Zendesk AI agents vs a custom AI support agent: which resolves more?
Zendesk's AI agents are the automation Zendesk sells inside its own product: configured in Zendesk, run on Zendesk's infrastructure, and billed on Zendesk's terms. A custom AI support agent is one built for your queue, wired into Zendesk through its API, and run on a model in your own cloud account. Aldenebai, run by Paul Rahme, builds the second kind, and one runs in production inside Zendesk at a US marina-software company, resolving about 60% of tickets with no human touch, with about one in ten of those reopened (rounded from production data, 2025–2026). This page says what each is good at, with Zendesk described only from its own public material.
The comparison
| DIMENSION | ZENDESK AI AGENTS | CUSTOM AGENT INSIDE ZENDESK |
|---|---|---|
| Where it is configured | In Zendesk's admin, from Zendesk's AI agents product. | In your repositories, as code and configuration your team owns. |
| Where the model runs | On Zendesk's infrastructure and providers. | In your own cloud account, for example AWS Bedrock; your data is not used to train the models and Aldenebai holds no keys. |
| Knowledge | Your help centre and the content you connect. | Your documentation plus any system it is allowed to read: account state, order status, known issues. |
| What it may do | The actions Zendesk exposes and you enable. | An allowed-action list you write: reply, tag, close, escalate. Refunds and account changes never, enforced by permissions. |
| Threshold and escalation | Zendesk's controls, as documented by Zendesk. | A confidence threshold your team sets, with escalations that carry the intent, sources, draft, and reason. |
| Cost model | Zendesk's published pricing for AI agents, on its pricing page; check the current terms there. | A fixed-price build you own, then model usage on your own bill with a ceiling. No licence. |
| Measurement | Zendesk's own reporting. | Ordinary Zendesk reports on tags the agent sets, so resolution and escalation rates are yours to audit ticket by ticket. |
| Time to start | Fast: switch on and configure. | A scoped pilot on the top ticket types; at the client, the first live ticket came about a month after kickoff, and your date is written into the proposal. |
| Best for | Teams that want a vendor-managed feature with no engineering, on Zendesk's terms. | Teams that need control over actions, data residency, and cost, or answers that read systems beyond the help centre. |
What a custom agent inside Zendesk adds
Three things a built-in feature cannot give you by design. First, an allowed-action list enforced by permissions rather than by settings, so a wrong answer can never become a refund. Second, a model running in your own account, which settles the data question with your security team in one sentence. Third, answers drawn from systems beyond the help centre, such as account state or order status, which is where the routine tickets that a help-centre-only agent cannot close tend to live.
The wiring is ordinary Zendesk: triggers, the API, tags, groups. It is written out in the Zendesk on AWS Bedrock guide, and the method around it in how to automate customer support with AI.
When Zendesk's own AI agents are the right call
When the queue is mostly questions your help centre already answers, when nobody on the team will own a threshold or read escalations, when procurement prefers one vendor invoice, and when there is no data-residency requirement. In that situation a vendor-managed feature is faster to switch on and cheaper to think about, and a consultant who says otherwise is selling. The honest test is the one the Automation Audit runs: which tickets are routine, what they need to be answered, and what a wrong answer would cost.
You can also run both. The built-in feature keeps deflecting pre-sales questions while a custom agent works the tickets that need systems access, and the tags tell you after a few weeks which is earning its place.
Questions Zendesk teams ask
Can a custom agent run alongside Zendesk's AI agents?
Yes. The custom agent sees only the tickets your triggers send it, usually one group or form, so the two can share a queue while the tags show which one resolves what. Nothing is ripped out, and the custom agent's kill switch is disabling one trigger.
Do we lose Zendesk reporting with a custom agent?
No. The custom agent sets tags for every outcome, so resolution rate, escalation rate, and first-response time are ordinary Zendesk reports and views, auditable ticket by ticket.
What does a custom agent cost compared with the built-in feature?
They are priced differently rather than higher or lower: Zendesk's AI agents follow Zendesk's published pricing, while a custom agent is a fixed-price build you own followed by model usage on your own bill with a ceiling. The Automation Audit puts both against your ticket volume before you decide.
What resolution rate should we expect?
About 60% of tickets resolved with no human touch is the production figure for the custom agent Paul Rahme built inside Zendesk at a US marina-software company, 2025–2026, and about one in ten of those tickets is reopened, the companion figure to ask any vendor for. Your own rate depends on how much of your queue has written answers; the audit estimates it from your queue.
Related pages
Bring your top ten ticket types. Leave with the answer.
The free Automation Audit sorts your queue into routine and not, and says which kind of agent each side needs. 30 minutes, written Automation Map within 48 hours.