Conversational flow modeling hybrid chatbots experiences (deterministic + AI)
4–6 minutes
Challenge
EY employs approximately one new employee every eight minutes, making scalable, self-service HR support a critical business need. The organization partnered with IBM to deploy a hybrid conversational AI experience which combined deterministic rule-based logic with machine learning to support employee onboarding and HR inquiries.
By the time Phase II began, the underlying challenge was no longer technical feasibility, but operational usability: stakeholders struggled to review, validate, and approve the conversational logic needed to expand the chatbot’s capabilities.
Although IBM provided conversational flow diagram proposals for prioritized HR intent areas, those artifacts consistently failed to support efficient review or confident decision-making. Approval timelines stretched, intent mismatches persisted, and user trust suffered.
The core risk was not the chatbot itself but rather the inability of stakeholders to reason about how the system would behave.
Project profile
My role
Conversational UI Designer
Duration
~3 mos
Team Composition
Small internal team of 3 + Stakeholder. Larger partner team consisting of 15+ conversation specialists (linguists, tech, AI).
Article (04/2019)
How EY’s journey created a major business impact
Role and responsibilities
Conversational UI Designer
I was accountable for reviewing, correcting, and approving partner-delivered conversational flow models produced by IBM for EY’s Watson chatbot.
Acting as a liaison between internal EY stakeholders and IBM, I was responsible for:
- Evaluating whether conversational flows accurately represented user intent handling
- Identifying structural and UX issues that could degrade user trust or satisfaction
- Translating technical logic into stakeholder-legible system diagrams
- Reducing approval friction without compromising conversational integrity
- Figma branded design system based on the Microsoft open-source Teams design kit
While IBM authored the initial flow diagrams, I owned their usability, clarity, and readiness for approval.
Skin up a branded front-end chatbot UI using Microsoft Team’s kit
While the backend was powered by IBM Watson, the chatbot’s front end was designed for deployment within Microsoft Teams. I was responsible for applying EY’s branded design system to create a customized, cohesive EY chatbot experience.
The core problem
The failure was both intent modeling and how intent logic was visualized.
IBM’s proposed conversational flows deviated from their own published best practices and exhibited several recurring UX architecture issues:
- Overloaded diagrams combining logic, copy, and system detail into single nodes
- Horizontal layouts that obscured conversational progression
- Inconsistent or missing termination states, implying scenarios where the bot would not respond leaving the user in the lurch (hard fall)
- No visual distinction between user actions, system decisions, or backend processes
As a result, stakeholders were forced to mentally simulate conversations instead of reviewing them which lead to confusion, prolonged feedback cycles, and reduced confidence in the system.
Best practices abandoned
The failure was both intent modeling and how intent logic was visualized.
IBM’s conversational flows deviated from their own published best practices and exhibited several recurring UX architecture issues:
- Overloaded diagrams combining logic, copy, and system detail into single nodes
- Horizontal layouts that obscured conversational progression
- Inconsistent or missing termination states, implying scenarios where the bot would not respond
- No visual distinction between user actions, system decisions, or backend processes
As a result, stakeholders were forced to mentally simulate conversations instead of reviewing them which lead to confusion, prolonged feedback cycles, and reduced confidence in the system.
Design intervention: Establishing a shared visual vocabulary for displaying and reviewing complex branching logic
To resolve and shorten the review cycle pain points, I built a custom Axure library and re-modeled the partner-provided conversational flows using Jesse James Garrett’s Visual Vocabulary as a neutral, vendor-agnostic abstraction layer.
This approach allowed stakeholders to reason about conversational behavior without needing to understand IBM Watson’s internal mechanics.
Key improvements included:
- Vertical flow orientation to mirror conversational progression
- Swim lanes to clearly separate user input, system logic, and backend actions
- Explicit decision points to clarify “can the bot answer?” logic
- Clearly defined success, fallback, and escalation paths to eliminate conversational dead ends
- Inclusion of user utterance data to ground decisions in real usage patterns
These changes transformed the diagrams from technical artifacts into decision-support tools.
Outcomes
The redesigned flow models eliminated stakeholder frustrations and shortened review cycles from ~1 week to ~2 days. These improvements also established a shared mental model across EY stakeholders and IBM, directly improving both internal efficiency and user outcomes.
Measured results included:
- Reduction in stakeholder review and approval cycles from approximately one week to two days
- Improved clarity around intent handling and fallback behavior
- A 17-point increase in initial Net Promoter Score (from 25 to 42), attributed to improved intent matching and reduced conversational dead ends
Project conclusion and strategic shift
The project concluded not due to delivery failure, but due to a strategic inflection point in the conversational AI landscape.
The emergence of large language models such as ChatGPT fundamentally altered the cost–benefit equation of hybrid rule-based conversational systems. EY chose to pivot toward AI-first conversational strategies aligned with these new capabilities, ultimately transitioning away from IBM Watson in favor of ChatGPT within Azure OpenAI.
The work completed during Phase II validated the importance of:
- Clear conversational system architecture
- Strong abstraction between logic and presentation
- Stakeholder-legible design artifacts
