Intro
AI meeting notes that clients actually confirm
Meridian Notes
Meridian Notes is a meeting workflow tool for financial advisors. It records a client meeting, generates an AI summary with per-claim confidence flags, converts it into calendared action items, and sends the client a plain-language version they must read, confirm point by point, and sign.
I designed and built the clickable prototype end-to-end — a five-screen flow that validates the core bet: no AI output reaches a client until a human verifies it, and no record counts until the client confirms it.
Challenge
Multiple meetings a day, and every recap is a liability
After each meeting an advisor must reconstruct what was said, extract the numbers that matter, turn commitments into tasks, and send the client a recap — manually, multiple times a day
Reconstructing notes across multiple meetings eats the advisor's evening
Dollar amounts, percentages and dates are exactly where memory errors happen
Recaps are sent, but never verified or acknowledged by the client
Wrong number damages trust
A recap with a wrong figure — even a small one — destroys the very client trust AI was meant to save time building
"…your retirement target of $3.1M remains on track, and we agreed to trim NVDA from 18% → 10% over two quarters."
⚠ Was it $3.1M — or $3.7M?Verbal advice has no audit trail
Advice given verbally with no written, client-acknowledged record is a compliance exposure for the firm
Solution
Turning every meeting into an accurate, client-confirmed record — without the evening of typing
A five-screen gated flow. Each step unlocks only after the previous one is done — the compliance value depends on the order
Screen 1 · Today
Prep is pushed to the advisor, not searched for
The day starts with the schedule, a hero card for the next meeting, and a week board of action items across all clients — this is the live prototype, click around
Prep card
Personal conversation hook
Last-meeting topics, a live portfolio donut, and one personal conversation hook per client. When social lookup fails, it falls back to the last CRM note with a visible warning
Daughter Emma just got into Berkeley — congratulate him before diving into numbers.
✓ Social lookup — instagram.com/m_michaelchenWeek board
Weekly action items
Action items are first-class objects shared between the Today board and the post-meeting Actions screen — edit in either place, it's one task
Thu 10
Enter scheduled NVDA sell-down (18% → 10%,…
Fri 11
Transfer $95,000 → 529 money-market sleeve
Screen 2 · Record
Numbers are verified in the moment, not after
A live transcript auto-scrolls during the meeting. Dollar amounts, percentages and key ages are highlighted as they appear, so the advisor can catch a wrong figure while the client is still in the room
Live transcript
Highlight what matters to you
- Figures auto-highlighted in the advisor's chosen color
- Recording is disclosed, consented, encrypted at rest
- Transcript feeds the summary's provenance trail
"So the plan keeps your target at $3.1M by age 65."
"And the NVIDIA position?"
"We trim it from 18% down to 10% over two quarters."
Pause
Pause is a compliance feature
Pause isn't a convenience — it's labeled as an off-record sidebar, letting advisor and client go off the record deliberately and visibly
Screen 3 · Summary
The AI states its own uncertainty — and the UI refuses to proceed
Every key point carries a category tag, a confidence level, and an expandable "What was said" provenance trail. Low-confidence claims render in red and each must be human-verified before the flow continues. This is the product's core mechanic — this is the live prototype, click around
Multilingual
Send the summary in another language
- Language picker offers English, 中文 (Mandarin), and Español
- Advisor's working view stays in English; only the client copy translates
- Translation generates at send time and ships alongside the original
Provenance
What was said
- Every key point expands to the transcript context it was pulled from
- Lets an advisor check a claim without re-reading the whole transcript
- Same mechanism supporting flag resolution: expand, compare, then verify
Trim NVDA concentration from 18% → 10% of portfolio over two quarters via scheduled sales.
Accuracy gate
AI confidence-tagged
- Each key point carries an investment-category tag and a confidence level
- Low-confidence points render red and need "Mark verified" before send
- "Approve & generate action items" stays disabled until every flag clears
Retirement-at-62 target requires ~$3.1M at a 4% withdrawal rate; base case crosses at age 61.
Relationship touch
Suggested reading
- Articles relevant to the meeting's topics, attached with one checkbox
- Shared with the client alongside the confirmed summary
- A static list in V1 — nice-to-have, not a required step
Allocation
Today vs. agreed target
- Current holdings plotted against the target agreed in the meeting
- Direct-labeled, so drift reads at a glance without a legend
- The advisor chooses whether this chart is shared with the client
US Equities
Current : 52%
Target : 45%
Fixed Income
Current : 22%
Target : 30%
Real Estate
Current : 12%
Target : 12%
Cash & Equiv.
Current : 14%
Target : 13%
Retirement goal
Live retirement projection
- Goal and monthly-contribution sliders recompute the projection live
- If the goal isn't reachable by 65, the caption prompts to adjust either input
- An in-meeting tool the advisor can walk the client through in real time
Screen 4 · Actions
Commitments become calendared work automatically
Approving the summary generates editable action items with due dates and reminders — inline edit, add, delete, and two-way calendar sync
Generated tasks
Editable, scheduled, reminded
- Inline text edit on every item
- Due-date picker scoped to the current week
- Per-item reminder toggle
Design decision
Shared state with the Today board
These are the same task objects the advisor sees on tomorrow morning's week board — the meeting's output lands directly in the daily habit loop, not in a separate list that goes stale
Two-way sync with Apple & Google Calendar
✓ 3 items scheduled this weekScreen 5 · Client view
Confirmation is per-point, because "I agree" means nobody read it
The client opens a PIN-gated page with a plain-language, first-person rewrite of the summary. They confirm each point, can dispute any point back to the advisor, ask questions, and sign by typing their name — producing the signed acknowledgment stored to the compliance audit trail
Access
A PIN, not another password
Low-friction access for a possibly elderly or non-technical client base, with an error state and a "Text me my code" recovery path
Enter your 4-digit code to open.
Plain language
Two versions of the truth, on purpose
"Trim NVDA concentration 18% → 10%" becomes "Sarah will begin a scheduled reduction of my NVIDIA stock…" — jargon-free, first-person copy is what makes confirmation meaningful
This isn't right — remove & send back to Sarah
Dispute path
Disputes are a feature, not a failure
Any point can be removed and routed back to the advisor with undo. A nonzero-but-low dispute rate is healthy — it proves clients are actually reading
"My retirement savings target remains $3.1M by age 65."
Audit trail
The record that closes the loop
"Signed acknowledgment stored to the client file (compliance audit trail)" — plus a next-meeting card with a real .ics download and Google Calendar link
Signed acknowledgment stored to the client file (compliance audit trail)
Next meeting · Tue, Oct 7 · 10:00 AMMetrics
One north star, and a metric that keeps the gate honest
North star: client confirmation rate — percent of sent summaries the client opens, checks every point, and signs within 7 days. It's the behavior the entire client screen is built to produce
Flag resolution integrity
Do advisors actually edit flagged claims, or click through? Detects whether the accuracy gate is real or theater
Dispute rate
Low-but-nonzero is healthy (clients are reading); rising means summary quality is slipping
Action item completion
Percent of generated tasks completed by their due date — proves the meeting output turns into real work
Reflection
What we'd validate next
The prototype validates the workflow, not the ML. Before investing in real speech-to-text and confidence scoring, four questions decide the roadmap
Riskiest assumptions, in order
- Does the accuracy gate produce real verification — or rubber-stamping? Seed summaries with one genuine error and watch 5–8 advisors
- Will clients complete per-point confirmation, or check-all in 20 seconds? Pilot with 15–20 real clients
- Is the personal hook welcome or creepy? If clients react badly, restrict to CRM-sourced notes only
- Can any STT vendor get dollar figures reliably right? If not, V2 leads with in-meeting tap-to-correct instead of post-hoc flags
Trade-offs I'd defend
- Gated wizard over free navigation — the compliance value depends on the order
- Blocking gate over advisory warnings
- Per-point confirmation over one signature
- Plain-language client copy generated from approved bullets, never independently
- PIN over full login — with rate limiting and real code delivery in production