Case · SupportBot · Conversational AI · OpenText

SupportBot: an OpenText-branded AI support agent

A conversational AI assistant embedded in the Webroot and Carbonite account portal that resolves the most common onboarding, account and troubleshooting questions in-chat, before they become tickets. The design is in limited release testing: live in production with targeted user groups and actively monitored ahead of a broader rollout.

Role
UX Lead & Conversation Designer
Surface
Web account portal, site-wide
Team
UX, engineering, product, support ops, privacy
Goal
Reduce customer support workload
Limited release
Live in the production account portal with targeted user groups, ahead of a broader rollout
Human in the loop
Agentic behavior is actively monitored in production, including human-in-the-loop oversight levels
Deflection
The primary success measure. Outcome figures are published here once monitoring confirms them
01

Measures

SupportBot is in limited release testing with targeted user groups, so this page leads with design rather than results. Success is defined by one question: is it taking repetitive load off the support team?

Deflection, the primary signal

Share of common questions resolved in-chat without entering the human queue.

Containment

Conversations that end solved versus escalated to a human.

Speed

In-chat answers versus the prior queue-and-wait.

Capacity reclaimed

Reduction in human-handled low-complexity volume.

Trust

CSAT and follow-up-contact rate on AI-assisted interactions.

Oversight

Human-in-the-loop oversight levels tracked in production.

02

The work

The assistant at the point of friction, and in motion. SupportBot is available anywhere in the Webroot and Carbonite account portal, so it meets setup and account questions wherever they arise, and answers them in natural language instead of routing users to a ticket queue or FAQ. The disclosure “AI-generated responses might vary” sits at the top of the panel, and three suggested prompts map to the scenarios that drive the most support volume.

SupportBot as a docked side panel in the Webroot and Carbonite account home, open beside the Backup, Security overview, Identity Protection, and Subscriptions cards. The panel header carries conversation history, an overflow menu, and close; below the greeting sit the “AI-generated responses might vary” disclosure, a line asking people not to enter personal information, and three suggested prompts.
The shipped treatment. SupportBot docked as a side panel beside the account home cards. Greeting, AI disclosure, personal-information warning, three suggested prompts, text entry, in that order.

The sequence it came from

The account home at rest. SupportBot is present only as a launcher in the lower right corner, so the page reads normally until someone needs help.
The page at rest. The assistant is present only as a launcher in the lower right, so the account home reads normally until someone needs help.
The launcher with its invitation label, “AI supportbot is ready to assist you,” which tells people what the control does before they commit to opening it.
The launcher and its invitation label. “AI supportbot is ready to assist you” tells people what the control does before they commit to opening it.
The assistant opened as a panel anchored to the launcher. Same content model as the docked treatment: greeting, AI disclosure, personal-information warning, three suggested prompts, and a text entry field.
Opened from the launcher. Same content model as the docked treatment, so the disclosure and the prompts are never conditional on how the assistant was opened.
The opened panel shown against the page content it overlays, with the pointer on the panel header.
The open panel against the page it overlays. The page behind stays legible, because the question a person is asking usually refers to what is on it.

In motion

Screen recording: SupportBot resolves an onboarding question in-chat, no ticket filed.
Demo 1: an onboarding question. Resolved in-chat, no ticket filed.
Screen recording: SupportBot handles an account task conversationally.
Demo 2: an account task. Handled conversationally, in the portal the person was already in.
03

Process

Repetitive tickets in the queue, users stuck mid-setup.

As Webroot and Carbonite consolidated into one OpenText account experience, support inherited a steady stream of repetitive, low-complexity questions: how to get started, which devices can be backed up, how to change an account setting. They clogged the queue and slowed response on urgent issues. Users didn't want to file a ticket and wait; they wanted an answer mid-setup, and the static FAQ wasn't being read.

  • Placed at the point of frictionCatching the question on whatever page the person is already on deflects the ticket instead of escalating it. Site-wide in the account portal, not in a buried support section.
  • Disclosure surfaced, not hiddenThe “AI-generated responses might vary” notice sits below the panel header, not in a footer, so expectations are set before the first answer.
  • Prompts mapped to real volumeThree task starters solve the blank-slate problem and steer users toward the questions the assistant resolves best, drawn from what actually drives support volume.
  • Scope-bounded on purposeLimited to Webroot and Carbonite products and account tasks, which keeps answers reliable and shrinks the hallucination surface.
  • Monitored in production, with humans in the loopBuilt with engineering, product, support operations and privacy. Agentic behavior is actively monitored, including human-in-the-loop oversight levels, ahead of a broader rollout.

Deflection only works if users trust the answers and stay inside the assistant's competence. That principle is behind all four design decisions.

Full detail

Situation and task

As Webroot and Carbonite consolidated into one OpenText account experience, support inherited a steady stream of repetitive, low-complexity questions: how to get started, which devices can be backed up, how to change an account setting. They clogged the queue and slowed response on the urgent issues. Users did not want to file a ticket and wait either; they wanted an answer mid-setup, and the static FAQ was not being read.

The task: design an on-brand conversational agent that absorbs the repetitive load and frees human agents for complex cases, without the failure modes that make AI support worse than none: confident wrong answers, an empty text box, hidden disclosure.

Three requirements followed from that:

  • Deflect repetitive tickets. Resolve the highest-volume questions in-chat so they never enter the queue.
  • Answer at the point of friction. Surface help anywhere on the site, not in a buried support section.
  • Stay trustworthy and on-brand. Carry the OpenText identity, be honest about AI limits, stay scoped to real product tasks.
What I did, in full

A.Placed the assistant at the point of friction

SupportBot is available site-wide in the account portal rather than in a support section. Catching the question on whatever page the person is already on deflects the ticket instead of escalating it.

B.Surfaced the disclosure instead of hiding it

The “AI-generated responses might vary” notice sits below the panel header, not in a footer, so expectations are set before the first answer. A line asking people not to enter personal information sits with it, in the same position in both the docked and launcher-anchored treatments.

C.Mapped the suggested prompts to real volume

Three task starters solve the blank-slate problem and steer users toward the questions the assistant resolves best. They are drawn from the scenarios that drive the most support volume, which is what makes them a deflection mechanism rather than decoration.

D.Bounded the scope on purpose

The assistant is limited to Webroot and Carbonite products and account tasks, which keeps answers reliable and shrinks the hallucination surface. Staying inside its competence is what lets a person trust the answer they get.

E.Designed the responsible AI behavior with the team that owns it

The work ran with engineering, product, support operations and privacy. Agentic behavior is actively monitored in production, including human-in-the-loop oversight levels, which is the condition on a broader rollout.

Status, in full

Limited release testing. The design is live in the production account portal with targeted user groups and actively monitored, ahead of a broader rollout.

Human in the loop. Agentic behavior is monitored in production, including human-in-the-loop oversight levels.

Deflection is the measure. It is the primary success signal, alongside containment, speed, reclaimed capacity, CSAT and follow-up-contact rate. Outcome figures are published here once monitoring confirms them, and no outcome figures are claimed on this page.