Zendesk Uses ChatGPT Enterprise to Build a Sales Agent with 1,000 Daily Active Users

How can enterprises make ChatGPT Enterprise truly work? Zendesk's practice offers a sample worth examining closely. The company not only opened ChatGPT Enterprise to all employees but also encouraged business users to build roughly 600 apps on their own. Among them, an agent for the sales team achieved about 1,000 daily active users—and its starting point was a simple question: "Tell me what you built."

An AI foundation available to everyone: business users are no longer "afraid" of ChatGPT

According to experience shared by Zendesk on YouTube, the company opened ChatGPT Enterprise to all employees, and users in general business roles such as sales, marketing, finance, and operations can now naturally ask ChatGPT questions directly. This sense of "comfort" is a key prerequisite for AI tools to move from pilots to scale—only when non-technical roles are willing to actively interact with the model does the mining of data value truly begin.

Unlike many enterprises that push AI "top-down," Zendesk shows a "bottom-up" vitality: about 600 apps were built by different people across the company. This means AI capabilities have penetrated to the edges of the business—frontline employees can put ideas into action whenever they have them.

ChatGPT Enterprise is the version OpenAI launched for enterprise customers. Its core differences from consumer ChatGPT are: by default it does not use enterprise data to train models; it provides an admin console (member management, SSO, usage analytics); it offers a larger context window and data analysis capabilities; and it can connect to internal enterprise systems via API or connectors. It is itself a general-purpose foundation and does not come with built-in industry-specific business logic, so how much value an enterprise can realize often depends on whether it can connect internal data and package it into applications tailored to job scenarios. This also explains why "opening it to everyone" is the first step—only when non-technical roles are willing to use it frequently will the enterprise have enough usage samples to discover scenarios worth codifying.

Sales agent: "calculating" expansion opportunities from data

One app highlighted in the video is an agent used by the sales team to identify customer expansion opportunities. Its logic is not complex but very practical: it proactively digs into company data to find customers who have "almost used up the seats they have" and who have "a high ticket volume per seat," then pushes the leads directly to salespeople, along with solid value arguments explaining why expanding Zendesk seats benefits the customer.

The value of this agent lies in turning usage data that was previously dormant in the system into actionable recommendations that sales can follow up on directly. It is not just a Q&A bot but an analysis assistant that can proactively scan, judge, and output business conclusions. Currently, the app has about 1,000 daily active users—quite substantial for an internal tool.

The key difference between an agent and large-model Q&A lies in "proactiveness" and "tool use": it does not just passively answer questions but can, according to preset goals, call data sources, perform multi-step judgments, and produce action recommendations. In this example, the agent first needs to understand the business meaning of metrics such as "seat utilization" and "tickets per seat," then filter and rank customer data, and finally generate persuasive value arguments—this usually requires connecting the enterprise data warehouse or CRM with model capabilities and codifying the judgment logic with prompts or workflows. Compared with general chat, agents have a higher barrier to entry, but their output is also closer to directly executable business conclusions.

The combination of data and scenarios is the real moat for enterprise AI

Zendesk's case shows that the effectiveness of ChatGPT Enterprise does not depend on how "smart" the model itself is, but on whether the enterprise can connect internal data with specific business scenarios. The reason the sales expansion agent is useful is that it clearly knows which metrics indicate opportunity—seat utilization and ticket density are products of Zendesk's own business logic.

A detail worth mentioning is how employees participate. The roughly 600 apps were not uniformly developed by a central AI team but grew up in a distributed way across business units. This model places higher demands on enterprise data governance and platform support, but it also brings AI applications closer to real pain points.

Data governance is the most easily underestimated part of "bottom-up" app building. When hundreds of apps are developed in a distributed way by different business users, enterprises need to solve several types of problems: first, permission boundaries, ensuring each app can only access data the user was already authorized to view; second, unified data definitions, avoiding conflicting conclusions caused by different teams defining metrics such as "seat utilization" inconsistently; third, traceability and auditing, clarifying what an app output and on what basis. On the platform side, it is usually necessary to provide unified identity authentication, data connectors, and an app catalog so business users can build self-service without crossing security red lines. The more solid the governance, the lower the trial-and-error cost of decentralized innovation.

"The cost of waiting is greater than trial and error": don't stand on the sidelines

There is a clear statement in the video: "Because if you wait, you'll lose more than you gain." This can be seen as Zendesk's direct advice to enterprises. At a stage when AI tools are iterating rapidly, waiting excessively for a "perfect solution" often means missing the window for organizational learning and data accumulation. Zendesk chose to first get everyone using it, then continuously refine it in specific scenarios.

Judging from the results, this approach has at least been validated on one sales agent: about 1,000 daily active users means it has become a genuinely functioning part of the sales process. For companies evaluating ChatGPT Enterprise or other enterprise AI platforms, Zendesk's experience points to a simple conclusion—putting tools in the hands of business users and letting them build from their own data and problems often creates more value than waiting for a perfect plan.