# Beyond the FAQ bot: AI agents that act need somewhere to take the stage

> AI agents in messaging have moved from answering basic questions to performing complex tasks. The interface hasn't caught up yet.

By [Adam Grohs](https://www.linkedin.com/in/adamgrohs/), Co-Founder / Head of Product - Stackable Labs · August 11, 2026

For years, "AI in customer service" meant one thing: a bot that could find your return policy. (Most of the time. On a good day.)

If it linked the right help-center article, someone somewhere popped champagne.

Extra credit if it was faster than just searching yourself.

That era is ending, and honestly, good riddance. The question customers bring to a messenger was never really *"what's your return policy?"* It was *"return this."* It was *"where's my order?"*, *"move my booking to Thursday,"* *"skip next month's box."* Those aren't questions. They're jobs.

AI agents can finally do those jobs.

The catch? Most of them are still doing it from inside a text box, like a headliner stuck performing in the lobby.

## The industry has moved from answering to acting

This isn't just us saying so. Over the past year the big platforms have pushed hard toward agents that take action.

- On **March 30, 2026**, Zendesk announced it would "expand access to our most advanced AI agent capabilities for all Zendesk customers," including "agentic reasoning, multi-step procedures, and external API integrations," in its Suite and Support plans. ([Zendesk announcement](https://support.zendesk.com/hc/en-us/articles/10487730059034-Announcing-expanded-access-to-AI-agent-capabilities-for-all-Zendesk-customers))
- At Relate on **May 19, 2026**, Zendesk introduced what it calls the Autonomous Service Workforce. The release says its AI agents "now operate across messaging, email, voice, and AI platforms like ChatGPT and Gemini," and it put a Zendesk MCP Client and Action Flows for AI Agents into early access, with an MCP Server to follow. ([Zendesk newsroom](https://www.zendesk.com/newsroom/articles/relate-2026/))
- On **January 26, 2026**, the Model Context Protocol maintainers shipped **MCP Apps**, the first official MCP extension. In their words: "Tools can now return interactive UI components that render directly in the conversation." ChatGPT, Claude, Goose, and Visual Studio Code had already shipped support. ([MCP blog](https://blog.modelcontextprotocol.io/posts/2026-01-26-mcp-apps/))

Put those together and the direction is clear. Agents are getting hands: tools, integrations, and protocols like MCP that let them reach into the systems where orders, bookings, and accounts live. And the people building AI clients have already decided that once an agent can *do* things, plain text isn't enough to show the result.

Customer messaging deserves the same.

## Your agent is performing in the dark

Here's what "where's my order?" looks like with a capable agent and nothing but text:

> Your order #48213 shipped on Tuesday via UPS, tracking number 1Z999AA10123456784, and is expected to arrive Friday. It contains 2 items: Merino Crew (Navy, M) and Trail Sock 3-Pack. If you'd like to start a return, reply RETURN. To change your delivery address, reply ADDRESS. To speak with an agent, reply AGENT.

Every fact in there might be correct, and it's still a wall of text. The customer has to parse it, find the part they care about, and then type a magic word to do anything. Reply RETURN. Very 1996.

Now picture the same moment as an **order card** next to the conversation: a status badge that says *Shipped*, a progress bar toward delivery, product thumbnails, a tracking link, and two buttons, **Start a return** and **Talk to someone**. Same data, same agent behind it. One of these feels like your brand. The other feels like a terminal session.

That's the gap. The intelligence got a massive upgrade. The stage it performs on mostly didn't.

## What an acting agent needs from its stage

When an agent can change things in the real world, the interface stops being a nice-to-have and becomes part of how the action works. In our experience it has four jobs.

**1. Show the state.** Before anyone acts, the customer should see what they're acting on: the order, the booking, the subscription. Fetched live, shown clearly, without making the customer ask for it first.

**2. Let the customer confirm, with a real control.** "Reply YES to confirm your $84.00 refund" is how you get accidental refunds and very confused customers. A good confirmation shows the exact items, the exact amount, and the exact date, then waits for a tap on a clearly labeled button. The customer decides and the system carries it out. For anything that's hard to undo, a model's guess about what someone meant should never be the only confirmation.

**3. Leave a receipt in the conversation.** When something happens, the transcript should say so: "Return started for 1 item, label attached." That keeps the record honest for the customer, the AI agent's next turn, and any human who picks the thread up later.

**4. Make the handoff carry context.** Some conversations need a person. When that happens, the agent who picks it up should already see which order, which issue, and what's been tried. Nobody should have to start the show over from the top.

## Enter Stackable, stage left

This is exactly the layer we built Stackable to provide. The Zendesk AI agent stays the brain of the conversation. Stackable extensions give it a stage and a set of hands right next to the messenger, and every one of the four jobs above maps to something real in the SDK:

- **Show the state:** extensions render into surfaces around the messenger and fetch live data through `data.fetch`, a proxied, allowlisted request. API keys are injected server-side, so they never touch the browser.
- **Confirm:** extensions are built from real UI components (cards, badges, progress bars, radio groups, buttons), and they inherit the brand's theme automatically.
- **Leave a receipt:** `messaging.send` lets an extension post text, images, files, or carousels into the active conversation, so a return label PDF can land right in the thread.
- **Hand off with context:** `actions.invoke` can start a new conversation with tags, custom fields, and metadata already set, so routing and your agents get the context up front.

It also works in the other direction. When an integration posts a postback button into the conversation through the Sunshine Conversations API, like *"Check my order,"* an extension listening with `events:messaging` can react and open the order view. The conversation decides what comes next, the extension puts it on stage, and the customer confirms.

There's also a quieter kind of trust here: trust for the brand. Every Stackable extension runs in a sandbox, declares the permissions it needs up front, and only gets those permissions. Marketplace submissions go through an automated scan and a review before they're published. If an agent is going to take action on your customers' behalf, the code it's working through should be held to that standard.

Our launch partners saw this early. [Kaan Tavli](https://www.linkedin.com/in/kaan-tavli/), Zendesk Practice Director at [Optimate.me](https://www.optimate.me), described Stackable as "evolving Zendesk's AI agents and messaging into a fully interactive experience layer." That's the idea in one line.

## A few house rules worth stealing

Whether or not you build on Stackable, these are the principles we'd put on the wall for any team shipping agents that act:

- **Never let free text be the only confirmation for something irreversible.** Show it, then make it a button.
- **Show the numbers.** Amounts, dates, quantities, and addresses belong in the UI, not buried in paragraph three.
- **One tap to a human.** An agent that can do everything except get you a person isn't an agent. It's an escape room.
- **Keep the transcript true.** If it happened, it should be in the conversation. If it didn't, the conversation shouldn't say it did.
- **Least privilege, always.** Give each piece of the experience only the access it needs, and make it declare that access up front.

## The experience is the point

We've spent two decades treating customer experience as *the* differentiator, and then a lot of us accepted that flatter, faster, and more automated was "good enough." The AI agents arriving now are better than "good enough" (or at least that's what the matrix is telling us to believe).

They deserve better than a gray bubble, and so do your customers.

Your agent's ready for its close-up. If you want to see what an agent-ready messenger looks like, [see it in action](https://stackablelabs.com/demo) or [browse the marketplace](https://stackablelabs.com/marketplace) for extensions you can install today. Stackable Messaging is free to get started.