Meesho · AI Services · Enterprise / Conversational UX
My role
Senior Design Manager
Timeline
2024-25
Type
Enterprise · AI / Conversational UX
Context
Meesho’s AI Services vertical runs a voice bot that handles inbound and outbound customer interactions across multiple client ecosystems. When MeeBot cannot resolve an issue, it escalates to a human agent. But escalation alone was never the problem. The problem was what happened the moment an agent picked up. Agents were working across five or more fragmented systems simultaneously: telephony, CRM, SOP documents, internal escalation chat, and a manual tracking sheet. The system functioned because agents adapted to its complexity, not because it helped them.
Research
Before any design work began, the team conducted foundational research across three client ecosystems [names detracted], each using the same Agent Assist infrastructure but operating in very different ways. Research included live call shadowing, contextual inquiry with L1 and L2 agents, and workflow mapping across every tool used in a single interaction. The goal was ground truth: how agents actually think, decide, and cope under real-time pressure.
Telephony
Live call control and call state.
CRM
Customer and ticket records.
SOP sheets
Search-heavy policy guidance.
Escalation
Supervisor support and approvals.
Manual tracker
Backup memory after the call.
Unity is not a convenience
A system requiring agents to hold context across five tools is a burden, not a tool. Any solution that does not unify the workspace inherits the same cognitive debt.
Every copy-paste is a system failure
When agents manually move data between platforms, the system is offloading its own integration failures onto people.
A knowledge base you must search has already failed
Intelligent systems should anticipate the question — surfacing guidance before the agent has to formulate a search.
Recall is not a documentation strategy
When post-call documentation depends on memory, accuracy degrades with every interruption. The system must capture context as it unfolds.
If it only works for the expert, it is not working
New agents memorise workflows over time because they have no choice. A well-designed system makes good performance the default — not the reward for experience.
Design direction
Early explorations tested a linear step-by-step layout, a segmented modular system, and a bento-style grid. Each round exposed a tradeoff — between guidance and flexibility, density and cognitive load, structure and adaptability. The turning point came when the framing shifted: the platform could not just be a better interface for existing tools. It had to be a unified execution layer — actively guiding agents from call start to close within a single environment.
Left panel
Context, not clutter
Auto-loaded customer data, ticket details, verification status, and conversation history — persistently accessible without manual retrieval.
Centre panel
Workflow as the product
SOPs transformed from documents into intent-triggered interactive workflows. Progressive disclosure keeps future steps out of view until needed.
RIGHT panel
Action without distraction
Escalation, case creation, and resolution controls isolated on the right — available when needed, never competing with the primary workflow.
Strategic outcomes
A shift in philosophy
Reframed Agent Assist from a support utility into an execution layer, changing the agent's role.
A unified platform architecture
The three-panel model consolidates five fragmented tools into one environment, reducing manual transfer and cross-platform switching.
Intent-led AI integration
AI moves from passive tool to contextual collaborator, surfacing SOPs and adapting panels based on interaction stage.
Designed for the new agent
Progressive disclosure and structured workflows make good performance the default, reducing ramp-up time and dependency on institutional memory.