Phase 0.5 — First-Principles Thesis (written before any external data)

What the meeting revealed (signal extraction)#

1. Dapper has ~800 unsolved tickets, 50–100/day baseline, spikes around drops/2FA settlement events.
2. Zendesk costs them ~$54k/year. Fin AI ~$30k/year. 34% AI auto-resolution rate, 42% blended resolution rate.
3. Pain is NOT in any single step (2FA reset takes ~1 minute) — pain is volume × headcount loss (Britney left, Dan moving to product, Kenny is VIP, Mark is the last general agent).
4. Top ticket drivers: 2FA resets, Disney products (which prohibit Fin entirely), missing-pack/missing-moment claims (often resolved by reimbursement now because engineering has no time).
5. Stack fragmentation: Zendesk + Fin + Retool (being deprecated) + n8n + Persona (KYC) + BigQuery + Author + Flow Scan (blockchain explorer).
6. Real failure mode: Fin answers correctly but users ignore it and demand a human — perception problem, not capability problem.
7. PII is the frontier — they CAN'T pipe sensitive data through Claude/LLM because storage on third-party servers is unacceptable.
8. Dan Carrero at Dapper is already trying to build a similar dashboard ("vibe-coded" Retool replacement). Internal demand exists.
9. Justin (Dapper) explicitly said: "We could work together some way... I f****** love that UI."

Structural reasoning (mechanism, human nature, economics)#

Why this category SHOULD work#

• Support tickets are mostly pattern-matching. A small number of issue types account for ~80% of volume. LLMs are state-of-the-art pattern matchers with semantic recall.
• The bottleneck is bipartite: (a) data retrieval (multi-tab, multi-tool), (b) draft generation (humans rewriting variations of the same response). LLMs collapse both into one step.
• Headcount is fixed cost; volume is variable. Every company that scales support faster than headcount has a structural reason to deploy AI.
• Existing AI support (Fin, Zendesk Answer Bot, Ada) is doc-only. They retrieve from FAQs but don't execute — they can't "look at your account." That's a 10x capability gap.
• The "human-in-the-loop draft" pattern is the right UX. It eliminates cold-start trust issues — agent learns by being corrected, supervisor stays in control.
• Skills / corrections / snippets self-improve the agent over time. This is a real flywheel: each ticket either becomes a learned skill or a correction. Defection cost compounds.

Why this category SHOULD fail (or be hard)#

• Crowded. Decagon (~$1.5B), Sierra (Bret Taylor, ~$4.5B), Forethought, Ada, Cresta, Intercom Fin, Zendesk's own AI, Salesforce Einstein. The investor narrative is mature and well-funded. AZ would be a late entrant.
• Support is a cost center. Support buyers (Dir of CX, VP CS) have small budgets relative to sales/marketing tools. They're often not the procurement champion.
• Switching from Zendesk is glacial. Ticket history is the lock-in. Audit trail, compliance, training corpus — the data graveyard outweighs UX upgrade. Many companies will buy a layer, not a replacement.
• Integration is the unit of work AND the moat. Each customer has a unique tool stack. Going wide kills margins; going narrow shrinks TAM.
• PII / sovereignty problem is real. Enterprise can't ship customer PII to third-party LLM. Solving this requires VPC deployment, BYO-LLM, or self-hosted inference — none of which a small product can do at scale early.
• LLM costs at 1.3k tickets/day × multiple model passes could eat margin unless caching and routing are aggressive.
• Selling to crypto specifically is a tiny vertical. And crypto buyers are the worst-paying buyers in SaaS (volatile, tight budgets, "we'll build it ourselves").
• Dapper "build it themselves" is a leading indicator. If Dan Carrero is already cloning Retool internally, AZ's product has to be 10x better than what Dapper can build with Cursor + a weekend. The bar is high.

Most underweighted insight#

The fact that users ignore correct AI responses and demand humans is not an AI capability problem — it's an UX/trust problem. The right product probably hides the AI/human boundary entirely (every response signed by a human name, agent does the work, human approves with one click). AZ's tool already does this. Most competitors do not. This is a real wedge.

Most likely structural shape#

• Not a Zendesk replacement (10-year war, lost before it starts).
• Not a horizontal AI agent (Decagon already there, well-funded).
• Plausible wedges:
  1. Vertical specialist for crypto/web3 support — pre-built integrations with Flow Scan, EVM explorers, Persona KYC, Dapper-style wallet ops. TAM is ~50–200 companies but ACV could be $50–200k.
  2. "AI co-pilot" for support that overlays your existing ticketing system — does NOT replace Zendesk, drafts in your existing UI. Lower switching cost, faster sales cycle.
  3. Internal-tool layer for fast-growing startups that lets them spin up a custom support dashboard in days instead of months — the "Retool for support" angle.
  4. Productize-with-anchor-customer-then-extract — sell to Dapper as engagement, build the multi-tenant version on the back of that revenue.

Conviction before research#

• Conviction that the product is real and useful: HIGH.
• Conviction that AZ should productize it externally as a venture-scale company: MEDIUM-LOW. The market is saturated and AZ is not a typical vertical SaaS shop.
• Conviction that AZ should productize it as a Dapper-style services-product hybrid (one or two anchor customers, software paid as services): HIGH.
• Conviction that the highest-leverage move is "Toby support tool → Dapper support tool" → maybe more: HIGH.

Anchor questions for the research swarm#

1. What is the actual current state of AI support tooling — what's commoditized, what's still open?
2. What are the 2026 wedge opportunities given Decagon/Sierra/Fin are already at scale?
3. What does the crypto/collectibles support stack look like across the industry?
4. What are the unit economics of an AI support agent — token cost vs. revenue per ticket?
5. What does Dapper's strategic situation tell us about AZ's bargaining position as a vendor/partner?
6. What are the failure modes of the "AI agent that drafts responses" pattern at scale?
7. What does the regulatory/PII landscape look like for AI support in 2026 (EU AI Act, state laws, Disney-style platform restrictions)?