Designing an Agentic AI Service Experience
How I helped shape a next-generation customer chat experience that moved beyond a legacy chatbot to support information seeking, guided action, trust, and human escalation across Allstate's authenticated account experience.
My work focused less on designing a chat window and more on defining the assistant's behavior — how it should interpret intent, guide action, confirm sensitive transactions, recover from uncertainty, and hand off to humans when needed.
Screens and details have been adapted or generalized to protect internal product information.
Why Allstate needed a new chat experience
Allstate's authenticated customers come to My Account with urgent, specific service needs. They want to understand their coverages, make a payment, check their policy status, get an ID card, resolve a billing question, or ask about reinstatement options. These are not browsing sessions — they are intent-driven moments where friction has real consequences.
The existing chatbot was limited in scope and less integrated with authenticated account context. It could handle basic query routing but couldn't understand nuanced intent, take action within the account, or adapt to the complexity of real customer needs. Customers who needed help often hit dead ends or were handed off to phone support without resolution.
The opportunity was clear: move from a legacy chatbot to a next-generation AI service assistant that could genuinely understand what a customer was trying to accomplish — and help them get there.
Turning natural language into safe, structured service actions
The design challenge was not simply to make chat smarter. It was to design an AI service experience that balanced several competing pressures simultaneously: conversational flexibility alongside structured task completion, customer trust and control alongside legal and compliance constraints, and graceful handling of unsupported intents alongside clear paths to human escalation.
"The assistant needed to understand natural language, but the experience could not behave like an open-ended LLM playground. It had to guide customers through bounded, confirmable workflows where sensitive actions stayed clear and controlled."
Insurance is a regulated environment. The assistant couldn't speculate about coverage, make unauthorized commitments, or process sensitive actions without explicit customer confirmation. Every agentic capability had to be designed with a failure mode — an escalation path, a disclosure, a confirmation step — that maintained trust even when things didn't go as planned.
UX design for the assistant's behavior, not just its appearance
I led UX design for key parts of this experience. My focus was less on the aesthetic of the chat window and more on defining how the assistant should behave across the full range of customer interactions — from the first message to the final confirmation.
I worked closely with product managers, engineering, journey teams, conversation designers, CX, and compliance and legal stakeholders — aligning on both what the assistant should support and how it should behave when it reached its limits.
Designing the assistant as a system, not a screen
The most important design work was not what the assistant looked like — it was how the assistant behaved. I helped shape a framework of reusable behavioral patterns that defined the assistant's decision logic across every type of customer interaction.
"In AI experiences, behavior is part of the interface."
These patterns governed the full range of the assistant's actions:
The result was not a library of one-off screen designs. It was a behavioral system — a shared decision layer that made the assistant's responses consistent, predictable, and trustworthy across every supported intent.
Defining the first launchable scope
The assistant could not launch with every possible servicing need. The team evaluated and prioritized a first set of customer intents based on a combination of factors: customer value, chat volume, technical feasibility, service risk, and product readiness.
Each intent was also classified by interaction type — information-seeking, confirmation, guided transaction, escalation, or fallback/recovery. This classification mattered because it determined the right design pattern: a coverage question and a payment action require fundamentally different approaches. A customer asking "what's my deductible?" needs a clear, structured answer. A customer saying "I want to make a payment" needs to be guided through a confirmable multi-step flow before anything happens.
From natural language to guided action
For high-impact tasks like payments, the design challenge was knowing when to let the assistant lead conversationally and when to shift into structured, confirmable UI. Open-ended dialogue is the right tool for understanding intent — but not for moving money.
The one-time payment flow demonstrated this clearly. When a customer expressed intent to make a payment, the assistant engaged conversationally to clarify timing and amount. Once those details were confirmed, the assistant transitioned into a structured review step, followed by an explicit confirmation action.
"Conversational intent became a structured, confirmable transaction flow. The assistant could guide the customer, but the customer stayed in control before money moved."
No payment moved without explicit review and confirmation. This is where trust was designed — not in the words the assistant chose, but in the structure it created around the action.
Making complex policy information scannable
For information-seeking journeys, the design challenge was different: how do you make dense, legally precise insurance information understandable without oversimplifying it?
When a customer asked "what are my coverages?", the assistant couldn't return a wall of text. But it also couldn't strip out the details customers actually needed to understand their protection. The solution was structured coverage cards — a designed response format that surfaced premium, coverage name, limits, and subcoverage details in a visual hierarchy built for scanning.
"The goal was not to make policy information feel simplistic. It was to make it legible, structured, and easier to verify."
Progressive disclosure let customers navigate from a high-level summary to coverage-specific detail without overwhelming them with everything at once. The card format also created a consistent pattern that could be reused across other information-seeking intents throughout the experience.
Designing for waiting, uncertainty, and escalation
Agentic AI systems are not instantaneous. Retrieving account context, checking eligibility, or processing a service action takes time — and that time is visible to the customer. Left undesigned, it becomes a trust gap.
"In AI experiences, waiting is not empty time — it is a trust state."
The assistant's delay and progress states were designed to be transparent and progressive. The assistant communicated clearly at each stage — working, gathering details, thanking the customer for patience — and flagged when a delay extended beyond the expected range, offering a specialist connection before the customer had to ask.
This escalation path was designed as a feature, not a failure. Handing off to a live agent was not a fallback of last resort — it was a transparent, dignified transition that preserved the customer's service context and maintained trust precisely when the AI had reached its limits.
Closing the loop — feedback and service recovery
The end of an AI conversation is both a measurement opportunity and a service recovery point. The end-of-chat feedback flow was designed to capture what happened after the interaction closed — and open a path to resolution when it hadn't.
The flow asked customers whether their need had been met, how likely they were to use the assistant again, and — if applicable — what specifically could have been better. For customers whose needs weren't resolved, the flow surfaced a live-agent option: turning a dissatisfied session into a continued service interaction rather than an abandoned one.
The feedback data gave the team signal across four dimensions: resolution success, dissatisfaction patterns, recovery needs, and future-use confidence. It also created the measurement infrastructure needed to prioritize improvements across subsequent releases.
Two tests. Approximately 100 customers. Greenlit for launch.
Two successful prelaunch customer tests were completed before the product received the green light for launch. Approximately 100 customers participated across both rounds, engaging with the assistant across real servicing needs inside their authenticated account experience.
Customers successfully used the assistant to retrieve coverage details, billing information, policy status, and document access — validating the assistant's ability to surface accurate, structured responses to real service questions.
Early guided transaction flows were validated in test conditions, confirming that customers were willing to take service actions — including payments — through an AI assistant inside an authenticated insurance account.
Live-agent handoff functioned as an important recovery path. Customers who reached the assistant's limits were successfully transferred without service failure — confirming escalation as a feature, not a fallback.
Following the second successful test round, the product received internal approval to move toward launch — validating both the user experience and the underlying agentic capability in a regulated insurance environment.
What this work proves
Designing for AI is not about designing better chat bubbles. The most consequential work on this project happened before any screen was built — in defining how the assistant should behave, where its boundaries should be, what confirmation looked like for a sensitive action, and what a good escalation felt like for a customer who needed more than the AI could give.
Agentic systems need structure around them. Natural language is the entry point, but structure, confirmation, and recovery are the experience. The customers who trusted the assistant enough to make a payment, check their coverage, or ask a complex policy question were trusting a designed system of behaviors — not just a text interface.
This project reinforced that the most important design skills in the agentic era are the ones that have always mattered most: understanding user intent deeply, designing for trust systematically, and knowing when to keep the human in the loop.
"The future of customer service AI is not a smarter chatbot. It is a service experience that can understand intent, guide action, preserve trust, and know when to bring in a human."
