Payflow AI Workspace
A finance and accounting AI workspace for asking questions, generating reports, and tracing every figure back to its source file, integration, and assumption, designed so a controller can trust a number before it reaches the board.
Finance, Accounting AI Workspace
A finance and accounting AI workspace where every answer traces back to its source file, integration, and assumption. The team credits that provenance with a 4× lift in time spent with the AI agents.
- Role
- Design Engineer + Product Designer (end-to-end)
- Team
- Two co-founders (CEO + CTO), a founding engineer, a junior designer I mentored, and direct client calls
- Timeframe
- 2H 2025 · shipped, iterated weekly with users
- What I owned
- Research, IA, interaction system, data-viz language, and the production-grade React build.
The problem
Finance teams don't distrust AI because it's wrong. They distrust it because they can't see why it's right.
In finance, a confident answer with no lineage is worse than no answer. A controller can't put a number in front of the board she can't trace to a bank feed, an invoice, or a stated assumption. Most finance AI buries the source in an export footnote, if it's there at all.
The cost is measurable: teams still lose 39% of their time to manual, automatable work1, reconciliation alone eats 20 to 50 hours a month, and half of teams take six-plus days to close.2 Automation earns adoption only when the operator can verify the output as fast as she could have produced it.
“I'll happily let software do the math. I will never let it do the math somewhere I can't see.”
Why now
97% of finance teams are experimenting with AI. Barely 1 in 7 has it running at scale.
Generative AI in finance is forecast to grow from about $2.9B in 2025 to roughly $26B by the early 2030s (~31% CAGR).3 97% of finance departments report some AI use, yet only ~14 to 17% have moved past pilots into core workflows.4 The blocker isn't appetite or accuracy. It's trust, auditability, and explainability. That gap is the design opportunity: the team that makes AI legible to finance operators wins the at-scale market.
Who I designed for
Maya is VP of Finance at a 60-person company: the human API between the systems and the board.
Maya runs finance for a Series-B company. She lives across QuickBooks, Stripe, Mercury, and Shopify, and turns six data sources into one board narrative every month. She isn't anti-AI. She's accountable. If a figure is wrong, it's her name on it.
Her jobs, in her words:
- "Answer the board's question before the meeting, not after." From a vague ask ("how's runway?") to a defensible number in minutes.
- "Let me check the work without redoing the work." See the source, integration, and assumption behind any figure in one click.
- "Don't make me babysit a chatbot." Recurring reports regenerate themselves and land in an inbox, already cited.
Note
Designing for Maya reframed the product: the hero feature isn't the answer. It's the trail behind the answer. Every screen had to make verification a one-tap reflex, not a research project.
From first principles
I started at the structure, not the screen: where does trust actually get earned?
Trust is earned at read time, so provenance can't be a modal you open later. It has to be a structural layer on every figure.
Low-fidelity wireframes
The flow, drawn before it was styled
The provenance-first structure, wireframed end to end: ask, attach context, get a cited answer, open it onto a canvas.
01 · Home
Ask Luca, or bring in context to ground the answer.
Real product surfaces, no scripted overlay
How Maya gets from a question to a board-ready number
A focus carousel across the real surfaces. Every frame makes the same point: the answer carries its receipts.
01 · Ask in plain language
No query syntax. Maya types the board's question as the board asked it.
The system, in full
One shell, consistent grammar, from a single question to a recurring board packet.
Every surface shares one layout grammar, so Maya moves between them without relearning where trust lives.
Dashboard: leadership KPIs, ask-this-chart on every widget.
Reports: board packets that regenerate on a cadence, pre-cited.
Context: the evidence base Luca draws on, dropped in once.
Settings: identity, team, and the accounting data Luca can reach.
Mobile-first, not mobile-after
The board question often arrives on a phone. The answer, and its receipts, had to fit there too.
The same provenance promise holds at 402px: composer thumb-reachable, charts reflowed, sources one tap away.
Mobile walkthrough
The same four moves, thumb-first
Ask, scan, verify, trust, at phone ratio. The answer keeps its receipts on a 402px screen.
01 · Ask from anywhere
Type the board's question, or tap a suggested starting point.
The deeper surfaces hold up too. Workspaces and the context library keep their structure on a phone, so Maya can pick up the operating model or drop a board deck between meetings, trail intact.
Workspaces: pick up the operating model or board packet anywhere.
Context: drop a board deck and it becomes part of Luca's evidence base.
What I built
A data-viz language and a provenance footer, both designed as systems.
One charting vocabulary. Waterfalls for revenue-to-profit walks, area for trends, donuts for composition, all on one restrained palette anchored on a single brand green. A chart never gets decoded twice.
Provenance as a persistent footer. Assumptions and Sources stay pinned to the report at the shell's header height, with an accent active state, never a hidden modal. Drilling back to a source is a reflex, not a hunt.
Ask-this-chart everywhere. Every chart and metric is a conversation starter, so verification and follow-up happen in place, not in a new thread.
Live components, recomposed
The system itself, pulled from the real app code
Not screenshots. The production React modules from /payflow, recomposed: the charting language, the KPI tiles, and the provenance footer, themed live to this page.
Revenue to net profit
Trailing six months
RevenueNet profit
Revenue
$4.2M
+12%Expenses by department
This quarter, in $k
Cash on hand
$1.84M
Gross margin
Up 3pts on the quarter
68%
Ask Luca
Ask anything about your finances
Context library
Evidence Luca can cite
Drop a board deck or export
PDF, CSV, or XLSX, up to 25MB
Revenue mix
By line of business
Subscriptions 62%Services 24%Marketplace 14%
Net burn
$310k
Provenance footer
Assumptions · 3Sources · 6
QuickBooksStripeMercuryShopify
Pinned to every answer. One tap from a figure to its source, integration, and assumption.
Cut from the build · A standalone 'Sources' tab
What we tried. Early on, lineage lived in its own tab: clean to build, easy to scope. It demoed fine because the demoer knew to open it.
Why it lost. In testing, nobody opened it. For a busy controller, provenance one click away doesn't exist. I moved it into a footer pinned to the answer. Verification had to be ambient, not a destination.
What it moved
Rapid iteration on the product drove a 4× lift in time spent with the AI agents.
This shipped. Fast, evidence-led iteration produced a 4× increase in time spent on platform with the AI agents in 2H 20255, verifiable in the interactive build at /payflow.
“Overhauled product through rapid iteration leading to 4x increase in time spent on platform with AI agents in 2H 2025.”
Note
The lift tracks the thesis: when verification is ambient, operators stop bouncing out to spreadsheets and stay in the product. The metric I'd chase next is time-to-trusted-answer: how fast Maya gets from the board's question to a number she'll stake her name on.
What I learned
Designing the trail mattered more than designing the answer.
Two more months
The temptation in finance AI is to make the answer prettier. The real work is making it checkable. Once I treated provenance as a structural layer instead of a feature, every other decision (the persistent footer, ask-this-chart, the shared data-viz language) followed almost automatically. People spend time inside a tool they can verify. If I restarted, I'd instrument time-to-trusted-answer from day one. The interaction model was right before I had the telemetry to prove it.
- Tidjane
Shipped with
A modern, front-end-forward stack
A production React app with a Recharts data-viz language: Cursor and Claude Code in the editor, Figma for the system, GitHub for delivery.
- Recharts
Shipped with
Notes
- 1
PwC Finance Effectiveness Benchmarking (median across ~1,000 companies): ~39% of finance resource time spent on manual, automatable tasks. Directional industry benchmark.
- 2
Month-end close benchmarks, 2025: cash reconciliation ~20 to 50 hours/month and the #1 time sink; ~50% of teams take 6+ business days to close. Manual-entry error rates run ~1 to 4%. Aggregated industry reporting.
- 3
Generative-AI-in-finance market sizing: ~$2.2 to 2.96B (2024 to 2025) growing to ~$13.8 to 25.7B by the early 2030s at ~31 to 37% CAGR (Grand View Research; The Business Research Company). Broader "AI in finance" estimates run higher ($38B to $190B by 2030, MarketsandMarkets). Third-party forecasts; figures vary by scope.
- 4
~97% of finance departments report some AI use, but only ~14 to 17% have reached at-scale deployment in core workflows; trust, auditability, and explainability are the cited blockers. Aggregated CFO/industry surveys.
- 5
"4× increase in time spent on platform with AI agents in 2H 2025" is a company-reported outcome, attributed to Payflow's Co-Founder & CEO. The supporting figures (one-click provenance, 7 surfaces, responsive to 402px) describe the design system as built and verifiable in the interactive representation at /payflow.
- 6
Lines of code and call minutes are the author's own estimates of effort as the lead Design Engineer on this work, not instrumented analytics. They reflect the production build and the research and demo cadence across 2H 2025.
More work
Other projects worth a look.
