Three agentic platforms launched into adjacent customer-facing categories this week. Zigment put an agentic layer across the marketing stack and promised sentiment-aware action on every message, click, and call. Waldo shipped what it calls agentic brand intelligence, with a Strategy Agent and one-click workflows. IrisAgent brought a support resolution platform that runs autonomously across chat, email, and voice, grounded in a customer's knowledge base and SOPs.

The lazy read is that agentic CX is becoming a category. That is true, and it is also boring. Three launches do not make a market, and any one of them could be repositioned in a quarter. I keep coming back to something smaller and more durable in the product copy itself. Each of these platforms is, in different language, selling the same thing: a structured way to decide when the agent should not act.

My read is that the moat forming under this category is not the agent. It is the stop contract.

The product feature nobody is naming yet

Zigment's pitch is that sentiment becomes a control input, not a dashboard metric. A frustrated tone on a call shifts the next action. A high-intent click escalates a journey state. Strip the marketing language away and what you are buying is a set of triggers that decide which interactions stay inside the agent loop and which leave it. The interesting variable is not how fast the agent can act. It is where the platform draws the line between confident action and required handoff.

Waldo points at the same surface from the brand side. "One-click workflows" sound like speed, but the moment brand intelligence becomes executable, somebody has to define which insights are allowed to turn into irreversible market-facing actions and which require an operator to spar with the Strategy Agent first. The product is partly the agent. The product is more honestly the geometry of that one click: what it can launch unattended, what it cannot, and who gets the choice.

IrisAgent makes the pattern unmistakable. Autonomous resolution across chat, email, and voice only works if there is a defensible architecture for when the system stops resolving and routes to a human. Voice is the giveaway. Voice adds ambiguity and recovery cost that text channels can hide, and the only way to operate it autonomously at enterprise scale is to ship a tight, legible escalation layer alongside the autonomy.

The agent is the marketing. The judgment gate is the product.

What buyers are about to start asking

The shift this implies for buyers is sharper than another round of automation-rate benchmarking. Automation rate is easy to game and easy to ship. A 50 percent ticket deflection number tells you almost nothing about whether the 50 percent the agent kept were the right ones to keep.

The questions that matter look more like product requirements than governance language:

  • What are the explicit triggers that stop the agent and route to a human, and are they exposed as an API the buyer can edit?
  • What confidence thresholds gate which classes of action, and how are those thresholds learned and revised over time?
  • Where does next-best-action selection live, and can the operator override or constrain it without breaking the rest of the workflow?
  • What does the audit trail look like at the decision boundary, not just at the output?

These are not safety questions. They are commercial questions about who owns the most consequential surface in the system. A buyer who can edit those triggers gets a platform. A buyer who cannot gets a vendor-defined operating model wrapped in their own brand.

This is also where the agentic category quietly differs from a chatbot category. Chatbots competed on fluency. Agentic platforms compete on the legibility and editability of their decision boundaries. That is closer to how serious teams have always evaluated continuity at the handoff than to how marketing has historically evaluated content tools.

The power question hiding inside the product question

Zoom out and the commercial implication starts to look structural. Whoever owns the escalation layer across customer channels owns the working definition of acceptable action under uncertainty inside that business. Sentiment thresholds, refund authority, claim handling, brand-safe response patterns, when to push a price, when to hold one. All of that lives inside the gate.

This is not a Machiavellian observation. It is an operating one. Customer-facing decisions used to be encoded in policy documents that nobody read and CRM rules that nobody updated. They are about to be encoded in agent control planes that run continuously and observably. Once those control planes are in place, changing them is a product change, not a memo. The platform that owns the gate has a quiet but compounding say in how the company shows up to its market.

That is why I would not read these launches as three competitors in a CX category. I would read them as three early entrants into a control-plane category that does not have a name yet. The previous wave of marketing platforms sold the workflow. The current wave is starting to sell the governed loop underneath it, and the place where that loop becomes legible is the stop.

Where this leaves builders

If you are building in this space, the strategic move is unsentimental. Stop competing on automation rate. Compete on the expressiveness and ownership of the stop layer. Ship the triggers as a first-class API. Make confidence thresholds editable and inspectable. Treat next-best-action selection as a contract between platform and operator, not a black box you defend with case studies.

If you are buying, the same logic runs in reverse. The vendors who deserve serious evaluation are the ones who can describe their stop contract crisply, in product terms, and demonstrate it on your data, not on theirs. The ones who reach for governance vocabulary when you ask are telling you they have not built the layer yet. There is a real difference between a deployable agent and a system you can actually run, and it sits almost entirely at this seam.

The broader thing this changes is the texture of agent procurement. For a year, the dominant question has been: can this agent do the work. For the next several quarters, the more useful question is: when this agent should not do the work, how is that decided, who can change it, and what does the system do next. That is a product question with a clean answer or a hand-wave. There is not much middle ground.

The next agentic CX or brand intelligence pitch you sit through will probably open with throughput, deflection, and revenue lift. Let it. Then ask one question and watch the room.

What exactly is the stop contract, and who owns it.

Sources and further reading