Jonathan LoPorto | Staff Product Designer

Home

Work

Investor relations assistant

A source of truth for questions the CFO's team couldn't get wrong

*UI recreated for portfolio purposes -

All data shown is fictional

Built for Meta's Investor Relations team ahead of earnings calls. I was recruited to the project, given a bare tech demo, and designed the entire Gen1 product alone.

Problem:

The CFO's team needed a way ensure better preparedness for quarterly earnings interviews. Enterprise Products needed a flagship GenAI use case to justify a GPU investment. There was no existing design, just a tech demo and a tight deadline.

Outcome:

I designed a Q&A document builder instead of a chatbot: verified, source-cited answers instead of a conversation to re-read. It shipped, and helped secure GPU investment for the org's GenAI work.

Problem:

Outcome:

I designed a Q&A document builder instead of a chatbot: verified, source-cited answers instead of a conversation to re-read. It shipped, and helped secure GPU investment for the org's GenAI work.

Role:

Sole design owner, Gen1

Timeline:

2024

Role:

Sole design owner, Gen1

Timeline:

2024

The work was time consuming and error prone

The CEO and CFO take live questions from the media during quarterly reviews. It's the investor relations' team job to make sure those answers are accurate.

At Meta, this process included pulling all the relevant documents into a shared drive, kept constantly updated. They read through it manually, highlight, take notes in side documents, and build a running list of questions they expect to face.

That question list become the interview scripts that the CFO and Zuckerberg learn ahead of earnings calls.

Information could change right up to the call. Keeping the script's source of truth current, this close to airtime, was a real source of stress on the team.

Starting from zero

The ask that reached the team was ambiguous: help the IR team move faster with AI, and figure out how.

When I joined this effort there was just a bare-bones tech demo of a coded chat interface on an engineers desktop answering questions form a PDF file that lived in a Google Drive folder.

It was unformatted. No layout, no verification, no sense of what a trustworthy answer should even look like on screen.

Solution: A source of truth builder

Everyone's instinct pulled toward having it be a chat, since that's what every other internal AI tool looked like.

Everyone's instinct pulled toward chat, since that's what every other internal AI tool looked like. But, chat forces you to review the conversation thread. More than one person would be viewing the same project, which also means a chat thread be meaningless to everyone but the person who asked the question.

I created a source of truth builder: Team members add their questions as batches, and the tool builds a completely cited Q&A document.

An answer isn't a conversation

Ask one or many question, get structured answers, not a chat thread

Someone prepping for an earnings call wants a document of vetted answers, not a thread to scroll back through. I put the composer at the top, and let answers build into a list, helping to separate mental model from AI tool = AI chatbot.

Source verification and updates

We addressed feedback from the IR team about being able to verify results from an early prototype. We did a detailed competitive analysis of 5 AI tools on how they highlight sources and citations, and pushed further into verification flows.

You can drill down to answer. A citation opened the source document and highlighted the exact passage. One level deeper still opened a chat scoped to just that document. If the AI used the wrong source, reviewers can flag it and the answer updates seamlessly

Answer with inline source citation

Source document view - Highlighted cited passage

Document scope chat one level deeper

Removing a source updates the AI answer

Source documents kept changing after an answer was generated. A confident but outdated answer was worse than none for a team facing investors.

What mattered was the principle: an answer this consequential has to know when it might be wrong.

Aware of when data changes

The tool flagged affected answers and let users refresh sources in one action.

Results

The tool shipped

and was used by the finance team in time for 2024 H2 quarterly results, and received positive feedback.

Secured org-wide AI investment

This was our early and visible success signal for Enterprise AI, it helped secure a GPU investment for Enterprise Products.

Jonathan LoPorto