Audio Capture → Service Delivery OS

A Plaud-style device as data wedge, Donna/mufu as the intelligence engine — collapsing the cost of servicing and delivery for field sales organizations
19 MARCH 2026 · R2
Conv. Intelligence TAM
US$1.8B
2025, 10.3% CAGR
Donna Insight
3 months
operating proof of cost collapse
Bob Pilot
1,000+
WA msgs captured, live
Team Candidates
5
Bosco, Mitchell, Wenhao, Bob, Vincent

I. The Real Thesis

This is not "B2B Plaud" or "coaching for field sales." The thesis is deeper and comes from three months of operating Donna:

Cost collapse in service delivery. Full-context CRM (like Donna) enables instant quality replies for servicing. Extracted codified expert workflows make delivery automatic via LLMs. Together, these collapse the cost of both service (account management, client comms) and delivery (the actual work product) in any service business. The audio capture device is the data wedge — it captures the one thing that doesn't exist digitally yet: in-person field conversations.

The architecture:

LayerFunctionDonna Proof
1. CAPTURE Device records field conversations — data that doesn't exist without the machine WA capture already live (bob-1: 1,000+ msgs). Audio capture = the in-person extension.
2. CONTEXTUALIZE Full history of every client relationship: past meetings, commitments, preferences, open threads Donna tracks 100+ relationships with warmth scores, cadence, action items. Working daily.
3. SERVICE Instant quality replies, proactive follow-ups, client reporting — all powered by full context Bob's servicing pivot (Mar 18): AI-assisted reporting + automated first-reply = 10-100x retention revenue.
4. DELIVER Codified expert workflows execute the actual service automatically via LLM Malaysia franchise data pack, JobsDB CV screening, brand book — all LLM-delivered from captured briefs.

Why this matters: Every field sales AI competitor (Rilla, VoiceLine, Gong) stops at layer 1-2. They capture and transcribe. Some do coaching. Nobody connects the captured conversation data to downstream service delivery. That connection — conversation data powering automatic delivery — is the moat.

The FieldCheck Parallel

This is the same architecture Eric and Wenhao built for blue-collar AI:

FieldCheck (Blue-Collar)This (Field Sales)
Capture device Camera on phone (image/video) Plaud-style recorder (audio)
What it captures Physical state of locations (cleanliness, pests, gel application) Client conversations (needs, objections, commitments, relationship signals)
Why device needed People don't have that visual data without the machine People don't have that conversational data without the recorder
AI processing Gemini Vision → checklist verification Whisper/Deepgram → Donna context engine → LLM delivery
Downstream value Manager dashboards, worker verification, SLA compliance CRM updates, instant servicing, automatic delivery of work product
Token cost ~$0.04/visit (Gemini Flash) ~HK$90/user/mo (transcription + LLM)

Eric has already built and deployed this pattern. The modality changes (video → audio), the domain changes (blue-collar → white-collar sales), but the architecture is identical: device captures what's invisible → AI processes → delivers value back.

II. Market Context

Market LayerSizeSource
Conversation Intelligence (global)US$1.77B (2025) → $3.47B (2032)QY Research3
Sales Enablement Platforms (global)US$5.23B (2024) → $11.3B (2030)Grand View Research4
AI Note-Taking (global)US$740M (2026) → $3.48B (2035)Precedence Research5
HK Insurance Agents78,786 licensedHK Insurance Authority7

But these TAMs miss the point. The addressable market isn't "conversation intelligence" — it's every service business with field sales teams where the cost of servicing and delivery is dominated by human labor. That includes:

Plaud Proves the Hardware Layer

Plaud: US$250M ARR, 1.5M devices, profitable, self-funded1. 50% paid subscription conversion. But Plaud is consumer-only — no team management, no CRM sync, no downstream delivery. The hardware works. The intelligence layer is the business.

Competitor Landscape (Compressed)

CompanyWhat They DoWhat They Don't Do
Gong ($7.25B)6Virtual call intelligenceNo field/in-person. No delivery automation.
Rilla Voice ($3.7M seed)9Field sales recording + coachingUS home services only. Phone app. No downstream delivery.
VoiceLine (€10M Series A)10Voice → CRM for field salesEurope only. No hardware. No delivery layer.
Catch AI11Field recording for home services$250/rep/mo. No servicing. No delivery.
activatedonna.com13Field sales AI assistant (EU)Name conflict. No hardware. No delivery automation.
Plaud ($250M)1Consumer AI note-taking hardwareNo B2B. No team mgmt. No CRM sync. No delivery.
Nobody does capture → contextualize → service → deliver. Every competitor stops at capture + maybe coaching. The full-stack play — using captured conversation data to power both servicing AND automatic delivery of work product — has no direct competitor. This is because nobody else has built a relationship intelligence OS like Donna and then asked "what if we pointed this at a sales team?"

III. Unit Economics

COGS Per User/Month

ComponentCostAssumption
Audio transcriptionHK$12600 min/mo, Deepgram $0.0043/min17
LLM inference (summaries + coaching)HK$78GPT-4o-mini for summaries, Sonnet for coaching analysis18
LLM inference (delivery automation)HK$120Codified workflows executing reports, proposals, follow-ups via Sonnet
Cloud/storageHK$15~2GB audio + outputs per user/mo
Donna engine (context + relationship intelligence)HK$200Amortized across team. PCRM loop, warmth scoring, briefings.
Hardware (Plaud NotePin amortized)HK$58$179 device over 24 months
Support/opsHK$150At 50+ users scale
Total COGSHK$633/mo~US$81/user/mo

Revenue Model

TierPriceWhat's IncludedGross Margin
Managed Service (Phase 1) HK$3,000–5,000/user/mo Device + capture + servicing + delivery automation + weekly briefings 79–87%
Platform (Phase 2) HK$1,500–2,500/user/mo Self-serve dashboard + capture + servicing + delivery templates 74–82%
Death cost: Eric's time at <30 users. The managed service phase has high margin on paper but Eric's time is the hidden cost. At <30 users, it's profitable as a Donna-biz deployment but not scalable. The path to scale requires (a) codifying the workflows that Eric currently runs manually, and (b) a team that can handle onboarding/ops without Eric in the loop. This is exactly what the team formation below addresses.

IV. Pilot: Bob / Hopeman Group

Bob is the right pilot because the infrastructure is already partially live and the business need is crystal clear:

LayerCurrent StatusWhat Audio Capture Adds
CAPTURE bob-1 WA pilot live: 1,000+ msgs, 56 chats, Supabase logging In-person client meetings (一品雞煲, 高德 sessions) captured automatically
CONTEXTUALIZE CRM analyzed: 18 sheets, 216 clients, $1.29M. Donna context engine operational. Every client relationship gets full conversation history (WA + in-person)
SERVICE Veaky pivot (Mar 18): servicing quality = #1 priority. Upsell/renew = 10-100萬/client. Instant post-meeting follow-ups. AI-assisted client reporting from Veaky's Report Step .xlsx.
DELIVER Malaysia franchise data pack, brand book, PR pages — all LLM-delivered from captured briefs. Any client brief captured in-person → LLM delivers first draft of work product same day.
Bob's quote: "請一個人我訓練他三個月其實某程度上都是做設置" — hiring someone and training them for 3 months is basically setup. The AI agent does the same thing, doesn't quit, and scales. Bob already thinks in this frame.

Pilot Proposal

V. Team Assessment

PersonBackgroundRole FitSignal StrengthRisk
Bosco Lam
STRONG FIT		 Past K-Pay: managed field sales teams, saw sales ops bottlenecks firsthand. Pickupp alumnus. Currently at Farmio (with Paco). Product / Sales Ops Lead. Knows the pain from the manager side. Can design the team management layer because he's lived it. Interested in mufu concept ("just like Tokugawa government"). Caught up Mar 15. Warm relationship. Currently busy with Farmio. Would need to commit part-time or transition.
Mitchell Suen
GOOD FIT		 Lead sourcing founder. Sold qualified leads through cold outbound. Clean CTO buyout recently. Lead Gen / Outbound. His own pain (lead-to-conversion gap) IS the upstream problem this product solves. Knows how to generate pipeline. 61-message deep catch-up with Eric (Mar 18). In transition. Active in Tue/Thu TT group. Not yet in people.json (new/renewed contact). Commitment level unclear. In transition = available but also exploring.
Wenhao Dong
STRONGEST FIT		 Built FieldCheck (camera → AI for blue-collar). SG SMBs. Sales background. Smilie (gifting) founder. Co-builder / SG Market. Already built the exact pattern (device → capture → AI → deliver). SG SMB access. Most effective co-build with Eric. Blue-collar AI kill date Apr 15. Strategic fork unresolved ("tech is thin margin"). Awaiting Eric alignment. Needs to see this as the evolution of what they were building, not a pivot. The framing matters: same architecture, better market (white-collar margins >> blue-collar).
Bob
PILOT + TEAM		 R&B marketing agency. 216 clients, $1.29M. Deep F&B/SMB connections. Donna pilot live. First customer + BD arm. Already selling to SMEs. Every Hopeman client is a potential deployment. "If you can sell, I'll handle backend." WA pilot live. Servicing pivot. "請一個人我訓練他三個月其實某程度上都是做設置." Already running Hopeman full-time. Would need this to feel like an extension of his business, not a side project.
Vincent Li
IF FUNDRAISE		 Super connector (HK/SG/VN/HCMC). BD/sales at Incorp Asia (Hillhouse-backed). First organic Donna customer. SEA expansion + fundraise. HCMC is massive untapped field sales market. Cross-border BD skills. Investor connections. Asked about Donna for buyer search across countries. New job at Incorp (corporate services = field sales intensive). Only relevant at fundraise/expansion stage. Currently busy with Incorp + trading. Don't over-index on Vincent until product-market fit proven.

Team Configuration Assessment

The founding kernel that makes sense today:
  • Eric — builds the intelligence engine (Donna/mufu). Already running.
  • Wenhao — co-builds the field deployment (device + capture pipeline). Same role as FieldCheck. SG market.
  • Bob — first deployment + HK SME channel. Every Hopeman client = potential customer.
  • Bosco — product design for sales ops layer. Knows exactly what managers need because he was one.
Second wave (once pilot validates):
  • Mitchell — outbound lead gen for the product itself. His exact skill set.
  • Vincent — SEA expansion / fundraise. HCMC market. Investor access.

The Wenhao Conversion Pitch

Wenhao is the most important conversion. Blue-collar AI kill date is Apr 15. The strategic fork ("tech is thin margin — 50/mgr, 500/SME") is unresolved. Here's the pitch:

Same architecture, better market. FieldCheck: camera → captures physical state → AI verifies → delivers compliance report. This: recorder → captures client conversation → AI contextualizes → delivers work product + servicing. Same exact engineering. But the margin isn't $50/manager or $500/SME — it's HK$3,000-5,000/user/month, because white-collar service delivery margins are 10-100x blue-collar verification margins. The "egg" is not an egg anymore.

VI. What 3 Months of Donna Proved

This isn't theoretical. Eric has been running the intelligence engine at personal scale for 3 months. Here's what works:

CapabilityDonna at Personal ScaleDonna at Team/Sales Scale
Full-context instant replies Eric replies to 100+ contacts with full history. Never drops a thread. Every rep has full client history before any meeting. Post-meeting follow-ups generated in seconds.
Codified workflow delivery Malaysia franchise data pack, CV screening, brand books — delivered same day from a brief. Client proposals, reports, analysis — auto-generated from meeting transcripts using codified templates.
Relationship warmth scoring Dunbar circle management: 75/150 tracked with cadence and warmth. Account health scoring per rep. Pipeline risk detection. Churn prediction.
Daily review loop (PCRM) Every morning: what happened, what's next, what's at risk, who needs attention. Manager gets daily team briefing: who met whom, what was committed, what's at risk, who to coach.
Action item extraction Every conversation auto-generates tasks tracked in action_items.json. Every client meeting auto-generates next steps. Nothing falls through cracks.

VII. Red Team

Bull Case

  • Intelligence engine already built and running (3 months proof)
  • FieldCheck pattern already built (device → AI → deliver)
  • Bob pilot already partially live (1,000+ WA msgs captured)
  • Measurable revenue impact (servicing = 10-100x lead gen)
  • Team candidates all have relevant domain experience
  • Hardware is commodity (no R&D risk)
  • HK TVP grant covers 75% of client cost
  • No competitor does capture → service → deliver
  • Wenhao conversion aligns with his kill date timing

Bear Case

  • Eric juggling 5+ active projects — bandwidth is the real constraint
  • Managed service doesn't scale past 30 accounts
  • "Donna" name taken in field sales AI
  • Wenhao might not convert (different problem space)
  • Bosco busy with Farmio, Mitchell is new contact
  • Recording consent is complex (legal/compliance)
  • Plaud/Gong could add delivery features
  • Team formation takes 2-3 months before first revenue

Verdict

Strong — this is the natural evolution of Donna-biz, not a new product.

The thesis is sound: 3 months of operating Donna proves that full-context CRM + codified workflows = cost collapse in service delivery. The audio capture device is just the data wedge for in-person conversations — the same pattern as FieldCheck for physical inspections. The competitive landscape validates the market (Rilla $1M ARR year one, VoiceLine €10M Series A) but nobody is building the full stack from capture to delivery.

What makes this different from the R1 report: This isn't about selling hardware + coaching. It's about using captured conversation data to power automatic service delivery — the work product itself. Coaching is a feature. Delivery automation is the business.

Concrete path:

  1. This week: Equip Bob's team with 3 Plaud NotePins. Capture the next client meeting. Auto-generate follow-ups + first-draft deliverable from the transcript. Show before/after to Veaky's team.
  2. This month: Pitch Wenhao the conversion — same architecture, better market. Frame it as "FieldCheck for white-collar." Align on Apr 15 decision.
  3. Next month: Approach Bosco for product design input (sales ops layer). Approach Mitchell for outbound once pilot results are in hand.
  4. Month 3: If Bob pilot shows measurable retention impact, pitch Chaky (insurance) and use Bob as case study.

The one thing that changes the verdict: If Eric can't free up bandwidth from the current 5+ projects. This needs 15-20hrs/week of focused attention for the first month. EVR is 30K/mo cash. Sourcy is contracted. Essai is signed. This competes for the same time as all of them. The team formation (Bosco, Wenhao) is specifically designed to distribute the load — but the first month is on Eric.

References

[1] Plaud Revenue, Funding & Analysis — Sacra, 2025. $250M ARR, 1.5M devices, profitable
[2] Field Sales vs Inside Sales 2026 — Prospeo. Field sales workforce stats
[3] Conversation Intelligence Market 2026-2032 — QY Research. $1.77B TAM
[4] Sales Enablement Platform Market 2030 — Grand View Research. $5.23B TAM
[5] AI Note Taking Market to $3.48B by 2035 — Precedence Research. AI note-taking growth
[6] Gong vs Chorus 2026 — Delverise. Gong $7.25B, 24M calls/yr
[7] HK Insurance Market Statistics — Insurance Authority, Sep 2024. 78,786 agents
[8] Rilla Pricing Guide 2026 — SalesAsk. $4K+/user/yr
[9] Rilla $3.7M Seed — TechCrunch. $1M ARR year one
[10] VoiceLine €10M Series A — ContentGrip. 100+ enterprise clients
[11] Catch AI — gocatch.ai. $250/rep/mo field recording
[12] aiOla Voice AI for Field Sales — aiOla. Fortune 50 case study
[13] Donna — AI for Field Sales — activatedonna.com. Name conflict, EU product
[14] Dooly Shutdown — Recapped. Failed: note-taking alone isn't enough
[15] crmCopilot Failed (YC) — Reforgers. "CRM update" = feature, not product
[16] Vivun $131M Pivot — Upstarts Media. Sales tools stall without AI replacement layer
[17] Transcription API Pricing 2026 — PkgPulse. Deepgram $0.0043/min
[18] LLM API Pricing 2026 — DDaverse. GPT-4o-mini $0.15/$0.60 per 1M tokens
[19] Plaud for Business — plaud.ai. Enterprise positioning (no team features yet)
[20] Pharma 57% Improvement — Quantified AI. Field sales AI coaching ROI