Designing a Conversational AI Assistant for Financial Planning and Analysis
User Research · Oracle EPM · Fintech · Development
PRODUCT OVERVIEW
A conversational AI assistant designed for financial analysis teams, capable of integrating GPT 4 with Oracle EPM. It enables users to query financial data in natural language, reducing dependency on manual reports and dashboards.
TIMELINE
6 Months
ROLE
-
UX Designer in an emerging technology innovation initiative
-
Researched generative AI landscape and reporting tools to identify opportunities
-
Designed conversational flows, integration architecture features, and prototypes
-
Partnered with developers and stakeholders to align on implementation
TOOLS
Figma · OpenAI API · Oracle EPM · REST API
Specifics of this project are confidential and cannot be disclosed. However, I have provided a basic overview of work I have done as part of this project. All assets and details displayed here-in are generic representations.
PROBLEM
Finance teams spend hours navigating dashboards, filters, and formulas. What if they could simply ask their ERP questions in plain English?
According to public industry studies, 41% of FP&A work remained manual, consuming up to 10 hours a week in spreadsheets and error-correction.
For analysts, that means hours wasted pulling reports. For executives, it meant delayed clarity on critical decisions. And for planners, it meant models and assumptions were rarely consistent across teams. Most FP&A teams have:
In 2023, as generative AI tools like ChatGPT were rising to prominence, Deloitte launched an internal initiative to explore how large language models could transform enterprise finance. I joined this effort as the UX designer tasked with defining the user interface and underlying integration for how financial planning and analysis (FP&A) teams can use AI to interact with Oracle EPM Cloud.
THE USERS
Our solution focused on three key user groups with distinct needs within the financial planning and analysis ecosystem.
Our focus centered on three core groups within FP&A, each with distinct needs we designed around:
-
Analysts: Analysts spent hours pulling numbers into spreadsheets, applying filters, and reconciling formulas. Their biggest pain point was speed: they wanted a faster way to get answers without drowning in manual work.
-
Executives: Executives needed high-level clarity for board decks and reviews but often received dense, technical outputs. Their challenge was translating raw numbers into clear insights, quickly enough to act on.
-
Planners: Planners worked across multiple scenarios and departments, often with inconsistent assumptions. They needed consistency and alignment, as well as a way to ensure everyone was working from the same numbers.
OUR VISION
Transparency, reusability, and approachability formed the core of the design vision for the conversational solution.
Our goal was to design an assistant that felt like a teammate, not a black box, delivering speed, clarity, and consistency without jargon or uncertainty. To ground our design process, we conducted a competitive analysis, looking at other generative AI tools to identify best practices and trends.
THE PROCESS
We followed an iterative process through research, design, prototyping, and testing with each stage shaped by direct user feedback.
We began with informal interviews and internal discussions to understand reporting pain points. To ground feasibility, I also explored REST API integrations and EPM query capabilities to see how data could realistically be pulled and surfaced.
IDEATION & DESIGN
Because this was an MVP exercise, we prioritized speed, simplicity, and confidence over breadth of features. Rather than trying to replicate every capability of Oracle EPM, the goal was to rapidly design and develop a version of the assistant that could start getting real-world use and stakeholder feedback.
PROTOTYPING
We translated the design concepts into a mid-fidelity prototype that simulated the core ask-and-answer flow. The goal wasn’t to showcase visual polish but to validate whether the conversational model felt natural and whether users trusted the responses. We built our design prototype to test the conversational flow, focusing less on visuals and more on usability and trust.
HANDOFF & DEVELOPMENT
Unlike many projects where design work ends at handoff, here I played a dual role as both the UX designer and the integration developer. I documented flows, states, and edge cases in Figma providing artifacts that served as the blueprint for the build. On the development side, I leveraged by knowledge as EPM SME to help shape the integration architecture of the solution.
I also had the opportunity to present this MVP to senior leadership, walking them through both the design vision and the technical feasibility of the tool. This visibility reinforced the impact of the work and opened discussions about extending the assistant into client-facing solutions.
OUTCOMES
“Thanks for your tremendous work in building the Oracle EPM Gen-AI solution. Your energy, enthusiasm and attention to details is amazing !!”
- Deloitte Award
2 Step ↓
SUPPORT DEPENDENCY
Early feedback suggested that non-technical users were able to complete insight queries independently without assistance from analysts or IT support, thanks to the intuitive interface and context-aware logic.
~75% ↓
TIME TO INSIGHT
Based on a benchmark taken during researching report extraction and insight synthesis process in traditional BI systems, the new system allowed users to extract relevant insights from data in minutes.
This project marked my first hands-on experience designing for an AI-driven product. I learned how critical it is to manage user expectations in generative systems and ensure traceability in AI outputs, especially within enterprise environments. Working on the integration between GPT and structured financial systems gave me a new appreciation for the complexity of data mapping and integration and deepened my confidence in designing AI-driven interfaces.
Check out this Case Study Next!
or