A support agent that follows policy and hands off hard cases
A custom support agent for order, delivery, sizing, return, and refund questions, with fixed policy checks and a clear path to a person.

- Client
- Dinsko
- Timeline
- 5 weeks
- Tools
- Claude Haiku, Voyage AI, PostgreSQL, pgvector, Supabase
Overview
The goal was to resolve customer work
Dinsko handles about 18,000 orders and 4,200 support tickets each month. Three agents covered questions about deliveries, sizing, returns, and refunds. At NocodeAssistant, we built a custom support agent in five weeks.
Dinsko had already tried Intercom Fin. It gave standard return windows for sale and custom items and sent customers to a tracking page instead of checking the order. A quick answer could still leave the full task with a person.
Challenge
A useful answer needed live data and exact rules
The agent needed access to live orders and carrier data, plus the exact return rules. A managed tool would have meant less software to maintain. A custom build gave the team more control over the work from a customer question to a completed request, with more responsibility for integrations, testing, and policy updates.
Approach
The model selects tools; code applies policy
Claude Haiku interprets a request and selects tools. Connected systems retrieve order and carrier details. Code checks item type, return window, and refund path. The agent asks the customer to confirm before it submits a refund request, and records that confirmation.
Failed or repeated requests go to a person with the conversation summary, order context, and a suggested category. Streaming shows when an order lookup or eligibility check is in progress.
Search combines keyword matching with Voyage AI embeddings over 23 help documents in Swedish, Norwegian, and English. The documents cover returns, sizing, care, warranty, and shipping.
Evaluation
Tests found mistakes in retrieval and instructions
An early retrieval benchmark compared document titles and scored correct chunks incorrectly. We changed it to compare document IDs. Another test showed the agent following a sales instruction found in retrieved product copy during a return conversation. We changed the prompt to keep reference text separate from instructions.
The team built 40 adversarial cases for ambiguous requests, multiple requests, prompt injection, and questions outside the agent’s scope. Release checks blocked changes that fell below set retrieval and adversarial thresholds.
Result
The agent resolved about a third of all tickets
After 90 days, it resolved 47% of tier-one tickets without a person, about one-third of all tickets. Median first response fell from 4.2 hours to 8 seconds. Monthly support team cost fell from €14,200 to €9,200, a 35% reduction.
One employee moved from support to customer retention, and their salary moved to that budget. The custom agent cost about €380 per month, included in the support cost figure.
Reflection
Keep policy fixed and make handoffs useful
The model can manage a conversation and select tools, but it should not set refund policy. The system works best when it completes repeatable work and gives support staff the context they need for the rest.