What Actually Changed

Per-resolution pricing became standard. In 2024, Intercom introduced $0.99/resolution for Fin. In 2025, Zendesk followed. By 2026, per-resolution pricing is the dominant model for AI support. You pay when AI resolves a ticket, not for access to AI features.

This matters because it aligns incentives. Vendors only get paid when their AI works. Customers only pay for results. The old model — pay $50/agent/month for AI features that may or may not help — is fading.

The downside: $0.99-$1.00/resolution is expensive at scale. But competition is pushing prices down. Classification-based tools offer the same resolution model at $0.20-$0.30.

AI agents replaced chatbots in marketing copy. Every vendor rebranded their chatbot as an "AI agent" or "autonomous agent." The terminology shift reflects a real change: modern AI can take actions (process refunds, create tickets, update accounts), not just answer questions.

But most "AI agents" in 2026 are still chatbots with better LLMs behind them. The ones that actually take autonomous actions — Ada, Siena, the Intercom/Zendesk agents — are the minority.

Voice AI entered production. AI-powered phone agents went from demo to deployment. Pizza chains take orders with AI. Medical offices schedule appointments. Insurance companies handle intake calls. The technology works for structured, predictable phone interactions.

It doesn't work for complex, emotional, or multi-issue phone calls. But for the 40-60% of calls that follow a script, voice AI is production-ready.

Intent classification became a commodity. What used to require custom ML engineering is now available as an API call. Pre-trained classifiers handle 300+ support intents out of the box. Small teams can get the same classification accuracy that large companies spent months training custom models to achieve.

What's Still Hype

"AI will handle 80% of support by 2029." Gartner's prediction gets quoted everywhere. What people miss: the "80%" refers to common, routine issues. The complex 20% — the tickets that actually need human judgment — isn't getting automated anytime soon. And that 20% is the most important 20%.

"Our AI agent is autonomous." Most AI agents in 2026 still need a knowledge base to reference, guardrails to prevent bad answers, and human oversight for edge cases. "Autonomous" in marketing means "resolves some tickets without a human." That's useful. It's not autonomous.

"AI will replace support teams." It hasn't happened. Support headcount at most companies has decreased slightly (5-15%) or stayed flat. What changed is the work: agents handle fewer routine tickets and more complex issues. The role got harder, not smaller.

"Generative AI is always better than classification." LLM-generated responses are more flexible but also more expensive, slower, and occasionally wrong (hallucination). For predictable questions with known answers, classification at $0.20 beats a $0.99 LLM-generated response every time. Different tools for different problems.

What to Pay Attention To

Multi-step automation. The real evolution isn't better chatbots — it's automation that chains multiple steps: classify the message → check the customer's order → verify refund eligibility → process the refund → send confirmation. This is what "agentic AI" should mean, and it's getting practical.

AI + human handoff quality. The moment AI hands off to a human is the most important moment in the support interaction. Good handoff includes context (what the AI understood, what it tried, why it escalated). Bad handoff means the customer repeats everything. This is where the best tools differentiate.

Pricing pressure. $0.99/resolution won't last. As more tools enter at $0.20-$0.30, the premium players will need to justify the 5x price gap. Expect prices to drop industry-wide over the next 12-18 months.

Vertical specialization. General-purpose AI works for general support. But healthcare, fintech, legal, and real estate have specific classification needs that general models handle poorly. Custom vertical models (trained for specific industries at low cost) are a growing category.

What to Do About It

If you haven't automated yet: start. The technology is mature enough, cheap enough, and fast enough to set up. You're not early anymore — you're behind.

If you've automated with a $0.99/resolution tool: do the math. At 500+ resolutions/month, switching to a $0.20 classification tool saves thousands per year for the same resolution rate on predictable questions.

If you're considering voice AI: test it on one use case (appointment scheduling, order status) before committing to a full rollout. The technology works but the edge cases are real.

If you're building custom AI: reconsider. Unless AI classification is your competitive advantage, the build-vs-buy math strongly favors buying. Ship your product, not your support infrastructure.