AI Agents for Customer Support: Market Map
1. Thesis
Customer support is shifting from "software that helps humans answer tickets" to "AI agents that resolve tickets end-to-end," collapsing the per-seat SaaS model into per-resolution pricing and forcing incumbents to cannibalize their own seat revenue before startups do it for them.
---
2. Segments
A. Incumbent Platforms Embedding AI Agents
Companies: Zendesk (AI agents / former Ultimate.ai), Intercom (Fin), Salesforce (Agentforce), Freshworks (Freddy AI), HubSpot (Breeze)
Dynamics: Distribution advantage plus data gravity — they own the ticket history and workflows. But they face a painful pricing transition: Intercom's Fin at ~$0.99/resolution is the template everyone is copying, and it explicitly trades seat revenue for outcome revenue.
B. AI-Native Agent Startups (Full Resolution Focus)
Companies: Decagon, Sierra (Bret Taylor's company), Ada, Forethought, Lorikeet, Gradient Labs (less certain on Gradient Labs' current traction)
Dynamics: The hottest venture category — Sierra and Decagon have raised at multi-billion valuations. They sell on resolution rate, not features, and win logos frustrated with incumbent AI quality. Key risk: thin differentiation as foundation models improve and incumbents catch up.
C. Voice-First AI Agents
Companies: PolyAI, Parloa, Replicant, Cresta (agent-assist heritage moving toward autonomy), Retell AI (Retell is more infrastructure/API — borderline)
Dynamics: Phone is ~50%+ of contact center volume and the most expensive channel, so ROI is largest here. Latency and interruption-handling were the moat; that moat is eroding as realtime voice models commoditize, pushing value toward telephony integration and enterprise deployment.
D. Agent-Assist / Copilot Layer (Human-in-the-Loop)
Companies: Cresta, Assembled (Assist), Observe.AI, Level AI, ASAPP
Dynamics: Sells to VPs of CX who aren't ready for full autonomy — coaching, real-time suggestions, QA automation. Structurally squeezed: it's a waypoint, not a destination, and most players are racing to reposition as full agents.
E. Contact Center Infrastructure (CCaaS) Adding AI
Companies: NICE, Genesys, Five9, Talkdesk, Amazon Connect
Dynamics: They own routing, telephony, and workforce management — the plumbing agents must run through. Slower to ship credible AI, but hard to rip out; likely acquirers of segment B/C companies.
---
3. Who Captures the Most Value
*Segment A (incumbents) captures the most absolute value; Segment B captures the most new value.*
The incumbents win the bulk because: (a) AI agents need the system of record — ticket history, knowledge bases, order data — and switching costs on that layer are enormous; (b) per-resolution pricing on their installed base is a revenue expansion if they execute (Intercom's Fin is the proof case); (c) enterprise buyers prefer one throat to choke. AI-native startups (Sierra, Decagon) will produce 2-3 large outcomes on complex enterprise deployments, but most of segment B gets acquired or squeezed as resolution quality converges.
---
4. White-Space Opportunities
1. Agent-to-agent commerce/support resolution. As consumers deploy their own AI agents (booking, purchasing, disputing charges), someone needs to own the machine-to-machine support protocol — authentication, negotiation, and resolution between a customer's agent and a company's agent, with no human on either side. No current player is building the trust/settlement layer for this.
2. Regulated-industry escalation and liability infrastructure. In healthcare, insurance, and financial services, the blocker isn't AI quality — it's audit trails, error attribution, compliance-grade logging, and insurance against wrong answers (who pays when the agent misquotes a policy?). A "compliance and liability wrapper" for AI support agents — potentially bundled with actual E&O-style insurance — is unbuilt. Existing players hand-wave this with "human escalation," which doesn't satisfy regulators.
---
Confidence notes: Company lists are real as of my knowledge cutoff; specific valuations and Gradient Labs/Lorikeet traction are the items I'm least certain about. The Ultimate.ai→Zendesk and Fin pricing facts are well-established.