Guides · Worked example

When does an AI support agent pay for itself?

Short answer

Roughly above 200 tickets a month, and comfortably above 1,000. Below that a better macro set and a tidier help centre will beat a build on cost. The return comes from deflection on repetitive categories plus faster handling on the rest, and it depends almost entirely on how much of your volume is genuinely repetitive.

~200/mo
Rough floor below which a build does not repay
30–50%
Faster resolution on our support triage deployments
Deflection
The metric to agree before anything is built
Updated August 2026 · Written for people scoping a project, not for search engines

Two savings, not one

Support automation returns money in two different ways and they behave differently. Deflection removes a ticket from a human entirely — the agent answers, the customer is satisfied, no one touches it. Assistance keeps the human but shortens the work: the ticket arrives classified, prioritised, routed, with a draft reply and the relevant history attached.

Deflection is the bigger saving per ticket and the smaller share of volume. Assistance is the smaller saving per ticket across nearly all of it. Most business cases that disappoint were built on deflection alone, and most that hold up counted both.

The arithmetic

Take a team handling 3,000 tickets a month at an average of eight minutes, loaded cost €30 an hour. That is 400 hours, about €12,000 a month.

A realistic deployment on a mixed ticket base deflects 20–30% of volume outright — order status, password resets, delivery windows, opening hours — and shortens the remainder by a couple of minutes each through classification, routing and drafting. That combination typically removes 30–40% of the total hours. On these numbers, in the region of €4,000–5,000 a month, against a build that is a small multiple of one month of that saving.

The deflection share is the number to be sceptical about in any proposal, including ours. It is knowable in advance: it is roughly the share of your last six months of tickets that fall into repetitive, answerable categories. Ask for it to be measured, not estimated.

What determines whether it works

01

How repetitive your volume actually is. A support base dominated by unique technical problems deflects poorly no matter how good the agent is. Six months of ticket history will tell you before you spend anything.

02

Whether the answers are written down. The agent grounds its replies in your help centre, macros and past resolutions. If the real answers live in senior agents’ heads, that content is the first phase of the project.

03

Where the escalation line sits. Refunds, account actions and anything legally sensitive should route to a person by rule. Every rule of that kind lowers deflection and raises trust; that trade is a business decision, not a technical one.

04

Whether quality is measured. CSAT on automated replies and reopen rate, tracked separately from human replies. Without them, deflection is just a number going up.

The cost side people forget

Model usage scales with ticket volume and is usually a small line, but it is not zero and it should be in the model. More significant is the review time in the first weeks: someone senior reads a sample of automated replies daily until the escalation rules are tuned. Budget a few hours a week for a month.

There is also a quality risk with a real cost. An agent that answers confidently and wrongly is worse than no agent, which is why confidence thresholds start conservative and are loosened only against measured reopen rates.

When not to build one

Under a couple of hundred tickets a month the saving cannot cover the build; better macros and a clearer help centre will. If your ticket base is genuinely non-repetitive, the deflection assumption collapses and so does the case. And if answers depend on judgement nobody has written down, the agent has nothing to ground on — documenting it is the project, and it has value even if no agent is ever built.

Follow-up questions

What people ask next.

That is your policy decision, and both options are defensible. What we will not build is a system designed to conceal it while implying a human, because it breaks trust the first time a customer notices.

Still unsure whether your process is worth automating?

Bring us the process. We take it apart with you at no charge and give you a straight answer, including when the answer is no.