Zendesk rebuilds for a world where bots handle your customer service

Zendesk, a customer-service software vendor of nearly two decades, has launched the Outcome AI Platform, a renamed and expanded version of its core product, built for a market the company thinks is coming: one where AI agents, not people, do most of the talking to support systems 1

1

2

2

.

Shashi Upadhyay, Zendesk's president of product, engineering and AI, forecasts that within a couple of years more than half of customer service traffic will be agent-to-agent 2

2

. Personal assistants such as Instinct and Meta Muse already book travel and shop for people. He believes agents like these will become the main way people consume support 1

1

2

2

.

Building for the agents

The new Zendesk Bridge connects outside AI agents, including ChatGPT, Claude and Microsoft Copilot, directly to the platform. An issue can then be resolved inside whatever tool a person is already using 1

1

2

2

. Upadhyay rejects the idea that this is just APIs with a new name. "I think the way people talk about headless, it's often like, 'I'm just going to repackage all my APIs and call it headless.' That is not what we have in mind. What we have in mind is, it's actually a full-on development platform," he said 1

1

. Per Zendesk, the system inherits company permissions, so an agent sees only what its user is allowed to see 1

1

.

Alongside it sits the Open Service Directory, an open specification and public directory where businesses register verified support channels 1

1

2

2

. It works whether or not a brand uses Zendesk 2

2

. Seven companies, including Betterment and ElevenLabs, are initial participants 1

1

. Upadhyay says the reason is fraud: without a verified directory, fraud and phishing would enter bot-to-bot service and ruin its potential 1

1

. "If we don't do something like this, then all the fraud, the phishing, all that stuff enters the space and mucks it up," he said 1

1

.

Getting paid for outcomes

The pricing changes with it. Zendesk is expanding outcome-based pricing that drops per-seat fees in favor of unlimited seats. What it charges ties to metrics each customer picks, such as retention, conversion or dollars collected in a collections operation 1

1

2

2

. "The unit of measurement is going to be outcomes," Upadhyay said 2

2

. Some of its autonomous agent products were already priced this way, so this is an expansion of an existing model 2

2

. Because Zendesk goes unpaid without results, he argues, specialization becomes the vendor's problem rather than the customer's. The company says the specialized agents it launched are trained on nearly 20 years of ticket data, one sector at a time 1

1

. Defining the right outcome to price on is an acknowledged challenge, and Zendesk's answer is to ask customers which metrics they already run on 1

1

.

Zendesk says the rename tracks the widening scope: the former Resolution Platform now covers back-office, mid-office and employee service, where resolution is not always the right measure of success 1

1

. Amazon Web Services, whose strategic partnership with Zendesk dates to December 2025, endorsed the approach 2

2

. "When AI does real work, the fairest way to measure its value is by what it accomplishes, not by how many people have a license," said Allison Johnson, a director there 2

2

. Zendesk customers already automate up to 80 percent of interactions using Amazon's Bedrock and Connect 2

2

.

The agent-to-agent future the whole platform depends on is still a forecast. The diginomica columnist argues that a basic agent gateway alone will not preserve the value of application infrastructure. He warns of a possible scenario where a personal agent that half-understands a problem meets a support agent lacking the resources to fix an edge case 1

1

. Zendesk has priced itself on that world arriving; the directory, the bridge and the outcome contracts are the wager that it will.