Focal by Four Bridges — Manav Hemnani

The impact

Voice recordings and WhatsApp messages become structured deal updates, tasks, and reminders — without asking partners to open a browser, fill a form, or change how they communicate.

The problem

Finance professionals communicate by instinct and voice. Every CRM, tracker, and task system was built for people who type. That behavioral mismatch is where deals fall through the cracks.

What I built

Reframed the brief first — the problem was data loss, not documentation — then designed and prototyped the full voice-to-CRM loop from scratch. 12+ months iterating inside the actual firm it was built for.

↺

The brief reframe

The original brief was: make CRM updates easier. The reframe: the problem isn't friction with the CRM — it's that updates never happen at all. Partners don't avoid the CRM because it's hard. They avoid it because the moment of information has already passed by the time they open it. Focal is built for a different problem.

01 — Impact

Zero behavior change.
Full operational visibility.

The finance professional who never updates anything now updates everything — because the update happens in the moment they're already speaking, not at the end of the day when they're trying to remember it. These are the outcomes Focal is built to deliver.

<30s

Voice input to structured deal update. Speak once. Focal does the rest.

Manual CRM fields required from partner-level users for standard updates.

100%

AI inferences reviewed before anything updates. Every change is proposed, not silent.

"The best workflow tool is the one that fits around the person — not the one that demands the person fit around it."

Design principle that shaped every decision in Focal

Mobile-first, deliberately

This isn't a desktop app with a mobile view. Every design decision started on a 390px screen. Finance professionals are on the move — between client offices, in cabs, on calls. The tool had to work in that context or it wouldn't work at all.

02 — Problem

Finance runs on voice.
The tools don't.

I spent months embedded inside Four Bridges. Partners ran Rs. 2–5 Cr deals entirely through calls, WhatsApp voice notes, and corridor conversations. Decisions were made in real time. Context shifted fast. The people who knew the most about any deal were also the ones with the least patience — and the least time — to update any system.

The result was a predictable failure mode: one analyst manually translating every spoken decision into spreadsheets and shared documents. Under pressure, that dependency cracked. Deals slipped — not from bad judgment, but from coordination failure. The information was there. The system just couldn't see it.

"The most valuable person in the room is also the worst at updating any tool."

Observed pattern across every financial boutique — the friction behind the firm

The problem wasn't a software gap. It was a behavioral one. Every project management tool, CRM, and deal tracker was built for people who communicate by text. Finance professionals communicate by voice. That mismatch is where deals fall through the cracks — and where Focal begins.

Partners won't open a laptop to log a deal update after a call Analysts spend hours manually transcribing spoken decisions Compliance-sensitive environment — AI errors in deal records have real consequences 3–5 person team with fluid roles and rapidly shifting deal state Finance-specific language that generic AI transcription doesn't handle well

03 — Solution

Speak. Review.
Confirm.

Focal is built around one loop: speak or upload, AI extracts and proposes, human confirms, system updates. Five screens carry that loop — the dashboard, voice input, AI proposals, tasks, and deal intelligence. Nothing in the interface exists outside of that sequence.

The central design principle: AI proposes. Human confirms. Always. The system never makes silent updates to deal records. Every inference surfaces as a proposed change with its source transcript visible — and it waits for explicit approval before touching anything.

Voice recording screen — tap and speak a deal update

Voice input — tap mic, speak the update. Or upload a recording. Or type a quick note.

Focal home — priority deals, AI changes bar, morning briefing

Dashboard — priority deals surface first. AI changes waiting for review. Morning briefing from the AI.

Tasks screen — AI-proposed items added after voice input

Tasks — AI-proposed items flagged amber. Confirmed tasks in the main list. Nothing silently added.

Deals screen — live deal cards with stage and risk indicators

Deals — live deal cards with stage progress, risk flags, and last-update attribution.

Voice, recording, or text — any input format

Speak a quick update mid-call, upload a recorded meeting, or type a note on the move. The AI handles all three formats the same way.

AI extracts tasks, deal changes, and risk flags

The system identifies what matters: stage changes, follow-up owners, deadlines, compliance signals — and surfaces them as discrete proposed items.

Human confirms before anything updates

Every proposed change requires explicit approval. The source transcript is always visible. In a compliance-sensitive environment, silent AI updates are never acceptable.

Priority surface — deals that need attention now

The dashboard doesn't show everything. It shows what's critical, delayed, or has pending AI changes waiting for review — so attention goes where it's needed.

Morning briefing — AI summarizes overnight context

Each morning, the system surfaces a plain-language briefing: what moved, what's at risk, and what needs action today — from the previous day's inputs.

Interactive prototype

Explore the full Focal interface

Voice input, AI proposals, deal cards, tasks, and insights — all interactive.

Open prototype →

Honest status

Active prototype. Not yet shipped.

Focal is currently being refined in real working conditions with the actual team it was built for. The voice pipeline, AI extraction logic, and confirmation UX have all gone through multiple rounds of iteration based on real usage. This is what 12+ months of working inside a firm actually looks like — not a shipped product, but a system that's earned its place in the workflow.

The Pivot — From Desktop to Phone

The product started as a desktop dashboard. It's now a mobile-first app. That shift wasn't a rebrand — it was the product finding where it actually belongs.

The earlier version of this product was built for a desktop browser. It had a command center, a deal tracker, an AI changes panel, and a full analytics view. It was well-designed. It was never going to get used.

The observation that drove the pivot was simple: the people who needed this tool most never sat at a laptop when they needed it. A deal update happens in a cab, on a call, walking between meetings. The desktop wasn't wrong — it was just not where the work happened.

🔄 Desktop app vs. mobile-first product — which is the right surface?

What we decided

Mobile-first, always. Not a responsive version of the desktop app — a product designed ground-up for a 390px screen, with the phone's native affordances (microphone, camera, notifications) at the center of the interaction model.

The desktop version tried to solve the right problem in the wrong context. Switching the primary surface from desktop to phone wasn't a design decision — it was an observation about where actual work happens, finally acted on.

The rebuild also gave us the chance to simplify. The desktop version had a lot of surface area — it tried to be everything. The mobile version started from a single constraint: what is the one thing this product has to do well? The answer was voice input to structured output. Everything else is secondary.

Product Architecture — The Core Loop

Focal is built around a single loop: input → extract → propose → confirm → update. Every screen exists to serve one stage of that sequence.

The architecture came from asking: what's the minimum viable interaction that converts a spoken thought into a structured record? The answer was a five-step loop that couldn't be shortcut.

The five-step loop

01 · Input

Voice recording, audio file upload, or quick text. All three formats enter the same pipeline. No special handling required from the user.

02 · Extract

AI transcribes and parses the input — identifying deal references, stage changes, task owners, deadlines, risk signals, and compliance flags.

03 · Propose

Extracted items surface as discrete proposed changes. Each one shows its source (transcript segment or timestamp) and a confidence signal.

04 · Confirm

The user reviews and approves or dismisses each proposed change. Nothing touches the deal record until explicit confirmation.

05 · Update

Confirmed changes write to the deal record, task list, and team activity log. The source input is preserved and linked for traceability.

Five screens carry this loop: the Dashboard (priority surface), Voice (input), Proposals (review queue), Tasks (confirmed + pending), and Deals (live deal state). A sixth screen — Insights — sits outside the loop and aggregates activity over time.

P01

Voice is the primary input — always

The central nav button is a microphone. Typing is supported but never the implied default. The product leads with the behavior finance professionals already have.

P02

Priority first — not chronology

The dashboard shows what needs attention now, not what happened most recently. Risk, delay, and pending AI changes surface before routine updates.

P03

AI proposes, human confirms — always

The loop requires explicit human approval before any deal record is updated. The architecture treats silent AI writes as a design failure, not a feature.

P04

Source always visible

Every AI proposal shows the transcript segment it came from. If you can't see why the AI proposed something, you can't evaluate whether it's right.

Key Decisions — Where the Hard Tradeoffs Were

The decisions that shaped Focal were mostly about what to leave out — and where to keep the human in control.

⚡ Should the AI update the deal record directly, or require human confirmation every time?

What we chose

Human-in-the-loop for every AI inference, without exception. No silent updates. No confidence threshold that bypasses confirmation. Every proposed change waits for explicit approval.

In a compliance-sensitive, high-stakes environment, the cost of a wrong AI inference reaching the deal record is higher than the friction of a confirmation tap. Trust in an AI system builds incrementally. A system that shows its work earns the right to do more — eventually.

🔔 Should the app surface everything or filter to priority?

What we chose

Ruthless prioritization over completeness. The dashboard never shows all deals or all tasks — only those with risk flags, delays, or pending AI changes. The full view is always one tap away.

Finance professionals have no patience for a dashboard that requires scanning. If everything is on the surface, nothing is. The product earns attention by respecting it — showing only what actually needs action.

📋 Onboarding — how much setup is acceptable before first value?

What we chose

Minimal onboarding, maximum first-session value. The onboarding flow imports existing deal context first, so the dashboard is populated before the user ever speaks their first update.

An empty dashboard is the fastest path to abandonment. The first session had to feel like the product already understood the firm — not like the user was starting from scratch.

Where It's Going — Current Status & Next Step

Focal is an active prototype being refined through real workflow observation. The next step is controlled — building a private internal AI voice agent to test, learn, and protect what matters.

Current status

The prototype is live and being used inside Four Bridges in real working conditions. Voice input, AI extraction, and the confirmation flow have gone through multiple rounds of iteration based on observed usage — not hypothetical feedback. The core loop works. The refinement now is around trust, speed, and the specific language patterns of finance professionals.

What's being refined

Real usage surfaces things no prototype session ever will: which AI proposals get dismissed and why, how long users take before confirming, which deal contexts the extraction model struggles with. Every round of real use makes the proposal quality measurably better — and makes the confirmation UX feel less like a checkpoint and more like a routine.

→

Next step

The next phase is building a private internal AI voice agent — not a public product, not a third-party integration, but a controlled internal system the team owns entirely. The goal is to test real daily usage, measure how the workflow improves, and keep all company data and deal context protected inside the firm. No sensitive business information exposed to external tools. No dependency on consumer AI infrastructure. A system that earns trust the same way Focal's confirmation UX does — one verified interaction at a time.

Focal —Voice-First Mobile CRM

Zero behavior change.Full operational visibility.

Finance runs on voice.The tools don't.

Speak. Review.Confirm.

Focal —
Voice-First Mobile CRM

Zero behavior change.
Full operational visibility.

Finance runs on voice.
The tools don't.

Speak. Review.
Confirm.