Research
The Agentic CX Stack
A field map of the layers every AI-first customer experience runs on — and where the current tools actually sit.
Working document — revised as the landscape moves. Last major revision: June 2026.
Every “AI-first CX” architecture I’ve reviewed this year decomposes into the same five layers, whether the vendor admits it or not. Naming the layers makes vendor conversations dramatically shorter.
The five layers
- Channel adapters — voice, chat, email, WhatsApp. Commodity by now; the mistake is letting a channel vendor own the layers above it.
- Conversation orchestration — turn-taking, interruption, handoff to humans. This is where most “agent washing” happens.
- Agent runtime — the actual reasoning loop: tools, memory, guardrails. The layer with the most churn and the least lock-in worth accepting.
- Knowledge substrate — the retrieval layer over your actual business truth. Underinvested everywhere. If this layer is bad, nothing above it can be good.
- Evaluation & telemetry — did the agent actually resolve the thing? Still mostly spreadsheets in 2026, embarrassingly.
Where the market actually is
The contact-center incumbents are strongest at layers 1–2 and weakest at 4. The AI-native startups invert that. Nobody sells a credible layer 5 yet — every serious deployment I’ve seen built its own.
What I’d bet on
- Layer 4 becomes the durable moat; layers 1–3 converge to commodity within two years.
- Teams that treat evaluation as a product feature — not QA — ship agents customers actually trust.
- The “one vendor, five layers” pitch keeps failing procurement for good reasons.
Open questions
- Does the knowledge substrate belong to CX or to the enterprise KM function? (My KMWorld argument: KM, and CX rents it.)
- What does human handoff look like when the agent is better than tier-1 support?