← Home
Enterprise agentic AI chat interface & workflow
An evolving AI assistant that absorbs the time-consuming work, so support reps arrive at every case ready to resolve, not research.
Overview
Support reps don't have a knowledge problem. They have a fragmentation problem. Every case requires piecing together information from email threads, historical cases, and internal tools, before they can send a single reply to a customer. This project reimagined what support work could feel like if reps across Billing, Learning, Hiring, Trust, and more had an AI assistant that lived inside the case, one that had already done the groundwork before they opened it. Not a smarter search bar. A system that shows up to every case already prepared. As the lead product designer, I was responsible for the end-to-end design of the agentic chat experience: defining the UX framework, designing the conversational interface, and shaping how humans and AI share decision-making across 30+ automated actions, a number that continues to grow.
On average, a support rep spends 5 minutes researching a standard case before the first reply. For complex cases, those requiring cross-referencing emails, historical tickets, and internal documentation, that climbs to 11 minutes.
Multiply that across hundreds of reps and thousands of cases, and you have an invisible but enormous drain on productivity. More importantly, you have reps spending the majority of their cognitive energy on information retrieval, not on the customer in front of them.
1.
Tool fragmentation
Reps navigate multiple disconnected platforms to gather context, SharePoint, email, case history, internal tools, with no unified view.
2.
Manual, repetitive actions
Common workflows like merging profiles, updating records, or escalating cases required reps to leave the case entirely.
3.
No shared context
Every case started from zero. There was no system that had already read the thread, pulled the history, and handed the rep a starting point.
🚧
The problem wasn't that reps lacked tools. It was that the tools required reps to do the work the tools should have been doing.
SOLUTION
The goal wasn't to build a chatbot. It was to design a system that felt like a knowledgeable teammate, one that was already caught up on the case before the rep opened it, could take action when asked, and always checked before doing something that couldn't be undone.
Four capabilities make up the experience:
Case overview summarization
When a rep opens a case, a summary automatically populates, customer name, core issue, previous actions, and related cases, pulled from across all content sources. The rep arrives briefed, not blank.
Agentic quick actions
Reps can invoke tools directly in the chat, merge profiles, update records, delete entries, without leaving the case. The AI identifies the right tool, collects what it needs conversationally, then waits for rep approval before executing.
Internal knowledge lookup
Reps can ask questions and receive answers surfaced from internal knowledge bases, similar resolved cases, and relevant documentation, without switching tabs or running manual searches.
Agentic chat experience
The conversational interface that holds all four capabilities together. Designed so reps can invoke any tool, ask any question, and take any action, in one place, in natural language, without context switching.
The hardest part wasn't any of that. It was the half-second before the AI acts.
When an action can't be undone, who decides and how does the AI ask for permission without making the rep feel like they're rubber-stamping a black box? Getting that one moment right, what the AI shows, when it pauses, how it earns the rep's go ahead, this took more iteration than the rest of the system combined. It's the difference between a tool reps tolerate and one they trust with real work.
HOW THE AI EVOLVED
The system didn't start agentic. It earned its autonomy in three phases. Each phase quietly changed the design problem underneath it.
PHASE 1
Deterministic flow
Fixed logic, predefined sequences.
The job: make a rigid system feel clear and trustworthy enough that reps would let it drive. The answer had less to do with the flow itself than with what we chose to make visible at each step.
PHASE 2
Single-tool agentic
Now the AI selected the tool and gathered its own inputs.
The job: shift design problem from clarity to control, keeping the rep in charge of a system that had started making its own calls.
PHASE 3
Multi-tool orchestration
The AI began to plan and sequence several tools at once.
The job: give reps visibility into something that reasons, not just executes. It's the open frontier of this work and the part I find most interesting to design for.
RESULT
↓ 44% - Resolution time
Average resolution dropped from 59.4 to 36.1 hours across every launched action.
18,500 + hrs / month - Rep time saved
Through automated case summaries and content suggestions.
↑ 30% - Rep productivity
Four times the original 10% target, within the first year.
📣
There's a lot I can't show publicly about how the human-AI handoff actually works, the approval flows, how the AI earns trust, what we learned the first time reps didn't believe it. That's where most of the real design happened. Happy to walk through it.
