A
AIOS Wiki
read-only · public mirror
Open AIOS
Wikiai-supportbriefingai-support/briefing/02-phase-2-brief.md

BRIEF FOR PERSPECTIVE AGENTS — Phase 0.5 + 1 + 2 Compressed Synthesis

Hand-authored·7 min read·10 sections·Last edited May 12 by initial import·View history

What Axiom Zen has built (the artifact under evaluation)

A custom support-agent UI on top of FreshDesk for their own product (Toby). Features:

  • Custom UI replacing FreshDesk's native interface
  • LLM agent that drafts every response (human reviews/approves with one click; signed by human name)
  • User-context side panel that auto-loads from product DB (Retool-replacement pattern)
  • "Copilot" chat — agent can ask questions about the user, the agent runs DB queries
  • History view showing the full thread
  • Quick actions / quick links into Stripe and FreshDesk admin
  • Predicted-next-action buttons (LLM tries to anticipate the next agent step and prepare it)
  • "Skills" — auto-extracted reusable patterns from past tickets
  • "Corrections" — when the human edits the draft, the agent learns
  • "Snippets" — stored facts (like Fin's snippets but auto-suggested)
  • KPIs/agency metrics

The lighthouse-customer-shape (Dapper Labs)

  • ~800 unsolved tickets, 50–100/day baseline, spikes to 100+ around drops/2FA-settlement events
  • Stack: Zendesk (~$54k/yr) + Fin AI ($30K-ish/yr at 34% AI-only resolution / 42% blended) + Retool (deprecating) + n8n (workflow automation, Mark Kingston building) + Persona (KYC) + BigQuery + Flow Scan
  • Top ticket drivers: 2FA resets (post-SEC settlement spike), Disney products (Fin BANNED by Disney contract), missing-pack/missing-moment claims (often resolved by reimbursement now because eng has no time)
  • Pain is volume × headcount loss: Britney left, Dan moving to product, Kenny VIP-only, Mark is the last general agent
  • Justin (Dapper) explicitly: "We could work together some way"
  • Dan Carrero at Dapper is already trying to vibe-code his own Retool replacement
  • Dapper has been laying off (3 rounds since Nov 2022 — 22%, 20%, 51 more)

Phase 0.5 first-principles thesis (highlights)

  • HITL pattern (human in the loop, agent drafts, human approves) is the right UX — solves trust + PII problem in one stroke
  • Skills/corrections are a real flywheel — defection cost compounds
  • Existing AI support is doc-only — they retrieve from FAQs but don't execute. Full DB access = 10x capability gap
  • Crypto verticalization may be the right wedge: Sierra/Decagon/Fin have ZERO crypto logos
  • Risk: services-firm-to-product is a hard cultural transition; Pivotal Tracker shut down April 2025

Phase 2 Research Findings — Compressed

Market & Competitors

  • TAM: $15–50B by 2026–2030, 25%+ CAGR
  • Tier 1: Sierra ($10B / $150M ARR / +354% YoY), Decagon ($4.5B / $35M ARR / +283% YoY), Fin ($100M+ ARR at $0.99/resolution), Zendesk AI ($500M ARR projected 2026 / 20K customers), Salesforce Agentforce ($500M+ ARR)
  • Mid-tier squeezed: Forethought acquired by Zendesk March 2026 ~$200M+ (Zendesk's largest deal in 2 decades)
  • NICE acquired Cognigy $955M Sept 2025 — largest pure-play CX AI exit
  • Niche: Pylon ($31M B, B2B-only, 150+ migrations from Zendesk), Lorikeet ($35M A, fintech/healthtech vertical), Wonderful ($100M A, multilingual), Maven AGI, Crescendo, Chatwoot (open-source)
  • Hybrid pricing winning: 27%→41% in 12mo (per-resolution + platform fee). Pure per-seat declining.
  • Wedges OPEN: vertical-native (crypto/web3 has NO leader), B2B-shaped, exception-handling, mid-market gap between Fin ($0.99 no-floor) and Decagon ($50K floor + custom), data-sovereign/self-hosted, voice-first non-call-center

Crypto Vertical

  • 125–160 plausible buyers globally, ~40–50 qualified ICP
  • Universe: 25–40 CEXs (Binance, Coinbase, Kraken, Bybit, OKX...), 15–20 wallets (MetaMask, Phantom, Trust Wallet 220M users...), 10–12 NFT marketplaces, 8–10 collectibles (Dapper, Sorare, Fanatics, VeVe), 6–8 prediction markets (Polymarket, Kalshi, Hyperliquid), 5–8 fan tokens, 6–10 stablecoin/RWA (Circle, Tether, Paxos, Ripple, Ondo), 5–7 institutional custody (Fireblocks, BitGo, Anchorage)
  • ZERO horizontal vendors have a crypto logo publicly (Sierra, Decagon, Fin all skew Notion/Duolingo/Rippling/Bilt)
  • Coinbase 2025 breach ($400M cleanup): bribed offshore agents copied data on 69,461 customers including KYC IDs, using legitimate support tools. Made support tooling = board-level risk vector, not back office. Budget unlocked.
  • Crypto-specific workflows generic AI handles poorly: 2FA recovery, wallet ops, blockchain tx investigations (Etherscan/Solscan/Flow Scan), missing NFTs, KYC/AML failures, withdrawal disputes, bridge failures, refunds (no chargeback equivalent), stolen-account claims (~30-50% are the hacker themselves)
  • MiCA full enforcement July 1, 2026 — every CASP needs Travel Rule + OFAC + KYC-aware support live by mid-2026
  • GENIUS Act (July 2025) — stablecoin issuers now under BSA, mandating CDD/SARs/OFAC in support workflows
  • Dapper-as-customer caveat: low margin, churn-risk. Real money in CEX + stablecoins + prediction markets.
  • Top-3 (Binance/Coinbase/Kraken) will never fully buy — license components. Target the next 30.

Unit Economics

  • HITL cost: $0.15 inference + $1.80 reduced human = ~$2/ticket (vs $4–5 human, $0.15 full AI)
  • Vendor blended GM: 60–70% (NOT 80% SaaS norm). FDE is COGS, not S&M.
  • ROI for Dapper-shape (50/day): savings ~$369K/yr at $100K ACV; payback <4 months
  • Volume floor: ~30 tickets/day to justify program. Sweet spot: 100–500/day. Above 2,000/day = bespoke compresses margins.
  • Pure per-resolution backfires above ~5K/mo — customer rage at scale. Hybrid (floor + meter + ceiling) is winner.
  • Inference NOT going to zero: GPT-5.5 doubled prices April 2026. Win = caching/cascading/bounded loops, not waiting.
  • $1M ARR achievable on $2–2.5M burn (4 engineers, 18 months).
  • $10M→$100M is roughly the same cost as classic SaaS, sometimes worse.
  • Cresta cautionary: $52M rev / 508 employees = broken cost structure.
  • Subtl.ai shut down July 2025; Forethought acquihire'd; ~40% of AI startups shut down 2024-2026.
  • Qualtrics: AI customer service fails at 4× rate of any other AI use case for end-customer satisfaction.

Regulatory & PII

  • EU AI Act Article 50 (transparency, AI must disclose itself) — effective 2 August 2026, narrow "obvious from circumstances" exception that customer-service bots don't qualify for
  • Colorado AI Act effective 30 June 2026 (postponed from Feb 2026)
  • MiCA full enforcement 1 July 2026
  • Texas TRAIGA effective 1 January 2026
  • California ADMT effective 1 January 2026 (opt-out + risk assessments for AI replacing human decisions about consumers)
  • Air Canada / Moffatt v. Air Canada (Feb 2024): companies liable for chatbot misrepresentation as if it were a webpage
  • Setzer v. Character.AI (settled Jan 2026): first ruling classifying chatbot output as a "product" — opens strict products liability
  • 88% of AI vendors cap liability at contract value or 12 months fees — far below deployer exposure
  • ISO CG 40 47 ("AI Exclusion") lets carriers exclude generative-AI claims from CGL early 2026; AIG/W.R. Berkley/Great American filing AI exclusions
  • Sierra's compliance stack (SOC 2 Type II + ISO 27001 + ISO 42001 + HIPAA + GDPR + CCPA + CSA STAR L1) = the moat. Decagon weakness: HIPAA enterprise-tier-only, gap is most-cited critique.
  • SOC 2 Type II–certified support agents close enterprise deals 2.3x faster, 67% higher ACV
  • Disney "license-or-sue" = Disney Pinnacle no chatbot at all without explicit licensure. Becoming template for sports leagues, studios, music labels.
  • OFAC settlement $3.1M (2025): customer-support agents recommended VPNs to disguise user locations — proving support is OFAC compliance surface
  • 89% of consumers want a human button; CA SB 243 (companion-bot private right of action); CFPB rulemaking one-button human escalation
  • Architecture defensible: VPC-deployed Bedrock or Vertex + RAG with row-level RBAC + tokenization middleware + agent-assist (not autonomous) + redacted prompts

Historical Precedent

  • Pivotal Tracker (closest analog: tool built inside consultancy → product) shut down April 2025
  • Basecamp (cleanest success): took >12 months post-launch before consulting could be killed; founders fully reframed identity
  • Slack (Glitch → chat): required killing the original game company entirely
  • Pioneer Square Labs / Atomic (cleanest venture-studio comps): separate cap tables, separate brands, separate leadership is the playbook
  • Thoughtworks Studios shut down 2020 after 14 years — services culture couldn't ship 5-yr SaaS roadmap
  • AZ → Dapper spinout (2018) is the prior successful AZ playbook
  • AI hype cycle: Klarna's May 2025 reversal validated co-pilot camp. Hybrid wins renewal; full automation wins demo. Forrester: 41% of AI tools deployed in prior 18mo were replaced/restructured within 12mo (avg switch cost $340K).
  • Buyer is now CFO + VP CS + CIO + Champion (CX Ops Lead). Sales cycle: Zendesk 4-7mo, Decagon 6wk-3mo, Sierra 3-6mo total
  • Switching cost from Zendesk: 3–6 months full cutover, 9–12 months stabilization, $340K avg cost (Forrester)
  • Power-user demonstrated WTP: $150–250 PEPM total stack spend in 50+ agent orgs

The decision space

Three coarse-grained options for AZ:

A) Productize broadly as venture-scale company — compete with Decagon/Sierra in mid-market AI support agent space. Late, crowded, well-funded.

B) Productize narrowly as crypto/web3 vertical specialist — Dapper as lighthouse, target the ~40 qualified ICP buyers. Exit comp: NICE→Cognigy $955M, Zendesk→Forethought $200M+.

C) Keep it internal — use Toby support tool only, perhaps share blueprints with Dapper as a partner gesture. No product transition.

D) Hybrid — services-product — sell to 2-3 anchor customers (Dapper + maybe 1 crypto exchange + maybe 1 collectibles peer) as a high-touch deployment. Software paid as services. Defer the venture-scale decision until $5M+ ARR signal.

Each perspective agent should evaluate these options through their lens.