Introducing Ada Labs: Applied AI research for agentic customer experience

For decades, the customer experience has been structured around channels, queues, and human workflows, where each interaction—whether support, sales, or customer success—was treated as a discrete event handled in isolation.

That model is now shifting, as AI agents emerge as the system that manages customer relationships end-to-end. They can understand context, take action, coordinate across systems, and improve continuously.

This is agentic customer experience (ACX), a model where AI agents do more than respond to requests. They manage outcomes across the entire customer journey and create measurable impact on cost-to-serve, customer satisfaction, and lifetime value.

However, reaching that future depends on solving a set of problems that most AI systems are not yet designed to handle.

The problems that will define agentic customer experience

The challenge now is building AI agents that can operate reliably across real customer interactions, handling complexity across conversations, channels, and time.

That requires solving a fundamentally different class of problems.

1. Adapting to real conversations

Customers don’t speak in clean prompts, they describe situations across multiple turns, with context that unfolds over time.

AI agents need to work within that reality to understand how intent evolves, retrieving information that reflects the current state of the interaction, and deciding what to do next. In practice, that can involve clarifying details, applying policies, or taking action across systems.

The goal is to move the interaction forward based on what the customer is actually trying to accomplish, not force it into a predefined flow.

2. Maintaining continuity over time

Many customer interactions extend beyond a single session, across hours or days, with new information and dependencies emerging along the way.

AI agents need to maintain context, track progress, and follow through as situations evolve. Without that continuity, customers end up repeating themselves and interactions restart from the beginning.

3. Measuring quality at scale

Enterprise AI requires continuous evaluation across production interactions. Without clear visibility into how the system is performing, it’s difficult to identify gaps or improve outcomes.

Teams need to understand resolution rates, performance across different interaction types, and where the system is falling short in order to improve it in a structured way. Solving these challenges is what enables AI agents to move beyond reactive interactions and operate as systems that manage customer relationships.

Introducing Ada Labs

Ada Labs is our applied AI research division. It brings together the machine learning scientists, AI engineers, and other product builders who are working on the systems that power agentic customer experience.

We’ve been doing this work for years. Across more than 350 enterprise deployments, Ada’s systems process over two trillion tokens each month. The architectures and evaluation systems behind Ada Labs are already running in production, supporting real customer interactions at scale.

Ada Labs formalizes that investment and makes it visible. We are now sharing more about how these systems are built, how they perform in production, and how they continue to evolve.

Our focus is straightforward: advancing the capabilities required for AI agents to operate reliably across the full customer journey.

Applied research that ships

Ada Labs is grounded in applied research that ships to production environments.

Every system we publish operates on live customer interactions at scale. This keeps the work focused on building AI systems that can operate reliably in enterprise environments, where consistency, accuracy, and the ability to continuously improve all matter.

The Reasoning Engine

The Reasoning Engine drives every decision Ada's AI agents make. It decides what to do at every turn of a conversation: when to answer, when to ask a follow-up, which procedure to apply, when to call a tool, and when to hand off.

Customer interactions are rarely single-step. A billing question can require checking account state, applying a policy, executing an action across systems, and explaining the result. Our work here focuses on planning, action selection, coordination across specialized models, procedure adherence, and balancing latency with reasoning depth.

Context Engine

The Reasoning Engine is only as good as the context it receives. The Context Engine is the system that assembles everything the Reasoning Engine needs at a given point in a conversation: the customer's profile, the conversation history, the knowledge that applies, and any coaching feedback that's been given on how to handle similar cases.

Our research here focuses on how that context gets retrieved, structured, and delivered.

ACX Improvement Loop

Every Ada AI agent has an ACX Manager responsible for ongoing performance. The ACX Improvement Loop is the set of tools used to measure, refine, and improve that performance over time:

• Topics and Intents surface what customers are asking and how the agent is handling it,
• Simulations test changes against scenarios the customer defines,
• The MCP server allows AI Managers to manage their AI agent from any LLM client, and
• Proactive recommendations flag where to focus next.

Our research here focuses on helping ACX Managers automate the loop from production signal to agent improvement.

Evaluation infrastructure

Evaluation infrastructure is how we check that changes to Ada's AI are working before they ship. Every model, prompt, and architecture change is tested against representative customer interactions, with checks for things like resolution, accuracy, safety, and policy compliance.

Our research focuses on automated evaluation that catches issues proactively so we can improve Ada's AI faster and ship with confidence.

Where Ada Labs is pushing next

The systems above form the foundation. Ongoing research focuses on extending what AI agents can manage across the customer lifecycle.

Current areas of focus include:

• Agent-agent collaboration: AI agents that share context, coordinate decisions, and manage complex interactions together, so customers don’t need to repeat information at each handoff.
• Voice experiences: Real-time voice AI supported by a dual reasoning architecture, designed to maintain both low latency and the depth required for more complex interactions.
• Long-running tasks: Systems that maintain intention across hours or days, supporting workflows like insurance claims, order tracking, and multi-step returns with ongoing follow-up.
• Background work: AI agents that analyze patterns, identify at-risk customers, and surface opportunities for improvement without waiting for interaction to be initiated.
• Industry-specific retrieval evaluation: Evaluation frameworks that measure retrieval quality within specific CX domains such as ecommerce, financial services, telecom, and travel, where context and expectations vary significantly.

These efforts reflect the broader goal of building AI systems that can manage ongoing relationships, not just individual interactions.

What this means for the enterprises betting on agentic CX

AI is becoming a central part of how enterprises engage with their customers. As adoption increases, the underlying systems that drive how AI agents reason, retrieve context, and are evaluated play a larger role in determining outcomes.

Organizations that invest in these foundations are better positioned to deliver consistent, high-quality experiences across the full customer journey.

Ada Labs makes our work in this space more visible. We're going to keep sharing how we build, evaluate, and improve the systems behind Ada's AI Agents, so teams evaluating different approaches have something concrete to compare. Many of the hardest problems in AI customer service are still ahead of us. That's what Ada Labs is for.