Disambiguation & Syntax Templates
Prompt suggestions and contextual controls to help users build effective queries
Intro
Overview
Bridging the gap between unstructured natural language input and structured relational data.
As agentic apps become the norm, users increasingly rely on natural language to query for data.
However, free-form NLP queries often suffer from ambiguity, and fail to clearly map user intent to the systems underlying data structure.
This project grew from simple query disambiguation for financial users into how get the answer right the beginning.
It’s part of a larger effort still to answer how do we support the users in the transition from manual interfaces to agentic products?
Impact
A new global component
- Designed prompt templates & syntax filters to address complex disambiguation and prompt guidance
- Shipped MVP templates & syntax filters in first release of Ledger Agent
- Collaborated with design system team to create standardized design pattern and specs
- Shipping in all 22 Fusion Agentic Applications across 4 Fusion products later this year.
Context
All about balance inquiries
General ledger
noun
a company’s log of all financial transactions flowing through their system every day. There are huge volumes of data, and lots of potential risk.
Month-end close checklist
Every month, accountants are responsible for the same set of accounts, performing analysis to ensure their accounts have no abnormal activity and all transactions are accounted for.
How were users previously performing balance inquiries?
Design Challenge
Problem
How do we support more natural queries with technical constraints?
An open input field promises the user to ask it anything. And for the most part natural language is easier.
But there is only so much the system can infer, and while some attributes translate well to natural language, others require precise specification.
Disambiguation is not enough
Providing assistance after something has gone wrong is insufficient.
Too many possible matches lead to inefficient disambiguation
Expensive backend processes with high risk of inaccurate responses
Slower task completion with extra clarification turns
Design Principles
Kept in mind
1
Quick access to standard terms and values without triggering disambiguation loops
2
Show the relationship between natural language query and the underlying data model
3
Present queries in reusable format that users can come back to
4
Multiple input touch points so templates are always accessible
Final Designs
Templates
Seeded prompt suggestions that help users shape requests into clear, effective prompts.
- Teach users how to structure questions to ask better questions faster
- Preserves the flexibility and ease of natural-language input.
- Prompts includes business objects as syntax filters, helping users refine or select the right values in context.
Syntax filters
Contextual controls that help users clarify, refine, or select business values within a prompt.
- They support structured objects such as Account, Entity, Period, Currency, Dimension, Source, Category, or Task.
- Syntax filters begin empty so users are not forced to choose values too early.
- As the user types and intent becomes clearer, they can suggest relevant values, narrow options, or help resolve ambiguity.
Touchpoints
Accessible from inputs on all pages
Recap
Next Steps
Expanding scope and functionality
- Real time auto detection of variables
- Predictive typeahead suggestions
- Expanding variables to include what app context or focus you are in
- Extending categories to support personalized content like recents and user created templates
Reflections
From one product to many
I love answering the question how do we support the users in the transition from manual interfaces to agentic products? And this is the perfect example of that. Users want the flexibility and freedom that natural language promises, but it becomes a difficult design challenge to make sure that the technology and design can back that promise up.
However designing for that gap is crucial, and it was a gap that many teams outside of ERP were facing. This took a project meant for a single product to 5+ ERPM applications, to collaboration with the Fusion Redwood Design System team to create a design pattern now being developed for all fusion agentic apps across HCM (Human Capital Management), ERPM (Enterprise Resource Planning), CX (Customer Experience), and SCM (Supply Chain Management).
Getting to propose a new design pattern that gets built was a super fun experience & there is lots more to come.
