On 10 September, Salesforce closed its acquisition of Fin. Four days later, Zendesk announced Specialized AI Agents. The two biggest names in helpdesk software both moved in the same week, and if you run support for a small team you are entitled to ask what, if anything, you are supposed to do about it.
Here is the honest answer for the Zendesk half: it is a serious release, it is aimed at a bigger company than yours, and the interesting part is not the feature list.
The release is dated 14 September 2026 and sits under “Press releases” in Zendesk’s newsroom. In their own words:
“Today, Zendesk introduced Specialized AI Agents, purpose-built for the work that matters most to a specific business or industry.”
It splits into two things, each with its own section heading in the release:
Zendesk says customers have run “more than 1 million Custom Agent executions within seven weeks of launch”, that some are “seeing automated resolution rates spike 10% post-deployment”, and that the agents can “automate up to 80% of workflows”. Those are Zendesk’s figures about Zendesk’s product, reported here as theirs. The release also says the agents can run inside Salesforce and ServiceNow, not only Zendesk.
One thing the release does not do is mention plans or pricing. Not a tier, not a rate, not a bundle. Anyone telling you what Specialized AI Agents cost is not getting it from this announcement.
It is this one, from Shashi Upadhyay, Zendesk’s President of Product, Engineering and AI:
“The era of generic, one-size-fits-all AI agents is over. The best agents are specialists: deeply grounded in the work, the industry, and the business they serve.”
Strip the launch language and that is a concession worth having in writing. A general-purpose AI agent pointed at your help centre does not know your refund window, your approval chain, or which systems it is allowed to touch. Zendesk’s answer is to sell you agents that are told, in advance, what job they own.
That is the right diagnosis. It also tells you exactly what the work is: somebody has to define the job, the permitted systems, the allowed actions and the approval points. Agent Builder being no-code changes who can do that work. It does not remove it.
Three practical readings.
1. Industry Agents are an enterprise commerce play first. The named integrations — Shopify, Narvar, Stripe, Riskified — describe a retailer with a returns policy, a fraud problem and enough volume for both to matter. If you are a ten-person SaaS company, the commerce agent is not built for your work, and the industries that might be are on a roadmap rather than in the product.
2. The unit of work is the workflow, not the ticket. The release is about agents that act across connected systems, not agents that answer. That is a real capability and a real obligation: every system you connect is a system an agent can act in, and every action you allow is one you are accountable for. The approval points are not a formality.
3. Ask what happens before the agent is specialised. A specialist agent is only as good as the knowledge and the permissions it was given. If your help centre is thin, or half your answers live in someone’s head, an agent builder gives you a faster way to automate an answer you have not written yet.
DeskLeap is not competing for the commerce-automation buyer this release is aimed at, and we would rather say so than pretend otherwise. What we do is narrower: ticketing, live chat, a knowledge base and SLA policies, with an optional AI Agent that answers from your knowledge base and your site content.
The pricing is the part most people want stated plainly, so: the platform is 100% free with unlimited human agents; an optional AI Agent add-on is $29/month each. Each paid AI Agent includes 500 replies a month. There are no seat tiers to climb, because there are no tiers.
What the AI Agent is genuinely for is the gap Zendesk’s release describes from the other end — not the multi-system commerce workflow, but the ordinary question that arrives when nobody is at a desk. It keeps working nights, weekends and holidays, and a first-contact review gate means a person may still be the one who replies.
If you are evaluating either, the useful question is not which AI is better. It is which of them is aimed at work you actually have.