Roku Evaluates 29 Vendors and Chooses Quiq to Scale Support for 145MM+ Customers

The Roku support team had looked at AI for years. First-generation chatbots weren’t the answer.

“The traditional approaches — gen one bots — meant building and maintaining thousands of static conversation flows, which just isn’t sustainable given how rapidly Roku is growing and how small but mighty our team is,” said Feinstein.

The fundamental problem with those earlier bots was architectural. Each one required manually creating and maintaining a specific flow for a specific intent. For a company with Roku’s product breadth, that meant thousands of flows — each of which had to be written, tested, and updated whenever policies, products, or procedures changed. The math didn’t work at Roku’s scale.

Several compounding problems made the status quo unacceptable:

• Content rewriting was a non-starter. Roku had hundreds of existing knowledge base articles that human agents relied on daily. These weren’t formatted for AI consumption. The team didn’t have the bandwidth to rewrite or reformat them, but they also couldn’t afford a solution that required starting from scratch.
• Scope explosion with no ceiling. Roku’s customers could ask about hardware troubleshooting across dozens of device models, billing issues across thousands of third-party channels, account management, content access, and more. A traditional bot would require a separate flow for each — an impossible maintenance burden for a lean team.
• No ability to handle vague or off-script questions. Roku’s customers are often confused when they reach out. They may not know the name of the product they’re asking about, or they may describe their issue in a way that doesn’t match any predefined intent. “Our customers, I’ll be honest — often when they reach us, they’re confused and they just want an answer, and they might have a vague question that a generation one bot just can’t handle,” Feinstein said.
• No visibility into what was happening. Competing solutions often required building analytics capabilities from scratch, or depending on the engineering team to dig through logs just to understand basic performance. Roku needed a single place where business users could track outcomes and find improvement opportunities without engineering involvement.
• The channel mix was working against efficiency. Roku’s largest contact channel by volume was direct phone calls. Many customers bypassed the website entirely, found the phone number on Google, and called in. That meant the existing AI agent on chat couldn’t even intercept the interaction — and phone contacts require one human agent per conversation, versus the two simultaneous chats a human agent can handle.

Roku chose Quiq after a rigorous evaluation process that began with 29 AI vendors, narrowed to five finalists (Quiq, Kore.ai, Netomi, Cognigy, and Amazon), and ultimately came down to a head-to-head proof of concept bake-off between Quiq and Amazon. Quiq won on three factors: demonstrated agentic AI capability, deep CRM integration expertise, and a collaborative professional services model that could do the heavy lifting while still giving Roku the control and transparency they needed.

The deployment was built on four capabilities specific to Roku’s situation:

• Data transformation without content rewrites. Roku’s existing knowledge base articles were written for human agents and included HTML formatting, customer support phone numbers, and other elements that would confuse an AI agent. Quiq’s platform transformed these sources — stripping unwanted content, adding question mappings, enriching articles with helpful context — while keeping them in sync with the original source. Roku did not have to maintain two versions of their content.
• Bidirectional platform integrations. To answer account-specific questions, the AI agent needed to know who it was talking to. Quiq integrated with Roku’s subscription and payment platforms, Zendesk, and Amazon Connect. When a customer logs in during a chat conversation, the AI agent pulls their specific account data — active subscriptions, device information, billing history — and uses it to answer their actual question rather than a generic version of it.
• An orchestration layer with process guides. Rather than hard-coded flows, Quiq uses “process guides” — sets of instructions, best practices, and available tools that the AI agent selects and applies based on the context of each message. These function similarly to the guides a human agent would follow. On every incoming message, the system evaluates not just the topic but whether the contact is a sales or service inquiry, whether the customer is new or existing, and how frustrated they appear — then selects the appropriate guide and moves the conversation forward. Guardrails run before and after every generated response to prevent off-topic, off-brand, or inaccurate answers.
• AI Studio for lifecycle management and observability. The AI agent was built, tested, and is managed inside Quiq’s AI Studio. This gave Roku the “best of both buy and build,” as Feinstein described it: Quiq handles the maintenance, scalability, and AI infrastructure, while Roku retains the observability, flexibility, and control to understand what the agent is doing and why. Every conversation is automatically scored across criteria Roku defined — contact driver, accuracy grade, resolution score, estimated customer satisfaction — and surfaced in a reporting dashboard that business users can operate without engineering support.

Two categories of inquiries drive the majority of Roku’s support volume, and the AI agent handles both.

General product troubleshooting: A customer opens a support chat and types something as vague as “it’s not working.” The AI agent doesn’t try to match that phrase to a predefined intent. Instead, it asks clarifying questions in real time — is this a product issue or an account issue? — then, once it understands the problem, generates tailored troubleshooting steps based on the specific issue described. A customer who mentions a particular error code gets instructions relevant to that error code, not a generic list. A customer who describes a symptom without knowing the cause gets a diagnostic conversation that walks toward an answer. At no point does the agent follow a scripted path.

Account-specific inquiries: When a customer asks about a charge they don’t recognize, the AI agent prompts them to log in. Once authenticated, it pulls the customer’s active subscriptions and surfaces the one most likely to match the charge in question, asking the customer to confirm. From there, it can answer follow-up questions — how long has this subscription been active, how do I add a PIN to prevent unauthorized purchases — without needing a hard-coded script for each. If the customer declines to log in, the agent adapts: rather than looping back to a sign-in prompt or immediately escalating, it provides general guidance from its knowledge base and continues the conversation.

When an interaction does require human involvement, the AI agent passes a full conversation summary, transcript, and pre-populated ticket fields to the human agent — so the customer doesn’t have to start over. “We spent a lot of time working with Quiq to make sure that handoff worked properly,” Feinstein noted. “That context being passed when handing off a conversation to a human advisor is critical.”

One finding from the Roku team stands out as a proxy for the AI agent’s comprehension quality: Feinstein described reviewing transcripts and being surprised by how the agent understood what customers were actually asking. “We’ve been really impressed looking at transcripts, thinking ourselves, ‘How on earth did this bot even understand what the customer was asking?’” That ability to handle the long tail of customer questions — even when they don’t follow a scripted happy path — is what makes the system usable at Roku’s scale.

For the Roku support team, the AI agent changed the composition of the work rather than the size of the team.

In a controlled A/B test, the AI agent was shown to a subset of US customers alongside the original friction-heavy web contact flow — the kind of experience where customers navigate forms and self-service options before reaching a human. The AI agent achieved resolved nearly half without a human agent. That exceeded Roku’s expectations and triggered a full US rollout.

Beyond containment, two shifts emerged that mattered operationally. First, the contact channel mix improved: with a more engaging conversational experience available in chat, the share of customers who chose chat over phone moved from 30% to 43%. Because human agents can handle two simultaneous chat conversations but only one phone call, that shift directly improves team efficiency. Second, the proportion of conversations handled by human agents that required genuine human judgment — what the team calls non-self-serviceable contacts — rose from 28% to over 30%. The AI agent was successfully pulling routine work off the human queue and leaving the complex cases behind.

The analytics dashboard in AI Studio gave the team a fast feedback loop they hadn’t had before. Quiq’s system automatically assigns a primary contact driver to every conversation and injects it into Roku’s CRM as a ticket field — letting the team track contact trends through their existing tooling without manual tagging. When Feinstein’s team sees accuracy scores drop or resolution scores flag in a particular category, they can drill into the underlying transcripts, identify what went wrong, and make targeted changes to the process guides or knowledge base content.

“It’s a fast feedback loop for us,” Feinstein said. “It lets us continually improve the agent and measure the impact of changes over time.”

Partnering with Quiq has given Roku the infrastructure to scale customer support without scaling headcount proportionally. A 52% AI containment rate — achieved in the first iteration of the agent — represents a meaningful reduction in human agent load across Roku’s largest customer market. The shift in channel mix is producing daily efficiency gains in the contact center. And the quality of work reaching human agents has improved, with a higher proportion of genuinely complex cases making it through.

Key outcomes:

• 52% AI containment rate in initial rollout, exceeding Roku’s expectations and triggering a full US deployment
• Contact channel mix shifted from 70% phone / 30% chat to 57% phone / 43% chat, improving human agent capacity utilization
• Human agent workload shifted toward genuinely complex contacts, rising from 28% to 30%+ of all handled conversations
• Full context passed on every human handoff — conversation summary, transcript, and pre-populated CRM ticket fields — eliminating the restart problem
• Knowledge base content reuse without rewrites, keeping Roku’s single source of truth intact
• Business-user-operable analytics dashboard, eliminating engineering dependencies for performance monitoring

Roku is currently expanding the AI agent’s capabilities: enabling it to take action on subscription billing adjustments (currently human-only), adding personalized device data so the agent knows a customer’s specific Roku model without being told, and building toward a voice AI agent to address the large volume of customers who call directly without visiting the website first.