The support AI stack

Support AI comes in distinct layers, each with a different risk profile and role. Knowing the stack helps you deploy the right AI in the right place (a 2026 snapshot — verify specifics, they change fast).

The layers:

1. Helpdesk / platform AI — AI built into your ticketing system (Zendesk AI, Intercom's Fin, Freshdesk's Freddy). The AI capabilities native to where your support already runs.
2. Customer-facing AI agents / chatbots — AI that talks to customers and resolves tickets (Fin, Ada, Sierra, Decagon, Salesforce Agentforce). Highest risk — it's customer-facing and can give wrong answers, so it needs the most guardrails (Module 2 and 4).
3. Knowledge base tools — AI that generates and maintains your help content, which grounds everything else (Module 3).
4. Agent assist / copilots — AI that helps your human agents (Zendesk Copilot, Intercom Fin Copilot, Cresta, Forethought). Lower risk — a human approves every customer-facing word — which makes it a great starting point for cautious teams.
5. QA / analytics / voice-of-customer — AI that scores conversations, analyzes sentiment, and surfaces trends (Level AI, Zendesk QA, IrisAgent) — now including QA'ing the AI agents themselves (Module 3).

A key distinction to internalize: customer-facing vs. agent-facing. Chatbots face the customer (high risk — wrong answers reach customers directly). Copilots face the agent (lower risk — a human reviews before anything reaches the customer). Many teams wisely start with agent assist (copilots) — getting AI's help while keeping a human in every customer interaction — before deploying fully customer-facing agents. Match the AI's autonomy to your risk tolerance and readiness.

The 2026 commercial shift: outcome-based pricing. The headline change in how support AI is sold: instead of per-seat licensing, vendors increasingly charge per resolution — you pay when the AI actually resolves a ticket (Intercom's Fin popularized ~$0.99/resolution; Zendesk and others followed with per-automated-resolution pricing). Why it matters:

• It aligns vendor incentives with actual value (they get paid when the AI works).
• It makes cost scale with usage — model your economics on your volume and resolution rate, not vendor headline numbers.
• **Teach the model, not the exact price** — the numbers change quarterly, but "you pay when AI resolves something" is the durable shift.

How to think about assembling your stack:

• Start where risk is lowest and value is clear — often agent assist (copilots) and knowledge tools — before fully autonomous customer-facing agents.
• Ground everything in an accurate knowledge base (the foundation — Module 2/3).
• Mind your existing platform's AI (Zendesk, Intercom, Freshdesk bake in a lot) before buying separate tools.
• QA the AI, not just the humans — as you deploy customer-facing AI, you need to monitor its quality (Module 3).

The takeaway: the support AI stack ranges from lower-risk agent-assist copilots to higher-risk customer-facing agents, all resting on an accurate knowledge base. Deploy deliberately — match AI autonomy to your readiness, start where risk is low, ground everything in accurate knowledge, and understand the outcome-based pricing that increasingly governs the economics. That deliberate approach is what separates support AI that helps from support AI that fuels the backlash.