AI support agent in a company with a legacy ERP
We categorised the support flow and built an agent that drafts answers. Most drafts could be sent as is — and the project showed where the real bottleneck was.
Project facts
- Area: Customer support
- Scope: categorisation, answer drafts, review flow
- Hand-over: completed, continued by the client's IT team
The situation
The support team handled a steady flow of customer requests by email. Agents wrote every answer from scratch, and the team wanted to reduce the time spent per ticket.
What we built
- Analysis of historical tickets and a category structure covering the main request types
- Handling instructions for each category
- An agent that drafts answers using the knowledge base and category rules
- A review flow where agents send, edit or reject each draft
Results
For the main categories, about 80% of drafts were good enough to send as is. The analysis also showed something the team hadn't measured before: most tickets required not only an answer but also manual operations in the company's ERP system, which had no usable API. That part could not be automated, so the gain in overall handling speed was smaller than the drafting quality alone suggested.
What happened next
The project was handed over to the client's internal IT team without issues — including the AI engineers we had hired and trained for them. They continued developing it on top of the categorisation, instructions and drafting pipeline we built.
Stack
Want something similar?
Results depend on each company's processes, systems and data. Project descriptions are not a guarantee of similar results.