The problem? Nobody could confirm those numbers were real.
Tickets had labels. But a ticket tagged "order status" might actually be a lost package complaint. A ticket tagged "accounting" might be a cancellation that also needs a credit note. Over time, these inconsistencies compounded and leadership found themselves unable to answer three questions that should have been straightforward:
Are we categorizing our tickets correctly?
Which issues are actually consuming the most of our team's time?
Where do we invest in automation first?
Without answers, a planned investment in AI-powered support had stalled. There was no validated data to build a business case on. And with volume growing across 11 languages, hiring more agents felt like the only option left on the table an expensive one that wouldn't fix the underlying problem.
Before writing a single line of code, we needed to know what was actually happening.
We started with a discovery phase alongside Just Carpets' support staff, mapping real workflows, not the documented process, but the actual day-to-day reality of how tickets were handled and how labels got applied.
Then we exported everything. 100,000+ tickets. Customer messages, agent responses, internal notes, existing tags. All of it structured into a single dataset.
Step 1 — Semantic ticket analysis at scale.
We deployed a custom LLM agent to read every ticket the way a human analyst would, understanding intent and context, not just matching keywords. For each ticket, the agent verified whether the existing tag matched the actual content, suggested reclassifications where it didn't, detected product and service mentions for more accurate filtering, and scored sentiment progression across the full conversation arc.
Step 2 — Root cause and complexity mapping.
Beyond classification, we looked for patterns. Which ticket types required the most back-and-forth? Where did customer sentiment deteriorate mid-conversation? Which issues clustered around specific carriers, regions, or time periods? This layer of analysis revealed the bottlenecks that tag-based reporting had been hiding for years.
Step 3 — Automation opportunity ranking.
With accurate tagging restored and complexity mapped, we could rank every ticket category by volume, resolution time, sentiment impact, and automation ROI. Instead of guessing where to invest, Just Carpets had a data-backed roadmap ranked by business impact.
The discovery phase didn't just fix the data. It created the blueprint for everything built next.
With the roadmap in hand, we built what Just Carpets actually needed: not a chatbot, but an intelligent AI agent system capable of handling customer support tickets end-to-end, across all 11 languages, around the clock.
The architecture: an orchestrator with specialized sub-agents.
Rather than one large model trying to do everything, we built an orchestrator that understands what a customer needs and routes the request to the right specialist. Each sub-agent is optimized for exactly one job:
The order status agent queries the order API and interprets shipment data to give customers accurate, real-time delivery updates.
The invoice agent navigates ExactOnline's document structure, retrieves the correct PDF, and sends it, a process that previously required multiple manual steps per ticket.
The credit invoice agent looks up the original invoice, inverts line items to create a credit note, posts it back to ExactOnline, and logs it, reducing what was a fully manual financial workflow to a single human approval click.
The orchestrator carries hard rules that no sub-agent can override: never surface raw API output to a customer, escalate to a human when something is unclear, and always respond in the customer's language with the right tone.
On top of the agent system, we deployed a permanent intelligence layer, a live analysis tool that processes every new incoming ticket in real time. It continuously validates tag accuracy, tracks agent performance, scores conversation sentiment, and surfaces emerging support topics before they become major problems. Just Carpets' leadership now has a living dashboard of what customers are actually saying, not just what tickets are labeled as.
A note on cost — because it matters.
Solutions like Intercom's AI charge up to $0.99 per resolved conversation. At Just Carpets' ticket volumes, that adds up to tens of thousands of euros per year for a system that doesn't integrate with your ERP, doesn't create financial documents, and doesn't adapt to your specific workflows. By owning the stack, the cost per resolved ticket drops to a fraction of that. The intelligence stays inside your business. And the system grows with you without the per-ticket price tag scaling against you.
The impact landed across the entire support operation, from the data layer up to leadership decision-making.
Ticket classification went from unreliable to trustworthy. For the first time, Just Carpets' reporting reflected what customers were actually writing about, giving leadership a foundation for decisions they could stand behind.
The AI agent now handles the ticket types that previously consumed the most manual time. Order status inquiries, invoice requests, and credit note creation, processes that each required multiple system logins, manual lookups, and hand-written responses, are now resolved automatically or with a single approval click.
The support team shifted their focus from repetitive lookups to complex cases that genuinely need human judgment. And with the live intelligence layer running continuously, the automation roadmap doesn't go stale, it updates itself as customer needs evolve.
As Gijsbert, COO at Just Carpets, described it: the What's Next team worked as an extension of their own team. The focus was never on deploying tools for the sake of it. It was on understanding the operation first, then building something that actually works inside it.
The total project took 12 weeks. Six weeks of discovery and analysis. Six weeks of building and deploying. Both halves were equally essential because automating the wrong thing at scale is more expensive than not automating at all.
Key Results
Reduction in time spend on tickets
Of all tickets touches by AI
Reduction in AI costs for a resolved ticket
