Amazon ConnectCX Hive

Amazon Connect Agent Assist Answers Back With Cards, Buttons and Actions

Agent assist can now reply with interactive cards, buttons and forms that run actions through MCP tools — here is how to design the guardrails around it.

Until now, agent assist in Amazon Connect Customer talked back in plain text. It suggested what to say, surfaced a knowledge article, drafted a summary — and then the agent did the doing. The November 16 announcement changes who does the doing: agent assist can now answer with interactive components — cards, carousels, buttons and input fields — and when the agent confirms, the AI agent runs the action itself through MCP tools in Connect Customer or third-party services.

What changed

Agent assist can now respond with interactive components instead of plain text only. The AI agent analyzes conversation context in real time, recommends actions, retrieves information and executes tasks on behalf of agents using MCP tools. It can respond proactively or when the agent asks, working from views you define first; if allowed, a custom AI agent can compose its own card, summary or short form.

Agents select options, fill in details and confirm actions that the AI agent then runs through MCP tools in Connect or third-party services. Responses can also open step-by-step guides, cases, customer profiles and third-party applications with one click. Each interactive response is checked before display, and responses with unsafe links or actions the AI agent is not allowed to use are rejected.

The example AWS gives is a charge dispute: the AI agent generates an interactive list of the customer's recent transactions, the agent selects the disputed one, confirms the details and approves, and the AI agent files the dispute and shows confirmation in the agent workspace.

The feature is available in all AWS Regions where Amazon Connect Customer agent assist is offered. Chat needs message streaming for AI-powered chat, and proactive voice responses need real-time conversational analytics.

For the full release record, see the Amazon Connect release intelligence page, the agent assist interactive responses update, and the official AWS announcement (opens in a new tab).

Pattern 1: Treat the card as a confirmation gate, not a shortcut

The most important design decision in this release is not in the feature — it is in the card. A button that files a dispute, issues a refund or updates an account is a real action behind a single click, so the card has to show what is about to happen before it happens. The dispute example gets this right: the card lists the transactions, the agent picks one, confirms the details and approves. The agent is still the decision-maker; the AI agent is the hands.

Design each card as a confirmation gate: it should name the action, the record it touches and the consequence — "File a dispute for transaction $42.18 on Nov 12" rather than "File dispute." If the card cannot show that much, it is not ready to carry an action. Write the rule down before the first pilot flow: no action without a visible target.

That rule is also what makes the pattern safe to extend. The same shape works wherever the tool behind the action lives — a billing system reached through an MCP tool, a CRM record in Salesforce or Zendesk, an order-management lookup, an ITSM ticket in ServiceNow. So the review question for each new card is always the same: is this an action you would let an agent run with a single click, with only the information shown on the card? If the answer needs a caveat, the card needs more detail — or the action needs a human step first.

Pattern 2: Views are the contract

The release draws a line that is easy to miss: agent assist works from views you define first, and only a custom AI agent that is explicitly allowed to do so can compose its own card, summary or short form. That line is the governance model.

A view is a contract: for this kind of situation, here is what the AI agent may show, which fields it may include and which actions it may offer. Treat views like a scripted response or a macro — reviewed, versioned and owned by someone. Who can create a view, and who approves a change? Cheaper to answer now, with one pilot flow, than later, when a dozen views have grown organically.

The "compose its own card" allowance deserves the narrowest start. Letting the AI agent compose a card is letting it decide the layout of the confirmation gate — which fields to show, which options to offer, what to leave out. Start with predefined views only, and add composed cards deliberately: one use case, one review of what the AI agent actually composed. The safety check that rejects unsafe links and disallowed actions is a backstop, not a substitute for deciding in advance what the AI agent is allowed to build.

Pattern 3: One-click openers turn assist into a launcher

The quieter half of the announcement is the one-click opener: an interactive response can open a step-by-step guide, a case, a customer profile or a third-party application. Read as a design pattern, that turns agent assist from an answer engine into a launcher — the place where the agent's next step begins.

The pattern to design around is the handoff. A card that surfaces a billing dispute should open the right case, not just any case view; a card that recommends a step-by-step guide should open the guide at the step that matches the conversation, not at the top. The value is not the click itself — it is that the click lands in the right place, with the context already attached. That is what makes the integration with the rest of the stack worth planning: the CRM profile that opens with the customer already loaded, the ITSM ticket that opens pre-filled from the interaction, the survey tool that opens the right post-contact questionnaire. Each opener is a small integration decision, best made with the team that owns the target system in the room while the view is being designed.

The failure mode to watch for is opener sprawl: every card offering five destinations, each one plausible, none of them the obvious next step. Keep openers per card to one primary destination, maybe a second. A card that tries to be a dashboard is a card an agent will stop reading.

A short checklist for the week

  1. Check the prerequisites. Agent assist needs AI agents with a domain and a knowledge base, a Connect assistant block in the routing flows, message streaming for chat and real-time conversational analytics for proactive voice responses.
  2. Pilot one flow with a short allowed-tools list. AWS names the AgentAssistanceOrchestrator agent for the pilot; run it on one flow, with a short list of allowed tools, and one use case — the dispute flow is the natural first candidate. A short tools list is a safety decision, not a limitation.
  3. Review every tool the AI agent can call. The tools exposed to the AI agent need the same review as any agent-facing integration: what each tool can change, which records it can touch, and what happens when it runs twice. The pre-display safety check rejects actions the AI agent is not allowed to use — make sure the allowed list is deliberate rather than inherited.
  4. Write the card rule down and name the view owners. No action without a visible target; each view gets an owner and an approval step for changes. Content without an owner drifts.

Interactive responses do not change what agent assist is for; they change how far its answers reach. The teams that get the most from this release will be the ones that treat a card with a button on it as a small piece of automation — designed, reviewed and owned — rather than as a nicer way to display a suggestion.