Most helpdesks now ship an AI agent, and every week another startup offers one that sits on top. So the question we get on discovery calls has changed from "can AI answer tickets?" to "do we use what our helpdesk has, buy a specialist tool, or build our own?" The honest answer depends on two things: what your tickets need the agent to do, and how many of them you have.
§ 01When the helpdesk's AI agent is enough
- Most tickets are questions with an answer in your help center or policies.
- The actions you need, such as tagging, routing or closing, live inside the helpdesk itself.
- Volume is moderate, so per-resolution pricing stays reasonable.
- You do not need custom approval rules or detailed logs of why the agent did something.
In that situation, building is usually a waste of money. Turn on the built-in agent, feed it a clean help center, and measure.
§ 02When a custom agent pays off
- The tickets need actions in other systems. Checking the order in the store, the subscription in billing and the parcel with the carrier, then issuing a refund, is several systems the helpdesk does not own.
- Your rules are specific. Refund limits by customer type, approval by value, different handling for wholesale accounts. Off-the-shelf tools tend to offer a few switches, not your policy.
- You need to explain every action. Regulated industries, finance and anything involving money need a record of what the agent checked and why it acted.
- Volume makes per-resolution pricing expensive. At scale, a fixed build plus running costs can come out cheaper than paying per resolved conversation. Do the arithmetic with your own numbers.
- You do not have a knowledge base or written SOPs. Off-the-shelf agents depend on them. A custom build starts by writing the rules down with you, which is often half the value.
§ 03Questions to ask any vendor
- Which actions can the agent take in which systems, and which are read-only?
- How do I set limits, such as value thresholds or human-only request types, and are they enforced in code or in the prompt?
- What exactly counts as a resolution for billing? Does a customer who gives up count?
- Can I switch off one workflow without switching off the whole agent?
- Where are the logs, how long are they kept, and can I export them?
- What happens to my data, and is it used to train anything?
A vendor that answers these clearly is worth considering, whatever the price. One that answers with a demo is not.
§ 04The hybrid that most teams end up with
In practice many teams use both. The helpdesk AI handles FAQ-style answers. A custom agent handles the action-heavy categories, such as refunds, cancellations and order changes, where rules and money matter. The important part is that each ticket type has one clear owner, so the two never answer the same customer. We describe the action-heavy side in AI agents for refunds, cancellations and failed payments and the handoff rules in when should an AI support agent hand off to a human.
§ 05Cost
A custom agent covering one or two ticket categories, integrated with your helpdesk and one back-end system, costs 3,000 to 5,000 USD and takes 2 to 3 weeks. Running costs are typically 30 to 250 USD per month for infrastructure and model usage. Compare that with the vendor's price at your real monthly ticket volume, not at the volume in the demo. The full numbers are in our cost guide, and the general support picture is in AI agents for customer support.
──
If you are choosing between your helpdesk's AI and a custom agent, a 30 minute discovery call with your ticket export will usually settle it. We will tell you if the built-in option is enough. Engagement models and prices are on our Hire page.