Financial Agents
Making AI Decisions Auditable
Institutional investors couldn't trust what they couldn't see. Redesigning the agent experience drove engagement across every metric that matters.
Impact
+489%
Agent activation rate
−35%
Support queries about agent behavior
+28%
Report re-engagement
+153%
Worklog open rate
Measured over 60 days post-launch vs. prior baseline
Role
Product Designer
Company
Boosted.ai
Year
Q1 2026
Team
PM, ML Engineers, CS
Context
Boosted.ai is a Canadian and U.S.-based company building AI-powered research tools for institutional investors. In 2025, we launched AI Analysts — autonomous agents that run recurring financial research and deliver updated reports on a schedule. The value proposition was clear: set it once, receive fresh intelligence every week.
But users had no idea what the agent was actually doing between runs. The product functioned as intended. Trust did not.
The Problem
The V1 worklog was built for engineers, not investors. It exposed raw function names, execution timestamps, and API latencies, which was meaningless to a portfolio manager deciding whether to act.
Users couldn't understand how outputs were generated or how calculations were derived. They ignored the worklog entirely because it did not provide value. The issue was simple: users couldn't verify the agent's work, so they didn't trust the output enough to use it.
"Legibility is a trust feature"
Users weren't asking for more data. They needed confidence in what the AI did.
"Two kinds of transparency"
How the agent works (plan) vs. what it did on a specific run (history). Different needs, different surfaces.
"Temporal accountability"
Users needed to compare runs over time. V1 had no history at all.
Constraints
Technical
Agent plans are LLM-generated dynamically. The UI had to flex around unpredictable step counts and tool types.
Bandwidth
Limited eng capacity meant no infrastructure rebuilds. Design had to work within the existing logging system.
Users
Professional investors expect precision and density. A simplified "it's working!" UI would lose credibility immediately.
Research
I interviewed eight analysts across four firms and partnered with Customer Support, who had been handling agent-related questions for months and had the clearest view of recurring issues.
Three patterns emerged:
Institutional-grade explainability
Institutional investors don't evaluate outputs the way typical chatbot users do. They need to audit the reasoning: what inputs were used, which formulas applied, and the exact date ranges behind the result. Without that, the output isn't actionable.
"I need to see how it got there, not just the answer."
Two kinds of transparency
Configuration (how the agent is set up) and execution (what it did in a specific run) are separate concerns. Users need both, but in different contexts. V1 merged them into a single stream and failed at both.
"I want to understand how it works overall — but when I check a run, I just want to see what happened that time."
Time is a trust dimension
Trust requires consistency over time. Investors compare runs to detect drift, validate outputs, and support compliance. V1 had no historical view, making that impossible.
"If I can't show what changed week to week, I can't rely on it."
Strategy
I decided to focus on transparency. Users needed to understand how the agent arrived at its answer: which inputs were used, which formulas applied, which date ranges mattered. They needed to verify the logic themselves.
To make this possible, I restructured the agent experience around three distinct surfaces. Each surface isolates a specific concern so users can quickly find the answer to one question at a time — avoiding the confusion of tangled information. This separation also makes the workflow auditable and the interface trustworthy.
Report
What did it find?
The report with its summaries and findings from the current run.
Workflow
How did it get there?
Step-by-step details: which tools ran, which inputs were used, and what each step produced.
History
What has it done over time?
A scannable log of past runs with key insights on top, letting users track consistency and spot drift.
Before proposing this, I mapped all questions from research and support calls against the V1 structure. No single surface addressed a need cleanly — everything was tangled. Separating these three concerns was the clearest way to make the interface legible.
Design Challenges & Decisions
Challenge 1
How to make the log understandable?
The first issue was the language. The raw agent log was technically correct, but not usable. I worked directly with ML engineers to translate each tool into something a user could understand and trust. This meant going beyond naming, understanding what each tool actually did, what success looked like, and what signal the user needed from it.
The format we landed on: Tool Name: Plain-language outcome
For example: News Analyzer: Found 3,127 news developments
This structure preserves the specificity that professional users need while making the meaning immediately legible.
Before
After
Challenge 2
How to make the Workflow scannable?
Once the language was clear, the next problem was structure. How should this information be presented? I explored three approaches:
- 1. Raw tool names — fast, but still unreadable
- 2. Prose descriptions — clearer, but hard to scan
- 3. Tool name + plain outcome — structure and clarity together
Option 3 won in testing. Analysts could scan the list quickly for anything unexpected, and the plain-language column gave them the confidence layer they needed.
Challenge 3
How to fit the runs inside history tab?
Each run had 20+ steps, making full visibility impractical. Testing showed users primarily cared about the main insight per run, and only drilled into details when needed.
I collapsed each run into an accordion with the main insight as the header, keeping full logs accessible on demand. This made it easy to scan differences across runs without overload.
Challenge 4
What to show during execution?
The initial loading state showed only the current tool, leaving users unsure how long the process would take. Many reported waiting without clarity on progress.
With the introduction of the Workflow tab, I shifted from a single loading indicator to full visibility. The interface now shows the entire plan upfront, along with the current step in execution.
This gives users a clear sense of progress and expected wait time, while also letting them verify the agent is working on the right steps.
What Users Said
"Now I can see what it looked at. I know it pulled the right stocks — I can defend this to my team."
Portfolio Manager
Royal Bank of Canada
"The history is what I actually needed. If it says no significant changes, I move on. If it flags something, I click in."
Senior Analyst
Global management consulting | McKinsey & Company
"It feels like a real analyst now, not a black box."
Investment Analyst
TD Bank
Results
60-day post-launch vs. prior 60-day baseline
What's Next
The History tab outperformed expectations. After release, users said it changed how they evaluated the agent's reliability. That signal drove the next phase:
"What's new" widget on Report Page
Surface the main insight from History tab directly on the Report page. The accountability layer shouldn't be hidden behind a tab — it should be the first thing users see.
Main insight on Home page
Every time the agent makes a new run, it creates an insight based on what changed compared to the previous run. We surface it on the home page so users can scan, identify what's most interesting, and click to research deeper.
One Year Later
A year after launch, behavior shifted. As tools like ChatGPT became mainstream, users no longer needed to audit every step. Trust moved from process to outcome.
Usage reflected this, and the Workflow tab declined. So we made a decision to remove it.
What changed
Removed the Workflow tab. Replaced it with lightweight hover overlays that show source and freshness inline — verification on demand, without a separate surface.
What drove trust instead
Embedded source links directly in the report: web sources, filings (10-K, 5-K), investor decks, and earnings calls. Every claim is traceable in one click.
Hovering a data point reveals its source, value, and up-to-dateness
Surfacing sources directly in the report (web links, filings, investor decks, and earnings calls)
As users mature and their expectations evolve, transparency needs to become more targeted. What builds trust early can become friction later.