro

Intro

Dobby, a Taiwan startup building an on-prem enterprise AI knowledge platform. With a small crew and paid users queued up, we ship fast—using AI, building AI—to deliver a no-code, preset-first experience with cited answers and approvals/audit for safe rollout.
Dobby, a Taiwan startup building an on-prem enterprise AI knowledge platform. With a small crew and paid users queued up, we ship fast—using AI, building AI—to deliver a no-code, preset-first experience with cited answers and approvals/audit for safe rollout.

Industry

B2B SaaS
B2B SaaS

Timeline

2025 Jun - Now
2025 Jun - Now

Team Size

1 Product Designer • 1 Product Manager • 3 Developers

Responsibility

End-to-end design
End-to-end design

Team Size

1 Product Designer • 1 Product Manager • 3 Developers

Problem

Non-technical admins wanted an AI assistant but hit a wall at setup—too many knobs, jargon they didn’t understand, and fear of “breaking something.” Enterprises won’t ship sensitive docs to public clouds, yet teams still need fast, trustworthy answers.

Design

1

1

1

Discovery

Short interviews and competitor analysis revealed two blockers: users either had to wait too long dig through files at company to find the answer they need, or the tool wasn't approachable if they didn't know how to code.

Short interviews and competitor analysis revealed two blockers: users either had to wait too long dig through files at company to find the answer they need, or the tool wasn't approachable if they didn't know how to code.

Short interviews and competitor analysis revealed two blockers: users either had to wait too long dig through files at company to find the answer they need, or the tool wasn't approachable if they didn't know how to code.

2

2

2

Synthesis

Research consolidated into user journey and affinity map converged on three needs: fast first answer, transparent approvals, plain language defaults.

Research consolidated into user journey and affinity map converged on three needs: fast first answer, transparent approvals, plain language defaults.

3

3

3

Decision

We cut the multi-step wizard to a single “first answer” screen, and split approval into its own focused panel for clarity and speed.

We cut the multi-step wizard to a single “first answer” screen, and split approval into its own focused panel for clarity and speed.

We cut the multi-step wizard to a single “first answer” screen, and split approval into its own focused panel for clarity and speed.

4

4

4

Delivery

One of the core goals was speedy delivery. To match the team's rapid shipping cadence, I used a hybrid AI-assisted workflow. (ChatGPT & Dovetail for research, Claude & Figma Make for ideation and prototyping)

One of the core goals was speedy delivery. To match the team's rapid shipping cadence, I used a hybrid AI-assisted workflow. (ChatGPT & Dovetail for research, Claude & Figma Make for ideation and prototyping)

Competitor Analysis

There is a gap in no-code workflow builder in the market to provide non-technical teams to chain actions and approvals more efficiently.

Read in full

Read in full

Read in full

User Interview

Five interviewees from regulated enterprises and Dobby's potential users. We learned that:

“Setting up AI chatbot in general has a complicated flows. I would prefer to just have preset like limited options/modes to choose from.” - Compliance & Risk Officer

“Setting up AI chatbot in general has a complicated flows. I would prefer to just have preset like limited options/modes to choose from.” - Compliance & Risk Officer

“Setting up AI chatbot in general has a complicated flows. I would prefer to just have preset like limited options/modes to choose from.” - Compliance & Risk Officer

Read in full

Read in full

Read in full

User Journey Map

Learning: ease the building step for non-technical users. Make publishing predictable with clear approvals. Default to on-prem with citations; keep collaboration safe via simple roles (admin/editor/view-only) and scoped visibility so non-admins only see their own chats.

Read in full

Read in full

Read in full

User Flow

I simplified the build flow with presets For publishing, I added change summaries, diffs, and a sandbox to test before release—all tracked in an audit trail. Role-based access kept it clean: Admins publish, Editors prep, Viewers chat. Inline flagging turned user issues into improvement requests.

Read in full

Read in full

Read in full

How Might We

HMW let non-technical users set up a secure, on-prem AI assistant in minutes using a guided flow?

Solution

Internal tool that surfaces reliability and trust

Upload, pick a preset, see ETA, ask, and get a sourced answer on one surface. Users said trust = seeing sources, not just confidence scores.

Upload, pick a preset, see ETA, ask, and get a sourced answer on one surface. Users said trust = seeing sources, not just confidence scores.

Upload, pick a preset, see ETA, ask, and get a sourced answer on one surface. Users said trust = seeing sources, not just confidence scores.

Approval you can skim and ship

Request → Review → Approve → Publish, with audit and rollback. Clear, reversible steps turn a demo into something leaders are willing to roll out.

Request → Review → Approve → Publish, with audit and rollback. Clear, reversible steps turn a demo into something leaders are willing to roll out.

Request → Review → Approve → Publish, with audit and rollback. Clear, reversible steps turn a demo into something leaders are willing to roll out.

Roles & scoped visibility

To get to market faster, we made the strategic decision to phase the rollout of this feature.

  • Phase 1 (V1 Launch): Will ship with a simplified model [e.g., all users are 'Admins'] to unblock our core 'build and publish' workflow.

  • Phase 2 (Fast-Follow): Will include the full, enterprise-ready permissions model I designed (Admin, Editor, Viewer), which is now a top priority on our post-launch roadmap."


Design System

I built a modular, enterprise-ready design system so the team could scale the UI while iterating fast toward launch: Reusable governance components shortened design/engineering cycles and kept weekly releases on track for the official launch. Preset-first patterns, standardized upload, and breadcrumbs let us evolve flows quickly without breaking consistency.

See Design System

See Design System

See Design System

Target Metrics

Current design is roadmap-aligned and includes capabilities slated for upcoming releases. Expect UI polish and naming adjustments.

No-code build completion

Target: ≥60% of first-time, non-technical owners create an assistant and send a first test prompt without touching manual settings, in a single session.

Admin approval for team use

≥70% of active workspaces complete at least one admin-approved publish within their first 30 days.

Time to first cited answer

≥50% of new workspaces see a source-backed answer in under 10 minutes.

Status & Up Next

Current

The Dobby AI agent is currently in its final development phase and is on track for a scheduled launch with our first enterprise customers in Q4 2025.

Measuring Our Targets

While we do not have post-launch quantitative data yet, my immediate priority upon launch will be to partner with the Product Manager to validate our design decisions against the target metrics