AI agents for customer service and operations
We build agents that read incoming tickets, emails and forms, pull the relevant order, policy or account data and draft a complete answer for your team. Your people handle fewer lookups and spend their time on the cases that need judgement.
What it is and when it fits.
Customer operations agents sit inside your support and back-office queues. They classify each request, gather the facts from your order system, CRM or policy documents, and draft a reply or next action. Depending on the risk, the draft goes to a person for review or is sent directly. Claims, returns, delivery questions and account changes are typical starting points.
This fits teams with a steady volume of repeat questions that still need data from several systems to answer well. If most of your tickets are one-line questions a help centre article already answers, a simpler self-service setup is cheaper. If every case is unique and high stakes, the agent should stay in drafting and research mode, and we design it that way from the start.
What we build.
Reply drafting
Drafts written in your tone from order, account and policy data, placed in the ticket for an agent to check and send.
Triage and routing
Classification by intent, urgency, language and customer value, routing each request to the right queue with a short summary.
Claims and returns handling
Agents that check a claim against policy terms and evidence, flag missing information and prepare a decision for review.
Order and account actions
Address changes, cancellations and refunds prepared by the agent and executed after approval, within limits you set.
Quality monitoring
Dashboards showing edit rates, escalations and answer accuracy per category, so you can see where the agent helps and where it struggles.
How it works.
- 01
Analyse the queue
We review a sample of real tickets to find the categories with the most volume, the clearest rules and the data needed to answer them.
- 02
Build against past tickets
Historical cases and your best human replies become the eval set the agent is measured against before it touches a live queue.
- 03
Go live with review
The agent drafts inside your helpdesk while your team reviews, and categories move to direct sending only when the numbers support it.
- 04
Tune as volume grows
We add categories, languages and actions over time, using the edits your team makes as the main signal for improvement.
Related work.
- Insurance
How LLM evals let an insurance claims assistant ship weekly at 94% accuracy
94%answer accuracy, tested on every release- LLM evals
- Langfuse
- Promptfoo
- Sales
AI lead qualification that cut first response time from a day to 4 minutes
4 minto first response, down from a day- Lead scoring
- Workflow automation
- n8n
Built with.
All technologiesFurther reading.
Human-in-the-loop AI agents: design patterns for production
The patterns we use to run AI agents safely in production: approval checkpoints, confidence thresholds, tool permissions, audit trails and fallbacks.
4 min read
How to build a production RAG chatbot: a practical guide
What it takes to turn a promising RAG experiment into a chatbot people trust: ingestion, hybrid search, citations, permissions and evals.
5 min read
Common questions.
It changes the work. Agents take over lookups, first drafts and routing, and your team reviews, handles exceptions and deals with the conversations that need a person. Most teams use the time for faster responses and harder cases.
We regularly work with Zendesk, HubSpot, Salesforce and Microsoft 365 mailboxes, and we build custom connectors for in-house order and claims systems. The agent works inside the tools your team already uses.
We minimise what is sent to models, redact where needed, keep logs in your environment and can use EU-hosted models so data stays in Europe. Access follows the same roles your helpdesk already enforces.
Yes. Current models handle Dutch, English, German and French well. We add evals per language, because quality can differ between them, and route languages with weaker results to human review.