Case study
An AI customer support agent in Zendesk for a leading German neobroker
S-PRO built a plugin inside Zendesk for a leading German neobroker. It classifies each incoming ticket and prepares a personalised draft reply. The operator checks, edits and sends it. Ticket handling became more than 40% faster, and the AI never writes to a customer on its own.
What problem did the neobroker need to solve?
Handling support tickets took too long. The task was to shorten handling time without changing the tool the operators already work in and without taking control away from them. Ticket volumes, periods and ticket types are not published for this client.
Why a draft for the operator, and not an automatic reply?
In S-PRO's view, the person who answers a broker's customer should stay accountable for what is sent. We designed the plugin as an assistant inside the existing workflow, not as a chatbot facing the customer.
Three design principles guided the work. This is S-PRO's view, not a project fact:
- A human is responsible for every message. The AI prepares, the operator decides.
- Work where operators already work. An agent inside Zendesk needs no new system to learn.
- Start with classification. It is the most measurable part, so it gives the first reliable signal.
For the wider approach to support automation see support center optimization.
How does the solution work?
The plugin follows a five-step flow. Steps 2 and 3 are done by AI, steps 4 and 5 by a person.
- A ticket arrives in Zendesk.
- AI classifies the ticket.
- AI prepares a personalised draft reply. The draft follows the neobroker's tone-of-voice guidelines, "BaFin-relevant disclaimers" and internal compliance rules, and uses account management, transaction history, compliance records and product databases.
- The operator checks and edits the draft.
- The operator sends the reply.
The final sending is always done by the operator.
What technology is behind it?
The plugin is a Zendesk app. Two tasks go to the model: classifying tickets and generating drafts. The model is from Anthropic, accessed via Amazon Bedrock. Model version, region and hosting are not published for this client. S-PRO is a Certified partner in the Claude Partner Network. For how we work with Claude in financial institutions see Claude implementation for financial services.
What were the results?
| Metric | Result |
|---|---|
| Agent draft reply not wrong | 85% of cases |
| Ticket classification | 90% correct |
| Ticket handling | more than 40% faster |
Figures as reported; the evaluation method is not published. The 85% means that in 85% of cases the draft reply written by the agent was not wrong, as reported. We do not publish a baseline, a sample or a period, and no baseline is published for the 40%.
What can a support team in fintech take from this?
S-PRO's general view, not a project fact:
- Start with classification. It can be measured quickly and feeds everything after it.
- Set the target before the start. Agree on what "better" means before the first ticket goes through.
- Measure several things. One number hides trade-offs between speed and quality.
- Keep the person responsible. In S-PRO's view, the operator's decision is the control.
What comes next?
The usual next step is the move from pilot to a wider production: quality gates, monitoring and a rollback path. This is S-PRO's general position, not a statement about this project. See production gates for AI agents in banks. The same approach carries over to another broker, bank or fintech with Zendesk and a steady flow of tickets. See AI agent development and fintech software development.
FAQ
Questions about this project
What did S-PRO build for the neobroker?
A plugin inside Zendesk that classifies incoming support tickets and prepares a personalised draft reply. The operator checks, edits and sends the reply. The AI does not write to customers on its own, so a person stays responsible for every message.
What results did the project report?
In 85% of cases, the agent's draft reply was not wrong; ticket classification was 90% correct and ticket handling more than 40% faster. These are the figures as reported. The evaluation method is not published, and we add no baseline or period to them.
Which AI technology does the solution use?
The plugin is a Zendesk app and uses Anthropic models via Amazon Bedrock for two tasks: classifying tickets and generating draft replies. S-PRO is a Certified partner in the Claude Partner Network. Model version, region and hosting are not published for this client.
Why does the AI draft the reply and not send it?
In S-PRO's view, a person should answer for what a customer receives: the AI should prepare the work, and a person should decide and send. That also lets operators keep using the tool they already know, with no new system.
Can a bank or fintech get a similar AI support agent?
Yes, if it has a support tool such as Zendesk and a steady flow of tickets. S-PRO's design thinking pilot starts from CHF 20k and takes 2 to 4 weeks. This is the general entry offer, not the price of this project.
Support tickets piling up? Let an agent draft, and keep your people in charge.
Tell us which tool your support team works in and how tickets arrive today. We will propose a first agent scope inside that tool.
© S-PRO. The client is not named. Figures as reported.
