Skip to content

Agents do the work. Advisors make the call.

We designed the AI wealth platform where agents draft the work and advisors approve it, from the task system to every chart.

Arca's task canvas: an agent finishes setting up accounts for a household, with the task and its actions on the left and the agent's summary on the right
Client
Arca
What they build
AI platform for wealth advisors
Our work
Product design, Design system, Mobile, Pitch deck
Timeline
Fall 2025 to May 2026

The company

Arca is an AI-native platform for financial advisors. Agents process meeting transcripts, draft client emails, create tasks and update records. Advisors review and approve anything sensitive, like a client email or a trade. It connects to the tools advisors already use, from CRMs to custodians.

The problem

The team had reached the point where building more without design might go in the wrong direction. Their co-founder called the task system almost the core piece: how advisors see the work assigned to them, and how human-in-the-loop work actually feels. One constraint shaped everything. The agent needs to see everything the user sees, so nothing could hide behind drill-downs or tabs.

Make the evidence visible before asking for trust.

The work, piece by piece.

01

The task canvas

Stacked task cards modeled on iOS notification stacks. Conversation context on the left, interactive components on the right, and auto-advance as each task completes.

02

Agent action review

One clear label, Action required or Auto approved, and one review button instead of approving every field. Updates show as before and after.

03

Citations everywhere

Each suggested field quotes the exact transcript line, email or document passage behind it, highlighted in place.

04

Diffs for every change

GitHub-style diffs for anything an agent proposes. Old value struck through, new value below.

05

Navigation the agent can read

Breadcrumb drill-downs became a flat, dynamic sidebar of intents the agent generates, plus global filters that highlight every component they touch.

06

Agent chat and workflows

Collapsible thinking blocks, a persistent review chip, stop, queue and interrupt controls, and a split view with stable context beside the agent.

07

A material system

Frosted glass panels and metallic buttons over day and night gradient landscapes, with exact blur and opacity rules.

08

The rest of the product

Meetings, inbox, calendar, contacts, a composable chart system, client onboarding, a document vault, mobile, and a full design system exported as variables.

Agent action reviewAn agent action review card for a brokerage transfer, with an agent note, a source citation, and approve, reject and edit controls
Frosted glass materialFrosted glass panels over a soft landscape, showing a holding's position details and performance chart
Composable chartsA weekly meeting breakdown: stacked bar chart, key numbers, client time list and a meeting-types donut

How it went

Two reviews a week, then weekly. When advisors started approving in bulk without reading, we redesigned review around fewer, clearer decisions. Ahead of launch the team moved to daily syncs, then we hardened the system: a line-item audit, type cut to a handful of styles, six standard dropdowns, and every radius and color exported as variables. Many components were built straight from our specs.

What other teams can learn from it.

  • A three-level model: workflow, task, actions
  • A workflow agent above task agents
  • Tasks go to agents by default and escalate to a person
  • Replanning mid-run, with cancelled tasks crossed out
  • A clear status system for tasks and actions
  • Every component built as a widget the agent can see
The last onboarding step: the company agent wraps up setup for a sample company and suggests what to ask it in SlackNext · AdaptA harness, designed by people who build one.

Building something like this?

Start a project