How AI Agents Cut Manual Service Desk Workload by up to 25% in e-Invoicing

The e-invoicing service desk handles a high volume of tickets each year, spanning hundreds of maintained projects and several integrated product lines. At Comarch E-Invoicing, it’s around 110,000 tickets a year. That scale, combined with a wide variety of business areas and clients, makes manual handling slow, inconsistent, and hard to scale without adding headcount. Rather than looking for a single generic automation shortcut, we decided to break down the entire scope of service desk work into categories, map each one to its existing handling procedure, and build a two-phase agent model on top of that structure – an approach that has reduced overall handling time overhead by about 20–25%.

Read on to learn:

• Why a single, generic automation approach does not work for the e-invoicing service desk.
• The four-step method for turning 150+ ticket categories into a realistic automation roadmap.
• How a two-phase AI agent model routes and resolves tickets without sacrificing accuracy.

What Makes E-Invoicing Service Desk Different

What makes first-line support in e-invoicing specific isn't just volume – it's variety. Service desk covers several integrated product lines, including e-invoicing, EDI, ECM, OD, and SFA, supporting numerous maintenance projects. That breadth of business areas and clients translates directly into a high degree of project fragmentation, which in turn drives a large number of highly specific handling procedures. A ticket about a document transmission error follows a completely different path than one about a partner onboarding request or an AS2 certificate change.

That diversity is the core problem for automation. Service desk receives more than 150 distinct ticket categories every year. Manually triaging all of them is slow and inconsistent by nature, and a misrouted ticket can lose hours before it reaches the specialist who actually knows how to resolve it. Any automation effort has to start by making sense of that complexity – and that’s exactly what we did at Comarch E-Invoicing.

How to Approach Automation in E-Invoicing Service Desk

1. Choosing What to Automate

Turning that complexity into something actionable required a structured, four-step method:

• Identify the categories: Over 150 ticket categories were identified, each specific to a different business area.
• Build templates: High-volume categories – those generating more than 1,000 tickets a year – were split into ticket templates that capture their typical structure and content.
• Map to procedures: Each ticket type was sorted and mapped to its specific, existing handling procedure.
• Review automation potential: Each handling procedure was then assessed individually for its potential to be automated.

This kind of structured, volume-prioritized view made it possible to build a realistic automation roadmap – one that targets the categories that generate the most manual effort first, rather than automating whatever happens to be easiest.

2. The Automation Model: Two-Phase Handling

The automation itself follows a two-phase model, triggered by tickets arriving through either e-mail or Jira.

• Phase 1 – generic procedures: Every new ticket is first handled by an agent following standard, generic handling procedures that apply regardless of category.
• Categorization: Once phase 1 is complete, the ticket is classified and routed to the handling procedure that matches its specific category.
• Phase 2 – category-specific procedures: From there, a dedicated agent takes over, following the procedure built specifically for that ticket category.

Technically, the flow is straightforward. Tickets originate in Jira and are picked up by a receiver component that reads new tickets. Business rules validate and route each ticket, consulting a shared document-reader utility whenever XML content or certificates need to be parsed. An AI assistant then classifies the ticket and drafts an answer, drawing on a knowledge stack for pattern matching and on relevant customer and EDI data sources to provide the specific facts a ticket requires. Some AI agents write the final answer back into the original Jira ticket, keeping the whole interaction in one place for both the customer and the service desk team, while others handle the full process, resolving the ticket end-to-end.

How AI Agents Support E-Invoicing Service Desk

Multiple AI agents built on this model are already running in production, each targeting a specific, high-volume slice of service desk work. The six most important include:

• Ticket Categorizer tags every new ticket with the right category and subcategory based on ticket summary, description, and attached XML, so it reaches the correct specialist team from minute one.
• Document Status Verifier retrieves EDI document logs and statuses to diagnose validation or processing errors.
• Partner Data Retriever looks up a B2B partner’s technical setup – routing, channels, certificates – using a single GLN/ILN from the ticket.
• Fazer Error Handler classifies recurring document transmission incidents into known error paths and drafts the customer reply.
• AS2 Certificate Handler verifies and packages AS2 certificate change requests so they reach the second-line team ready to action, without back-and-forth over missing data.
• Similar Issues Finder attaches links to the most similar past tickets to every new ticket, giving service desk staff ready context and known resolutions in seconds.

How Do These AI Agents Work in Practice?

Each of these agents delivers significant time savings for everyday service desk processes. Here's what that looks like in practice, based on the example of the Document Status Verifier and the Similar Issues Finder.

Previously, service desk staff spent about three minutes per ticket manually looking up document traces, statuses, and validation errors in EDI Tracking before they could respond to a customer. The Document Status Verifier agent now finds the message ID or document number in the ticket, queries EDI Tracking for the full event log, and posts a ready-to-read result directly as a Jira comment.

On the other hand, Similar Issues Finder embeds the ticket content on every new ticket and queries a vector database for the four most similar past tickets above a defined similarity threshold, then posts that list as a Jira comment, so the engineer immediately sees how similar cases were resolved without having to look for them manually.

The Results of E-Invoicing Service Desk Automation

• Based on 2025 ticket volumes, the agents already in production are estimated to save close to 6,000 hours of manual work per year – roughly the equivalent of three full-time employees (FTE).
• An average handling time reduced by about 20–25%.
• An estimated 8.8% of tickets are fully automated end-to-end, with no manual intervention.

A Model Built to Scale

The biggest lever here wasn’t a single powerful automation tool – it was the discipline of mapping ticket categories to procedures before building a single agent. That structured, volume-driven approach turned an unwieldy mix of 150+ categories into a realistic, prioritized roadmap.

What makes this two-phase model powerful is that it's open-ended by design. We keep extending it with new agents – the number already running in production has grown to well over a dozen. Every new agent means lower costs, faster resolution, and better, more consistent handling for our customers – and there's still plenty of room to keep pushing all three further.

Wojciech Nowak

Head of AI LAB at Comarch