Company & team

International Business Machines (IBM): w3 Publisher

domain

Technology

Scope

User research
Prototyping
Product design
Project management

Platform

Desktop

Context:

Understaffing necessitated a trade-off.

Despite multiple years of growth, Publisher, IBM’s proprietary site builder, operated with a relatively small team, creating ongoing tension between product development, user support and documentation maintenance. To protect development velocity, the team centralized support through dedicated Slack channels managed by rotating on-call developers, allowing the rest of the team to focus on business-critical initiatives at the cost of slower documentation updates.

Problem:

The trade-off created as many problems as it solved.

Over time, the w3 Publisher support channel replaced the documentation as the primary source of self-service, increasing strain to on-call developers and creating an unsustainable support model where development hours were increasingly spent on repetitive support rather than product advancements.

Research:

Because we couldn’t increase headcount, the impact was focused downstream.

From interviews with on-call developers, we confirmed understaffing as their primary pain point but because additional team-members was outside the project scope, we had shift our focus. It was interviewing the other side that painted the full picture.

“It was interviewing the other side that painted the full picture.”

Inquirer interviews revealed the intertwining journey between user groups. This casual loop illustrated how the pain points of both parties exacerbated one another and what the strongest intervention point was: unreliable documentation.

Opportunity:

Leveraging automation could accelerate the discovery and maintenance of documentation.

The opportunity was to design an automated workflow that could leverage existing artifacts to answer support inquiries, help maintain documentation and, ultimately, preserve developer hours for progressing Publisher’s roadmap. The goals realized from this were:

Short-term:

• Pilot a proof of concept to determine if agents could automate duplicative inquiries and, ultimately, justify a native architecture.

Long-term:

• Design localized workflows to progress agents capabilities and scale towards automated documentation maintenance.

Short-term:

Testing the hypothesis.

Before investing in a native workflow, we needed to determine whether an AI agent could reliably resolve common support questions using the documentation we already had.

Long-term:

Native deliverables.

With localized workflows, agents could now draft changes directly to documentation, further automating the self-service cycle. However, more automation necessitated additional tools for observability and collaboration to avoid hallucinations.

Impact:

89

development hours preserved during the first three months.

65.15%

of support inquiries resolved without developer intervention during the first three months.

32.9%

of AI-generated documentation approved without revisions during the first three months.

Challenges & trade-offs:

Project momentum vs. limited capacity

Because the team was understaffed, I had to step outside my role to recruit a developer, manage the project and establish a regular cadence for collaboration. Rather than treating limited capacity as a blocker, I created a working structure that allowed the project to progress alongside the team's existing responsibilities.

Technical complexity vs. scalable simplicity

Early in the project, we discovered that allowing an AI agent to re-open conversations in legacy support threads would introduce significant engineering complexity. Instead, we chose to have the agent summarize previous interventions and provide that context as a PDF in a new chat when re-initiating conversations. This allowed us to preserve the intended workflow without introducing unnecessary complexity, reinforcing the value of shipping the simplest solution while leaving room to scale.

AI capability vs. content quality

Inconsistent formatting across documentation made reliable parsing and retrieval more difficult, revealing a constraint that wasn’t immediately obvious when we began exploring AI-assisted workflows. We discovered that limiting widget configurations to predetermined groups improved parsing while leaving room to introduce additional modularity in the future.