Agent Experience Design
SAAS / Finance
Designing a solution to support analysts and investment
managers across a wide range of financial use cases. This
agentic AI platform leverages the latest large language
models and supports more than 120 workflows. The
platform was designed by Acuity Analytics’ Design Team for
finance professionals.
As a Senior Product Designer, I led this project from 0 to 1,
conducting user research and interviews, facilitating
workshops, collaborating with cross-functional teams, and
creating interactive prototypes.
Timeline
7 month (On-going)
Teams
Engineers, Data Scientist, PM, BA, QA, Domain Experts
Tools
Figma, Cursor, Figma Make,
Acuity Assistance
Clients
Acuity Analysis
Acuity analysts supported clients across multiple financial services. However, 60% of their work was still completed manually, requiring frequent switching between platforms and consuming significant time and resources. In the fast-paced financial industry, analysts must work efficiently and respond quickly to avoid missing time-sensitive opportunities.
Why there is need of Building this, If so many other available in the market.
We came to this decision by conducting research and we wanted to offer something that is built for finance work, that understands the language used by Analysts.
Every client has their unique way of working and requirements; the work should be tailored to their needs.
Data Security First: Acuity treats client data with strict privacy; public platforms are entirely off-limits.
Users want total auditability. Knowing the exact origin of every data fetch isn't just nice to have; it's essential for building trust.
A number behind our
motivation for design
This was the first agentic solution for Acuity analysts, and it has very
good impact on the business and workflow.
Impact
Sales / Marketing Team
40%
Reduction in time spent on first drafts
75%
Accuracy in initial outputs
50%
Reduction in operational costs
2M+
Market sources analysed instantly
100+
Workflows supported
10x
Faster investment research and reporting
Design Process.
Primary / Secondary / Auditing / Observation
Internal audit
Audited 7 internal products and documented the workflows running through them
Mapped 120+ recurring workflows across 5 business units
Broke each one into its steps to find where the same patterns repeated
Competitive benchmark
Benchmarked AlphaSense, Rogo, Hebbia and other GenAI platforms built for financial services
Compared them on sourcing, traceability, output formats and workflow coverage
Identified where each one hands the work back to the analyst
Analyst interviews
Interviewed 25 analysts across 5 business units
Walked through their real workflows step by step, from data hunting to final upload
Probed what they check, what they don't trust and what they redo by hand
Understand their pain-point and things that can be easily automated
Ideated, Prototype and iterate
Prototyped the modular block layout and tested it with stakeholders
Ran moderated and unmoderated sessions with foundation clients on the beta
Fed findings back each release and tracked what worked, what needed fixing and what was still unresolved
Research / User Interview
Primary Research
When we began conducting interviews, they helped us validate our research hypotheses
and findings.
We only had the chance to interview analysts from a few business units; with more time we could have conducted more for a clearer picture.
Mapping Out Existing Workflow...
Research Findings
Yes, we heard, noted, observed, Sympathise and analysed. Here are the findings.
Automating repetitive tasks could save analysts up to 50% of their time. Every team wants automation, but each has a different workflow.
A spreading team and a company research team needed different steps,different sources and different outputs.
A survey of FP&A professionals found they spend only 25% of their time on value-added analysis, while 75% is consumed by routine administrative and data tasks.
In Bloomberg polls, 27% of senior UK finance decision-makers said a lack of explainability is why they neither use nor trust AI tools for financial use cases.
shhhhh!
A hypothesis that got proven wrong
Hypothesis
Our Data Science team and we spent considerable time improving the accuracy of the data, believing that higher accuracy would encourage users to trust and adopt the product.
Fact
Adoption depends heavily on analysts’ trust in the data, which in turn relies on the quality of its sources and the availability of reliable validation methods.
Research Gap: We have not yet interviewed any users outside the analyst group, which raises questions about the product’s scalability.
User Persona
I’ve been working as an analyst for the past eight years, and I have to say that hunting for and gathering data hasn’t changed much. It remains one of the most tedious parts of my job.
Investment bank · 8 years in financial research · Company and industry research
What success looks like for him ?
Spending more time assessing quality rather than collecting data.
As a delivery lead, I ensure high-quality, timely client delivery and identify data-driven opportunities for future investment. However, traditional processes often delay task completion and require additional resources as deadlines approach.
Runs a research delivery team · Owns Portfolio monitoring, data ingestion, etc
What success looks like for him ?
Delivering financial work with greater accuracy and identifying last-minute investment opportunities
Ideation
An Extension of Thoughts from a Creative Mind
Trust Layer
How might we let analysts
trust AI output as fast as
they can verify it?
Scale Layer
How might we design a
flexible Workflow that
supports 120+ use cases
without requiring a
separate template for
each one?
Control Layer
How might we automate
the repetitive work off
analysts without taking the
judgment away from them?
This project allowed me to think and work beyond the screen,
approaching the experience at an ecosystem level.
Setting up goals and Matrix
Business Goals :
A system should support a higher number of databases from the market at least 80%.
Business wanted to achieve a higher number of project completions with lower headcount.
Business head says: reduce manual intervention in the first launch by at least 30%.
User Goals :
Give analysts their time back by removing repetitive tasks from their workload.
Provide AI assistance with full context so personas can perform their work more effectively.
What is Success for this Product?
01
Analysts spend less time preparing data and more time analysing it.
02
Outputs withstand audits without requiring manual
re-verification.
03
New workflows can be added without rebuilding the system.
04
Teams continue using the solution after one time admin configuration.
Service Design
“Fin's Agent” – Human + Tech
Ideation
User Journey
Back-end Engine of the Platform
Yes, your reading experience matters to me too.
You made it this far—oops, you seem tired. Here’s some coffee, boss!
↓ Next
Designing the backend of an AI system to ensure
structured, high-quality results
The product experience is divided into two critical layers: Design Time and Run Time each serving a distinct strategic purpose.
The Design Time experience enables analysts and domain experts to define repeatable workflows using configurable blocks. Where Run time is Agent routes tasks to specialized agents based on source type and workflow structure.
Not only limited to interface
Testing
Design Iteration – Testing
Before finalising anything, we tested it.
imp: The screens shown above use the older design system.
Iteration
We tested and we got real feedback
We conducted moderated and unmoderated sessions with foundational enterprise clients across
multiple domains, testing against live workflows and real documents rather than sample data.
Worked
The option to configure multiple data sources from the Workflow Configuration screen worked well.
Users can easily navigate through and audit the findings for each data point.
When setting up a new workflow, users can easily configure a prompt dedicated to that workflow.
Creating a workflow during the design phase.
Needed work
Enhance the UX copy and give the screen a fresher, more polished feel.
Workflow configuration needs an instruction-based flow — even though the design was intuitive, users felt stuck.
In the first version we missed out on empty states of many screens.
Most issues involved clarity: people understood the system, but weren't always sure what a specific control would do.
User Suggestion
They wanted an option to check version history; they often needed to refer back to it.
They suggested refining the system’s jargon and more fintech focused training materials.
Visibility into the AI’s reasoning while it works.
Design System
We have enhanced Our Design System based on the company branding chnages
Enhancing the design system helped us introduce new components to support AI features and use cases that were missing from the previous version. We also aligned it more closely with our brand language.
Design Decision
Design that got shipped
01
Prompt based experience...
Value Proposition
Designing a unified system for diverse financial use cases saves users the time they would otherwise spend learning new tools for each task.
One way to perform 100+ Workflow
We designed a prompt-based screen to support a wide range of financial use cases. We gave it a dedicated screen because we wanted to position it as a standalone way to accomplish tasks - not as a bot or just another option for financial work.
02
The Fleet of Agent is on work / Runtime
Value Proposition:
The screen builds user trust from the outset by showing the analyst’s task status, task dependencies, and citations for each section. By making this information visible, it also gives users greater control over the process.
The system maintains transparency by posting live status updates for every agent.
This screen shows the model runtime, giving users upfront visibility into the sources used to retrieve data for a financial use case. It also illustrates how the workflow is divided into multiple subsections and how individual LLMs are assigned to specific tasks. Task dependencies are clearly highlighted in the status indicators, while the status of each section helps users estimate the total completion time. Source visibility and validation further increase users’ confidence in the generated results.
03
Set up your workflow before running anything.
Analyst can configure end - to end workflow for the better output, from section to the question and his trusted source
Workflow, company, sources, and questions are all configured on a single screen. Analysts can see exactly what the AI will use, including internal data, licensed feeds such as CapIQ and FactSet, uploaded documents, and specified web domains. The Template is also organised into predefined sections with standardised questions.
Value Proposition:
The extensive set-up of workflow ensuring that two analysts performing the same task receive answers in a consistent format. and it saves users time in repetitive task.
04
Every data is mapped to it;s Source
Value Proposition:
Frictionless data validation gives users control and trust in the information. Displaying source icons upfront provide transparency at a glance and confidence.
User can check any number or retrieved data against its source.
Every section shows how many sources it used. Claims carry small numbered tags, and clicking one opens the actual source document beside the output, searchable and paged. Nobody has to leave the workflow or go hunting to verify a figure.
Yes, it still has room for Improvement, Something that we missed.it
It’s important for AI-powered tools to be transparent about what they can and cannot do, so users clearly understand their capabilities and limitations.
The workflow doesn’t show users how long a task is expected to take. And how much time it has been taken.
No version history : We shipped with no way to look back at earlier versions
✺
Reflection
Fin'sAgent was built around the workflows of Acuity's in-house analysts, making it effective for that team but never testing whether it worked for analysts elsewhere.
We assumed users already understand finance. I’d also test it with people who are new to investing to see whether it can teach them, not just help them work faster.
✺
Learning
I enjoyed working on this project because it challenged me to think beyond the screens and understand how it could change the entire way we work.
It was interesting and challenging to understand the data journey, work closely with the Data Science team, and learn how we trained the system to improve data accuracy.
Thank's For your Time,
If you have any feedback for me, please send it to me : )