Work/Adobe
I led Adobe Help as a product for seven years: the strategy, the roadmap, and the organization behind an experience with 400M+ visits a year. The goal was resolution: customers solving their own problems without needing support. In 2026, self-resolution became Help’s primary measure.
Customers were reaching Help from new places. Search visits were falling by millions a year, links inside Adobe’s apps had become Help’s largest source of traffic, and more search questions ended in an AI-generated answer.
Help was still built to publish documentation, and its measures focused on cost and deflection. A customer could find the right article and still be stuck, reading a long page to begin fixing an error they hit inside the product.
I proposed rebuilding Help around what customers were trying to do. The goal was resolution: a customer who leaves Help with the problem solved.
The measures caught up in two steps. From 2023, product engagement was Help’s primary measure, with support contacts as the guardrail, and the rebuild launched against those measures at Adobe MAX 2025. In 2026, self-resolution became the primary measure, and we began measuring it directly.
The 2026 scorecard, shown here, gave product, content, engineering, operations, and analytics one definition of success.
How the purpose of Help, and the measures used to judge it, changed over six years.
Each investment depended on the others. AI answers drew on the rebuilt content, and the content work needed every partner team. The same measures applied to all three.
Every article was organized around one customer task, with task navigation, numbered steps, and a way into the product. We migrated Help to these templates, added schema.org structured data so search engines and AI assistants could read each task, and retired thousands of older articles as part of managing the collection over time.
Moving old articles into new templates would have carried old problems onto a cleaner site. I required every migrated article to be rewritten around one task. Peers agreed with the direction but worried about capacity. I held the standard; contractor funding and a structured content repository, which let us automate parts of the rewrite, made it workable.
We added AI-generated answers to the top of Help search, grounded in Adobe Help content and protected by guardrails in production. The same structured content let search engines and AI assistants find and cite Help.
After launch, timeouts, fallback routing, and unsupported questions could all look like the same error. Engineering and partner teams fixed the configuration and routing, and I worked on distinct messages so customers could tell a missing answer from an outage. A plausible answer isn’t enough; quality means grounded, relevant, and useful across the conversation.
Keep the underlying help content accurate and current.
Check that each AI-generated answer is supported by that content.
Carry those safeguards forward when another assistant uses or rewrites the answer.
After launch, timeouts, fallback routing, and unsupported questions could produce similar errors. Engineering and partner teams fixed the endpoint configuration and routing. I worked on the messages so customers could understand what had happened.
Infrastructure timeouts, conflicting fallback routes, and unsupported queries appeared as similar failures to customers.
Engineering and partner teams corrected endpoint configuration and routing. Production errors declined after the fixes.
A generic error concealed whether the assistant lacked an answer or the service was unavailable.
I worked on distinct messages for each guardrail response, so customers could tell a missing answer from an outage.
A plausible answer can still be unsupported, incomplete, or unhelpful across a conversation.
The evaluation team assessed grounding and source relevance alongside accuracy, coherence, repetition, and the customer’s experience.
Recovery and evaluation were shared team efforts.
My organization owned the platform, templates, authoring tools, analytics, and the engineering backlog. It covered product management, SEO, production, visual communication, and developmental editing, and I led most of it through five people managers.
Peer directors owned the content organization and the rollout, which moved product by product: a Premiere Pro pilot in late 2023, more web apps through 2024, the legacy desktop apps by Adobe MAX 2025, and the long tail in April 2026. Our teams planned against the same measures, so platform work, content migration, and the launch at MAX moved as one program.
People at peak. Seven direct reports, five of them people managers. Product managers, writers, editors, designers, and web producers.
Contributors across Adobe in product and content, engineering, operations, editorial and design, and analytics.
People from five partner organizations who worked on the launch at Adobe MAX 2025.
My organization and the 80+ partner contributors were separate groups. The 40+ launch collaborators were counted for the Adobe MAX 2025 launch alone, from five partner organizations.
Each result is listed with what it measures and where it came from.
Adoption, answers displayed, and click-through show how customers used the new experience. Self-resolution is the measure of whether their problem was solved.
Relevance beat volume. General recommendations drew little engagement; help tied to a specific problem performed better. That is why the rebuild went deep on real customer tasks.
Over the same year, estimated organic search traffic to Adobe Help fell 24%. Being the cited source began to matter more than ranking for a click.
Three sources: purchases made through Help, the growth team’s product engagement models, and personalization experiments with signed-in visitors.
We tested recommendations, personalized account help, and shorter answers. General recommendations attracted little engagement. Help tied to a specific problem was more promising.
Machine-learning recommendations on product and article pages showed little or no engagement.
Content tied to the customer’s task and the product they were using.
The personalized subscription card test showed flat overall support-contact rates. A targeted payment-error treatment showed more promise for self-service.
Specific account problems, each with an action the customer could take.
Several shorter-content and faster-answer tests reduced support conversations or improved positive ratings.
Shorter answers in task-first Help, with continued testing of whether customers resolved their issues.
Adobe Help experiment readouts.
Customers increasingly encountered Adobe Help through in-app links, search, and AI-generated answers. We needed reliable, task-focused content that could support all three. That meant changing how we structured, published, and maintained it.
Migrated the site to a consistent article experience with left-hand navigation and clear task instructions.
Added schema.org markup to article pages. The Photoshop article shown above identifies the task, five steps, product, and publisher.
Retired thousands of older articles as part of managing the Help content collection over time.
Product navigation places the task in context. Numbered instructions guide the customer through it.
The current Photoshop article shown above includes HowTo, HowToStep, and BreadcrumbList schema, linked to Photoshop and Adobe as the publisher.
Everything in this section is future work: in development, prototyped, or planned. The goal: more customers resolve issues on their own, and fewer need to contact support.
A designer asks the in-app assistant how to replace part of an image with generative AI. She has used all 250 of her monthly credits.
Today’s assistant gives correct instructions she can’t use. In the prototype, the assistant checks her account, finds the blocker, and offers options that work: buy credits, or use Content-Aware Fill.
“The answer was right. It just wasn’t right for her.”
These capabilities work together, in whatever order a problem needs. Many conversations end with an answer or guidance; an automated action is one option among several.
See the customer’s app, version, plan, and history before they type, starting with the most common account and product issues.
Use that context to find what is actually stopping the customer, as in the scenario above.
Most problems need clear guidance. Where an action helps, prepare it and let the customer confirm, such as reassigning a license from inside the help page.
Check whether the problem was solved, and feed that into self-resolution measurement.
Pass the issue and the customer’s context to support, so nobody has to start over.
Prototype design: Shruthi Chandra, Kate Lampe, and the Adobe Help design team.
Future vision, shown with sample accounts. Prototype design: Shruthi Chandra, Kate Lampe, and the Adobe Help design team.
Customer entry points
Adobe.com · Adobe Home · Creative Cloud desktop app · Adobe Help
Connect each entry point to the same help service.
Guidance grounded in help content, matched to the product version.
Use customer context to diagnose the problem and suggest a fix.
Pass the issue and customer context to support.
Carry out a change after the customer confirms it.
Proposed architecture. Entry points connect through shared routing and customer context to the same help service.
Proposed investment sequencing for a multi-year plan, separate from the timeline of shipped work. Later phases build on capabilities from earlier ones.
The product managers, writers, editors, designers, and web producers in my organization. They rewrote and migrated the articles, designed the new Help home and the vision prototypes, and worked with engineering, operations, and analytics partners on the launch.
Jack
Neetu
Wayne
Parul
Parichay
Mehak
Kate K.
Erin
Rich
Kate
Dharini