Zendesk Custom Agents:
Specialized AI agents are only as good as the actions behind them

On September 14, Zendesk introduced Specialized AI Agents in two parts: Industry Agents, ready-made for specific industries and starting with commerce, and Zendesk Custom Agents, which businesses build around their own processes in Agent Builder. As Shashi Upadhyay, Zendesk's President of Product, Engineering and AI, put it: "The era of generic, one-size-fits-all AI agents is over."

We agree, with one condition. A custom agent is only as good as what it can see and what it's allowed to do. For a commerce brand, that means live order, return, subscription and loyalty data, and real actions in the systems that hold it. Here's what Zendesk announced, what's generally available and what's still in early access, and what custom AI agents will need from your commerce stack.

Industry Agents and Custom Specialized Agents

The Specialized AI Agents announcement has two halves:

• Industry Agents are prebuilt for common, high-value work in one industry. Commerce comes first: helping customers shop, manage orders, handle returns and exchanges, resolve delivery issues and process refunds. Zendesk says they connect to Shopify, Narvar, Stripe and Riskified, and that financial services, media and technology will follow.
• Custom Agents are agents a business builds around its own processes, policies and expertise in Agent Builder, a no-code environment. The business defines the agent's job, the systems it can use, the actions it can take, and where a human has to approve.

Both can run in Zendesk or in other service environments, including Salesforce and ServiceNow, and agents can run proactively on business events or on a schedule. Zendesk also says specialized agents "automate up to 80% of workflows," and that customers ran "more than 1 million Custom Agent executions within seven weeks of launch." Those are Zendesk's numbers, not ours.

Can you use Zendesk Custom Agents yet?

None of this started on Monday. Zendesk announced Agent Builder at Relate 2026 on May 19 as "a no-code interface that enables companies to build, test, deploy, and optimize custom AI agents," marked "Early access now." The Custom Agent EAP announcement went out the day before, with rollout on May 19.

So here's where things stand...

• Custom Agents and Agent Builder are in early access. The September 14 announcement doesn't say they've reached general availability, and Zendesk's help center describes them as "currently in an early access program (EAP)" that admins turn on in Admin Center (Understanding custom agents and turning on the agent builder EAP).
• Industry Agents were announced on September 14. Zendesk hasn't published availability details for the commerce agents yet.

We've been trying Custom Agents ourselves in early access, and we like where they're headed. For real commerce work, the gap today is less about the agent itself and more about what it can reach: the order, the customer and the back office behind them.

Put together, Zendesk's Specialized AI Agents is the family: Industry Agents are the prebuilt commerce specialists, and Agent Builder is where you build Custom Agents for everything else. Zendesk's own pitch for Custom Agents is that your greatest automation opportunity might be where your business is different, and for commerce that's a fair bet. Returns, exchanges and refunds look the same on paper, but every brand's policies, platforms and exceptions don't, and those differences are exactly where a generic agent stalls.

Thomas Verschoren's Relate 2026 write-up on specialised AI agents is still the best short explanation of why Zendesk moved this way: from one general-purpose bot to a coordinated team of specialists.

How Zendesk Custom Agents work

Zendesk's Custom Agents launch post describes each agent as inputs, instructions and outputs. Its tools can include Zendesk actions, custom actions, action flows, external actions such as Slack, Jira and Salesforce, knowledge articles, and other custom agents. One agent can call another. The help center adds that "all services that are connected to the action builder are also available to be added to custom agents," and that custom agents get no tools or knowledge until an admin adds them.

One detail matters for anyone planning a rollout: custom agents are started from action flows, and they need at least a ticket ID as input. Thomas Verschoren's hands-on test of Zendesk's new Custom Agents walks through both routes: a ticket trigger that hands off to an agent, and a Run on demand flow that conversational AI agents and Copilot can call mid-conversation. He also built a post-solve audit agent that reopens a ticket when a promised action wasn't finished, which is a good example of a custom agent working behind the scenes rather than in the chat.

Custom AI agents need commerce data and actions

Zendesk's own refund example makes the case. In the Custom Agents launch post: "A refund may require checking order history, loyalty status, previous refund activity, and the company's latest policy." That's at least three systems before the agent can answer, and then it has to act, by issuing the refund, cancelling the subscription or creating the return.

The commerce Industry Agents start with Shopify, Narvar, Stripe and Riskified. That's a solid foundation. The brands we work with also run on BigCommerce, commercetools, VTEX, Adobe Commerce and more, with subscription, loyalty, reviews and returns providers alongside. A custom agent for one of those brands needs all of it: the order, the subscription, the loyalty balance and the return, plus the actions to change them.

That's what separates agents that act from agents that answer. Writing the instructions is the easy part. The hard part is the plumbing underneath: authenticated access to each system, a consistent shape for an order whichever platform it came from, and actions that do what they say. That plumbing is what we're building next at agnoStack: one secure connection to the whole retail back office, from commerce and payments to shipping, subscriptions and loyalty, so a custom agent can do more than settle a ticket or issue a refund faster. With the full customer and order picture and truly actionable options in front of it, the same agent can save the order, offer the exchange instead of the refund, or turn "where's my order?" into the next sale.

Action flows are how AI agents act

Almost all of that runs through the Zendesk action builder. Custom Agents start from action flows and can call them, conversational AI agents call Run on demand flows, and the action builder is where custom actions and connectors live. We covered last month's changes to connectors and pricing in Zendesk action flows got cheaper. The short version: native Zendesk actions no longer use action credits, including actions AI agents and Copilot invoke.

agnoStack already sits on one end of that. With Actionable Options, released in 5.34.0, a support agent can start a Zendesk action flow from the agnoStack Order Details panel, with the order in context. Because custom agents are started from action flows, a flow that a person starts from the order panel could hand that order to a custom agent for the reasoning, then write the result back to the ticket.

What's next

Custom Agents are the right idea: specialists built around how a business actually works, with clear inputs, tools and controls. For commerce brands, the work now is to make sure those agents can see the full order history, across every platform, and act on it.

Our own connectors for Zendesk AI agents and the action builder are nearing general availability, and we'll share more as they roll out. In the meantime, if you want to see how agnoStack puts live order, return, subscription and loyalty data in front of your team today, and how Actionable Options start action flows from the order panel, schedule a demo.