Turning a fragmented, months-long due-diligence process into a transparent GenAI research workflow built on trust and visibility.
TL;DR:
Marsh, a Fortune 500 asset and wealth management firm, relies on researchers to create investment and operational due diligence structured, compliance grade documents reports on asset managers. Marsh wanted to accelerate due-diligence process using AI.
What’s the problem?
Existing AI workarounds were not built for this. Producing one of these reports. Manually extracting and synthesizing factual information from dozens of documents, then adding researcher narration on top. Nobody could tell what was written by a person versus generated by AI, and citations could not be checked making it difficult to trust the output.
Whats our solution
RateSense unifies that fragmented workflow into one AI-assisted platform, keeping a human in the loop. RateSense generates the report with clear attribution (written by AI or a researcher), citations tagged against every fact, and a confidence score that gives complete transparency. This meant researchers were able to easily generate and validate AI-generated reports, improving report turnaround from 6 to 8 weeks to 1 to 2 days.
My Contribution
Owned the 0-1 UX design for RateSense, from discovery through shipped delivery.
Discovery workshops with 15 SMEs.
Synthesized findings into a future-state service map unifying 3 fragmented legacy platforms and where AI could take over.
Partnered with 3 Gen-AI Engineers
to design 5 human-in-the-loop features behind the AI-Report creation: Attribution, citations, and confidence scoring.
Design & Shipped 2 end-to-end flows
Demoed the platform to senior leadership, including global managers, to secure buy-in for broader rollout
A peak into my Design file: Here’s what & how I shipped this project!
The Problem
Researchers had little to no trust because a single unverified A.I generated claim isn’t an inconvenience. It’s a compliance risk!
Lack of Verifiability
Researchers had little to no trust because a single unverified A.I generated claim isn’t an inconvenience. It’s a compliance risk!
This is an example from a well-known app
Invisible authorship
Finished report reads as one voice, with nothing on the page showing weather something is written by A.I or human. Which matters most during an audit.
by AI?
by a Researcher??
Lack of control
ChatGPT, Claude etc. are general-purpose AI assistants, Not made for iterative workflows. Researchers were stuck switching between A.I and word just to improve content for a portion of the report
The Solution
Unifying 3 disconnected platforms into an AI-assisted workflow. Every source gets pulled into Source library, and a report can be generated directly into Marsh’s standard template
Play
Play
Inspect Mode and Citations
Clicking into Inspect Mode surfaces all the meta-data behind a claim, along with a citation, confidence score covering language, tone, and how reliable that source actually is.
How this solves the problem
This gives a compliance reviewer exactly what the invisible authorship problem was missing: a clear, immediate answer to who wrote a given line.
Attribution
Hovering over any line shows whether AI wrote it or a person did. The distinction that used to be invisible across the whole report is now visible one hover at a time.
How this solves the problem
This adds a trust and transparency layer on top of the report. A claim is never just asserted, All the metadata is shown
Play
Inline editing and versioning
Iterate the AI-generated content without losing context generate multiple versions of the reports
Play
Select any part of the report, give it a quick instruction, and it updates in place. No full reprompt required.
Play
Lock a section and then regenerate the rest of the report from new documents without touching what is already approved.
Impact
RateSense turned a 6-to-8-week manual process
into a 1-to-2-day AI-assisted workflow
Pilot Adoption Rate (n=50)
Before company wide-roll out
Faster report generation
Compared to Industry standard AI tools
User satisfaction Metrics
20% Increase after iterations
