A
AIOS Wiki
read-only · public mirror
Open AIOS
Wikiai-supportresearchai-support/research/06-phase-4-perspectives.md

Phase 4 — Multi-Perspective Reasoning (Compressed)

Hand-authored·14 min read·5 sections·Last edited May 12 by initial import·View history

Sarah Chen — Skeptical Seed VC — VERDICT: SKEPTICAL

Persona: Decade in seed, ex-a16z, backs vertical SaaS in regulated industries. Believes most 2026 AI startups are zombies because LLM platform players will collapse 70% of app layer into native features within 24 months.

First-principles read (pre-brief): "We built it for ourselves" is the most over-represented failure pattern. AI customer support is the most crowded application-layer category. But "crypto vertical" is a real wedge hypothesis — Sierra/Decagon won't chase 40 logos in regulated headline-risk category. Initial thesis: $30-80M outcome at best, not venture-scale.

Structural problem: A wedge without a horizon. Crypto = 40 qualified ICP × 30% win × $150K ACV = $1.8M ARR ceiling. Even doubled = $4-6M ARR ceiling. Adjacent verticals (fintech/healthtech) already taken by Lorikeet ($35M Series A). Becomes $20-30M ARR feature, acquihire'd to Zendesk/NICE for 2-3x revenue. Forethought outcome at best.

What surprised her: HITL architecture is genuinely differentiated and survives the post-Setzer products-liability environment. Crypto truly has zero horizontal logos. Skills/corrections flywheel is a real defensible data moat. $1M ARR on $2-2.5M burn is genuinely capital-efficient.

Top 3 concerns:

  1. Org DNA / services-to-product transition (Pivotal Tracker base rate)
  2. Wedge ceiling — no expansion vector to $50M+ ARR
  3. Lighthouse customer pathology — Dapper is financially distressed, technically capable of self-build

Top 3 underappreciated opportunities:

  1. Agent-assist architecture is regulatory-tailwind native (EU AI Act, Colorado AI Act, Setzer, AI Exclusion in CGL)
  2. Skills/corrections flywheel as defensible data moat
  3. MiCA + GENIUS Act create forced-buy window (6-month sales window where buying is non-discretionary)

Recommendation: Option D (services-product hybrid). Don't raise venture round yet. Sell to 2-3 anchor crypto customers at high-touch deployment. At $5M ARR with three logos and ISO 42001, then spin out, hire external CEO, raise Series A from vertical SaaS specialist. Most likely outcome: Cognigy/Forethought-style strategic exit at $200-500M.

Term-sheet conditions if forced to invest: $4M on $16M post (20%), 2x liq pref, separate cap table, AZ capped at 15% strategic shareholder, external CEO hire as closing condition, milestones (SOC 2 by m12, two non-Dapper paid logos at >$100K by m18, named expansion-vertical thesis with design partner by m24).

Single question: "If Dapper Labs went bankrupt tomorrow, what does the next 18 months of this company look like?"

Marcus Tan — Visionary Founder — VERDICT: EXCITED (conditionally)

Persona: 3rd-time founder, $80M + $220M exits, crypto-native KYC/AML expertise. Believes vertical AI in regulated/painful spaces is the most undervalued opportunity of 2026.

Hidden asymmetry: AZ has 14-year proprietary dataset and operational muscle for crypto support. Sierra cannot acquire this; Decagon cannot replicate in 18 months even with $500M more. Operator-builder-investor combo in exact vertical with most ARR-per-buyer leverage. Same pattern as Plaid, Stripe — "built for ourselves" is the most underweighted moat in B2B SaaS.

The 10x vision (36 months): Brand "Sentinel" or similar (telegraph compliance, NOT AI mascot). Default AI support layer for any company touching crypto rails, stablecoins, RWA, tokenized assets. Mentioned same breath as Chainalysis (compliance) and Fireblocks (custody) — NOT Zendesk. 22-28 paying logos. $35-45M ARR by month 36. NRR 135%+. Acquisition: $1.2-1.8B by Chainalysis/Fireblocks/Salesforce/Zendesk, OR IPO 2030 at $4B+.

Pricing trick: As "AI support" = $200M business. As "compliance-grade AI support layer for regulated digital assets" = $1.5B business. Same product, 7x multiple.

Wedge sequence:

  • M0-6: Sign Dapper at $300-400K ACV with lighthouse rights (NOT revenue, case study). Hire Compliance Engineer (ex-Persona/Chainalysis/Fireblocks). Ship SOC 2 + ISO 27001 + ISO 42001 in flight. VPC-deploy. Spin out — separate cap table, brand, leadership.
  • M6-12: 4-6 logos. NOT Binance/Coinbase/Kraken (license components only). Target Gemini-tier, Bitstamp-tier, Bullish-tier. Ship MiCA Travel Rule + OFAC + KYC-aware module before July 1, 2026. Hybrid pricing $150-300K floor + $1.50-2.50/HITL ticket + 1.5x cap.
  • M12-24: Adjacent regulated (institutional custody, tokenized RWA, fintech-with-crypto, fan tokens, sports prediction). Ship voice channel (non-call-center). KYC re-verification co-pilot.
  • M24-36: Exit decision — sell to Chainalysis/Fireblocks at $1.2-1.8B (most likely), OR Series C at $1.5B post pushing toward $100M ARR, OR strategic merger with Lorikeet for regulated-vertical leader at $2.5B+ in 2029.

3 capabilities Sierra/Decagon can't replicate in 18mo:

  1. Compliance-Native Action Layer — every action auto-logged as compliance event with OFAC/Travel Rule/SAR metadata, surfaced into Chainalysis/Elliptic/TRM
  2. Hacker-vs-Victim Disambiguation Layer — trained on 30-50% inversion problem (stolen-account claimants ARE the hacker)
  3. Regulator-Ready Audit Bundle — one-click export for MiCA/GENIUS/state regulator subpoena

Anti-patterns:

  1. Trying to be horizontal AI support tool with "we also do crypto" splash page
  2. Letting AZ-the-agency culture leak into product company
  3. Anchoring pricing to Dapper

The boring truth: The Customer Success Compliance Lead role. Single named human at AZ owning each customer's compliance posture, attending AML/legal reviews, shipping custom regulator export. Sierra outsources this. Nobody at AZ has drafted the JD. Company that staffs this in month 4 wins.

Single question: "Are you willing to spin this out with separate cap table, separate brand, and a CEO who is allowed to fire AZ as a vendor on day one — or is this an AZ business unit dressed up as a product company?"

Mark Kingston — Power User — VERDICT: CAUTIOUS

Persona: The actual support agent at Dapper. Last general agent (Britney left, Dan to product, Kenny VIP-only). 800 unsolved tickets pile up nightly.

A day in life: 8:15 → 73 new tickets. Triage: macro-and-close (~30), needs-investigation (~25), oh-god (~18). Macro→deep-dive→escalate-to-Kenny-who-doesn't-respond rhythm. n8n workflows catch most patterns but missed a Lagos-IP fraud last week (workflow auto-approved Persona-clean account that started draining within 4 minutes).

5 ticket types he'd happily delegate:

  1. Password reset (non-2FA, non-suspicious)
  2. "Where's my pack?" / reveal-delay
  3. Account closure / GDPR-CCPA deletion
  4. Marketplace listing/de-listing questions
  5. Standard 2FA reset (Persona-clean, device continuity, no withdrawal flags)

5 ticket types he'd NEVER trust to AI alone:

  1. Hacked/stolen account claims (30-50% are the hacker)
  2. Withdrawal disputes with Flow Scan tracing
  3. Disney Pinnacle ANYTHING (contract violation)
  4. Disputed minor purchases / parental refunds (COPPA, age laws)
  5. OFAC/sanctions/KYC-fail edge cases

What he actually needs (the unsexy list):

  1. Keyboard-driven everything (j/k/a/e/r//)
  2. Bulk actions on similar tickets
  3. Real escalation queues with SLA timers
  4. Audit trail that survives discovery
  5. "Why did the AI suggest this?" button (show retrieval, rule, data)
  6. Snippet/skill versioning + 30-second kill switch
  7. Native "this AI is wrong" feedback loop with visible behavior change next ticket

What Justin and Dan don't understand: The marginal ticket isn't the median ticket. AI handles median fine. The marginal ticket — 45-min one — is where customer trust, regulatory exposure, IP-licensor relationships live. Replace Mark with median-tuned AI: save $80K, lose $50M Disney contract because bot said something dumb on Pinnacle.

Honest reaction to AZ tool:

  • ✅ "Predicted-next-action" was the ohhh I get it moment — first vendor that models workflow, not just retrieval
  • ✅ HITL with human signature default — every other vendor either has bot sign as itself (Air Canada precedent) or "suggestion" mode
  • ⚠️ Skills/corrections is a black box — needs skills inspector + rollback
  • ⚠️ Demoed on Toby (productivity tool), not Dapper (50K angry crypto users with KYC PII + OFAC exposure)

Disney specifically: Per-IP-licensor routing rules with provable enforcement. NBA/NFL/UFC/LaLiga writing similar clauses into next renewals. Marketing angle: "We're the AI support tool that won't get you sued by your league partner." Sierra/Decagon don't have this because their target customers don't have league partners.

3 deal-breakers:

  1. Audit trail that legal will sign off on
  2. Per-IP-licensor / per-jurisdiction routing with hard guarantees
  3. <2 second latency for agent draft

3 surprises (AZ might have right):

  1. Predicted-next-action models workflow not retrieval
  2. HITL default with human signature
  3. Skills auto-extracted from past tickets with corrections

Single question: "When the LLM drafts a reply that would have violated a Disney IP clause, an OFAC rule, or a state ADMT opt-out, what specifically prevents it from reaching the customer — prompt instruction, regex, hard-coded routing rule, or separate compliance model — and can you show me the audit log of a time it caught itself?"

Decagon CEO — Competitor — VERDICT: SKEPTICAL with CAUTIOUS undertone

Persona: Decagon CEO. $35M ARR, +283% YoY, $4.5B post on $381M raised. Biggest risk: compression by Anthropic/OpenAI/Zendesk.

Initial reaction: Pulse didn't move. AZ is artifact, not company yet. Single deployment (Toby), lighthouse customer fighting for survival (Dapper), venture-studio cap table. Not threatened today, possibly interesting acquihire in 18 months, useful signal to revisit crypto thesis.

Why no crypto customer yet:

  1. TAM math — entire vertical $40-80M ARR ceiling. Investors didn't underwrite $4.5B for $80M TAM.
  2. Brand contagion — OFAC settlement precedent + Travel Rule + 30-50% hacker-as-victim problem. One bad headline torches Notion/Affirm/Chime relationships.
  3. Buyer behavior — crypto-native build their own, run lean, CFOs treat $50K floors as offensive.

What would change calculus: Single Tier 1 anchor (Coinbase, Kraken, Circle) signing $1M+ ACV with public reference willingness.

What AZ has Decagon doesn't:

  1. Deep DB integration as default vs configuration
  2. Copilot-first UX with HITL as primary mode (post-Klarna reversal validated)
  3. Crypto-native domain knowledge baked in (workflow library + lived-experience taxonomy)

3 ways to crush AZ:

  1. Build crypto vertical pack — 4 engs × 2 quarters × $1.5M. Etherscan/Solscan/Flow Scan + MiCA + OFAC + 2FA recovery + Travel Rule + KYC handlers
  2. Acquire Lorikeet ($35M Series A, fintech-vertical, one step from crypto). 80% of crypto for free
  3. Partner with Coinbase or Circle — sign Tier 1 anchor at discount, public reference

Acquisition calculus:

  • $8-15M ARR + 5-8 crypto logos including Tier 1 + 80%+ logo retention + productized codebase = $80-150M (8-10x ARR with vertical premium)
  • Below $5M ARR = acquihire $10-25M
  • Above $30M ARR = too expensive, just compete
  • Structure: 50% cash / 50% earnout tied to crypto-vertical ARR retention at 24mo

Existential risk to Decagon: NOT crypto. HITL-with-skills-flywheel pattern becoming dominant UX paradigm AND AZ open-sources/productizes the skills layer = Cursor-for-support play. CS leader runs AZ for 90 days, accumulates 200 skills, switching-cost locked in a way autonomous-resolution doesn't replicate. If AZ horizontalizes the skills layer at $30K ACV against Decagon's $200K floor, Decagon has to defend core market. The move that hurts: not vertical-wedge, but UX-wedge with defection-cost flywheel.

Sierra reaction: Bret won't even glance at AZ. Hunting elephants ($1M+ ACV voice deployments). AZ TAM is irrelevant to Sierra's $10B math.

Recommendation to Anthropic: Don't ship "support agent" first-party. Ship primitives — HITL middleware, canonical skills-extraction API, prompt caching tuned for support traces, audit/compliance tooling for Article 50.

Single question at SaaStr 2026: "What's your renewal rate on customers that aren't Dapper, and how many do you have?"

Priya Bose — CMO/GTM — VERDICT: CAUTIOUSLY EXCITED

Persona: MongoDB → Notion → vertical AI startup ($0→$20M ARR in 14mo). Distribution > product is real. 90% of B2B AI founders blow GTM by starting with "best product wins."

Narrative position (the one sentence): "The only AI support agent built for support teams that handle money, identity, and irreversible transactions — deployed inside your VPC, supervised by your humans, and audited for crypto, fintech, and regulated commerce."

Anti-narrative (fatal): "We built a better Decagon — with slicker UI and human-in-the-loop." Puts AZ in feature comparison vs $4.5B / 100+ engineer company. "AI-powered customer support, reimagined" is a Sierra ad — don't audition.

Hero customer: VP Customer Operations or Head Trust & Safety (NOT CIO/CFO). Dual-hat: half ops, half compliance. 80-800 employees, $30M-500M ARR, crypto-native/fintech/regulated. "I am one bad ticket away from a Reuters headline." Reads The Block, Bits & Bips, a16z crypto blog, CCW Digital, Support Driven Slack. Findable via one good podcast + one ex-colleague reference. NOT findable via paid LinkedIn ads.

Wedge story (3 paragraphs):

  1. Coinbase 2025 breach + MiCA enforcement July 2026 = budget unlocked. Sierra/Decagon/Fin can't enter. We solve crypto-native ticket workflows with HITL that survives audit + breach forensic.
  2. Sell to 30-40 mid-tier qualified ICPs first. NOT Binance/Coinbase. Target Kraken-tier-2, Bybit-tier-2, prediction markets, stablecoin/RWA, 6-8 collectibles operators. Dapper is lighthouse, not market. Market is the 40 logos watching Dapper to see if we're real.
  3. Once default support layer for 40-logo crypto ICP, narrative travels to fintech (Mercury/Brex/Ramp shape) → neobanks → regulated marketplaces (insurance/healthcare/aviation). At every step compliance + HITL + breach-resistant aren't claims, they're battle scars.

Channel 0→$1M ARR (founder-led, ~$80K hard spend):

  • Dapper anchor + reference (non-negotiable)
  • 1:1 founder outreach to 40 ICP names over 8 weeks (5-8% close = 2-3 customers)
  • 4 podcasts in 6mo: Bits & Bips, Bell Curve, The Breakdown, Support Driven podcast
  • ETHDenver + Token2049 SIDE DINNERS (12 VPs Ops), no booth, $20K each
  • Warm intros from Dapper alumni network (Sorare, VeVe, Fanatics, Coinbase)
  • One contrarian content piece: "Why Fin Resolves 34% of Crypto Tickets and Why That's the Problem"
  • Expected: 8-12 customers @ $80-250K ACV = $1-2M ARR over 14-18 months

Channel $1M→$10M ARR:

  • First AE at $1.5M ARR (crypto/fintech background, $180K + 0.5% equity)
  • Second AE at $3M; First SE at $2M (FDE = COGS)
  • Content engine flips on at $1.5M
  • Conferences scale up: Consensus, Token2049, ETHDenver as full sponsors at $4M+
  • Channel partnerships at $4M+: Persona, Chainalysis, TRM, Fireblocks — adjacency, AZ becomes their "support AI partner"
  • Paid: ZERO LinkedIn until $10M (most overpriced channel for support AI)

Anti-marketing:

  • NO demo videos with fake "Sarah from Acme Corp"
  • NO "94.3% deflection rate" hero stat (Decagon/Fin poisoned this metric)
  • NO paid LinkedIn for 18 months
  • NO "AI-first" or "agentic" in headline
  • NO G2/Capterra review-farming year one
  • NO Gartner Magic Quadrant chase until $15M ARR
  • NO category creation deck — earn category leadership by being undeniable

Conferences:

  • Tier 1 (year 1): ETHDenver, Consensus, Token2049 — sponsor 12-person VP-of-Ops dinner per event, NO BOOTH
  • Tier 2 (year 2): CCW Las Vegas, ICX, Money 20/20
  • Tier 3 (selective): Permissionless, Mainnet, Devcon
  • Skip: SaaStr, Dreamforce, Web Summit

Content slate months 1-6:

  • M1: "Anatomy of the Coinbase Breach: Why Support Tooling Is Now a Board-Level Risk" (3,500 words, founder byline, The Block guest column + LinkedIn)
  • M2: "VPC-Deployed Support AI: The Architecture That Survives a MiCA Audit" (engineering-led)
  • M3: "Why Fin's 34% Resolution Rate Drops to 12% in Crypto" (data-driven, names competitors, factual, RISKY high-leverage)
  • M4: Dapper case study, co-published with Justin
  • M5: Podcast tour (4 podcasts, 6 weeks)
  • M6: First "State of Crypto Support" annual report — 30 pages, original survey of 100+ ops leaders. THIS BECOMES THE MOAT. (Year 3, every analyst/journalist/buyer cites it. How MongoDB owned "developer database," Notion "all-in-one workspace.")

Pricing (3-tier, hybrid floor + meter + ceiling):

  • Starter $4K/mo: up to 1,500 res/mo + $1.50/over, ceiling $8K, SOC 2/GDPR/CCPA, email support — 50-200 tickets/day
  • Growth $12K/mo: up to 6,000 res/mo + $1.20/over, ceiling $24K, multi-team + custom integrations + audit log export + SSO, HIPAA available, ISO 27001, named CSM — 200-800 tickets/day
  • Enterprise $30K/mo+ (annual): VPC, ISO 42001, MiCA/Travel Rule, OFAC tooling, custom indemnification ABOVE 12-month-fees cap (competitive wedge), dedicated FDE + named architect

Brand: Standalone brand, NOT "AZ Support" or "Toby Support." Buyer trust requires focus. Crypto wedge needs vertical-coded name. Optionality on cap-table side. Spend $30K with naming agency. Working candidate space connotes trust under irreversibility.

Lighthouse strategy — break Dapper-dependency narrative by month 6:

  • Anchor #2: mid-tier prediction market (Polymarket, Kalshi, Hyperliquid)
  • Anchor #3: stablecoin issuer / RWA platform (Paxos, Ondo, Agora/Usual)
  • In every public asset: name three customers, never one
  • Justin = customer advisory board member, NOT face of product

Sales objection-handling — "we're evaluating Decagon":

"Smart — they're the strongest horizontal player and worth the bake-off. One thing to pressure-test: ask them to indemnify above contract value for an Air Canada-style misrepresentation claim, and ask which of their 80+ logos handles irreversible financial transactions in your regulatory regime. If you're handling 2FA recovery, KYC-aware tickets, or anything OFAC-touching, the bake-off you actually want is on the bottom 5% of your tickets — the ones that get you sued. We can stand up a 14-day pilot on those tickets specifically. Decagon will quote you 6 weeks to start."

Top 3 opportunities tech team is missing:

  1. Indemnification as marketing weapon — 88% cap at 12mo. AZ can offer 2x-3x cap for regulated, turn into homepage line.
  2. "Human button" mandate is a FEATURE, not constraint — CFPB + CA SB 243. Sierra's pure-automation narrative struggles. AZ's HITL is natively compliant.
  3. Disney "license-or-sue" precedent will spread to sports leagues, music labels, studios within 18 months. AZ's HITL architecture clears licensure bar by default — unlocks sports/gaming/entertainment.

Single question: "If Dapper churns in month 14, do you have a business — and if you don't, why are you building it?"