The biggest mistake companies make when building AI agents

The starting point is the job

One of the first assumptions we challenged while building Industry Agents was that the starting point was AI. It wasn’t. The starting point was the job.

When organizations begin exploring AI, they often ask, “What kind of AI agent should we build?” We think that is the wrong question. The better question is: What job are we trying to automate?

That one shift changes almost every design decision that follows. Customers do not measure success by how intelligent an AI agent sounds. They measure it by whether the work gets done.

The problem is not understanding the question. It is doing the work.

Take something as simple as a return. Nobody actually wants to automate “returns.” They want to automate the decisions behind the return.

Does the customer qualify? Is the product still within policy? Should an exception be made? Is inventory available to fulfill an exchange? Is there a risk signal that should change the outcome? Does someone need to approve it? What action should happen next?

That is the job. The conversation is simply how the work begins.

A general-purpose AI model can understand language and reason about a request. But understanding what a customer said is not the same as understanding what the business should do next. The work that creates the most value is often the work behind the conversation.

We think of this as the messy middle of service. It is where business context, policies, systems, actions, and judgment come together.

AI does not need to learn your industry every time

Retailers do not all operate the same way. But they all process returns. Banks do not all have the same products. But they all investigate disputes. Technology companies do not all build the same software. But they all troubleshoot issues and manage accounts.

The job is often remarkably consistent. The business rules are what make each company different.

That is why Industry Agents start with expertise for the work businesses in an industry have in common. Our first Industry Agents focus on ecommerce work such as shopping and product discovery, order management, returns and exchanges, refunds, fulfillment and delivery, billing and payments, promotions, technical support, and account authentication.

The goal is not to force every business into one operating model. It is to give them a head start.

Our first Industry Agents focus on ecommerce work such as shopping and product discovery, order management, returns and exchanges, refunds, fulfillment and delivery, billing and payments, promotions, technical support, and account authentication.

Industry expertise gets you started. Business expertise finishes the job.

A retailer might allow returns within 30 days while another allows 60. One may automatically approve a replacement for a long-standing customer while another requires approval above a certain order value.

Those are not unusual edge cases. They are the business.

Industry Agents are designed to work with that reality. They can use business knowledge and policies, connect to the systems where the relevant data lives, and operate within configured guardrails and approval paths.

This is also where Industry Agents and Custom Agents fit together. Industry Agents provide the starting point for common industry work. Custom Agents handle the policies, decisions, exceptions, and workflows that are unique to a specific business.

Industry Agents understand what the industry has in common. Custom Agents understand what makes your business different.

If I were deploying Industry Agents today

I would not start by asking which AI capability to deploy. I would start by mapping the work.

Which requests does the team handle hundreds or thousands of times a week? Which decisions require business context? Which workflows cross multiple systems? Which jobs have a clear definition of done? Which processes require interpretation or a decision rather than a fixed sequence of steps?

Those are the areas where I would look first.

If every step follows exactly the same path, traditional automation may be the better answer. If the workflow requires context, interpretation, or a choice between paths, an agent can be a better fit.

The goal is not to maximize the number of agents. It is to automate the right work.

A closer look: a refund request

Consider a customer who says, “My mountain bike’s rear derailleur broke on my first ride. I’d like to return it and exchange it for the newer TrailPro model.”

The agent first identifies the customer and the order. It then gathers the context needed to make the decision: purchase history, delivery details, return and warranty policy, inventory, payment information, and relevant risk signals.

The Industry Agent can use Shopify for product and order details, Narvar for delivery and returns information, Stripe for payment information, and Riskified for risk intelligence. Zendesk provides the conversation and customer context, along with the knowledge and business rules that determine how the request should be handled.

If the request meets the configured criteria, the agent can initiate the return, provide the customer with the relevant instructions, and communicate what happens next. If it does not, the agent can ask for more information, route the request for review, or hand it to a human with the gathered context intact.

The customer sees one conversation. Behind it is a coordinated workflow.

Four things we learned building Industry Agents

1. The policy matters more than the prompt. A useful agent needs to know who qualifies, what counts as an exception, when it should act, and when a person should review the request. Those decisions need to be explicit.

2. The work spans more systems than you think. A request that looks simple to a customer can require CRM, order management, payment, inventory, fraud, delivery, and internal policy systems. The customer does not care where the data lives. They care that the job gets done.

3. Specialization makes AI easier to control. A defined job gives you a smaller surface area to test, measure, govern, and improve. It also makes it easier to understand where the agent should stop and a human or another specialist should take over.

4. Build for change, not just launch. Building an agent is not the finish line. Businesses change. Policies change. Products change. Customers change. Industry Agents are built within Zendesk so teams can build, test, deploy, monitor, and improve them in the same environment. Testing and simulation can expose problems before deployment, while traceability helps teams understand how an agent reached an outcome and which knowledge or tools it used.

Once an agent is in production, performance and outcome data show where it is succeeding, where it is getting stuck, and where it needs to change. Those signals can feed the Zendesk Resolution Learning Loop, helping teams continuously improve the agent’s instructions, knowledge, policies, and actions as the business evolves.

The opportunity is bigger than ecommerce

We are starting with ecommerce because the value of specialized service is particularly clear. Retailers manage high volumes of repeatable work that directly affects revenue, cost, and customer relationships.

But the model is broader. The same approach can apply to financial services, technology, healthcare, and other industries. The work is different. The design principles are the same.

Start with the work an industry has in common. Build the expertise into the agent. Connect it to the systems required to finish the job. Then give the business control over the parts that are unique to it.

What we are really building

We do not think the future of service is one general-purpose agent trying to do everything. We think it is a connected workforce of specialists.

Each specialist has a job. Each one understands the context required to do that job. Each one can use the systems and actions it needs. Together, they can move a customer request from conversation to outcome.

That is why we built Industry Agents. Not to give businesses another AI framework to configure, and not to make them rebuild expertise that already exists across their industry.

Start with the job. Give the agent the right context. Connect it to the systems where the work happens. Define the rules. Make the outcome clear. Then keep improving it.

That is how industry expertise becomes action. “The unit of design is not the model. It is the job.”