AI Insights

How Fiege's claims AI agent cuts 70% of admin effort

Chris Jon Graf
Chris Jon Graf · AI Strategist & CEOPublished on 16 November 2026
How Fiege's claims AI agent cuts 70% of admin effort

In short

Yes, complaints can be automated with AI agents. Fiege proves it in production: the agent handles over 70% of carrier claims, up to 11,000 cases a month. The real lessons: set the financial threshold first, plan for the last 30%, and build human-in-the-loop in from the start.

Can an AI agent really handle complaints on its own? Yes, under one condition: the process must be tightly defined. European logistics provider Fiege, one of the largest in Europe, has been running Carrier Claims AI in production for over a year. The agent handles more than 70% of incoming damage, loss and delay notifications, up to 11,000 cases a month. This is not a pilot; it is day-to-day operations.

Carrier claims are notifications exchanged between shippers, freight forwarders and recipients around transport issues. They arrive as emails with scans, photos and PDF attachments, come from 20 to 50 European parcel carriers and are written in several languages. These unstructured inputs are exactly where an AI agent creates value. Fiege describes the setup in its official Carrier Claims Management section.

70%

effort reduction achieved by Fiege's AI agent

up to 11,000

cases per month

Why hiring more staff was not the answer

Before automation, up to 14 employees spent most of their time on claims processing. Each case took 15 to 25 minutes. The problem: for many reimbursements, the payout does not cover the personnel cost. B2B liability in Italy, for example, can mean around 1 euro per kilogram. A lost designer shirt worth several hundred euros costs more to process manually than the reimbursement will ever bring in. Adding staff would only have made the problem bigger.

Lesson 1: The financial threshold comes before the first line of code

Fiege did not run the business case after the project started; it defined it beforehand. The requirement was clear: at least 70% effort reduction, otherwise the project does not pay off. This threshold shaped every design decision. If you want to follow the same path, first check which processes really pay off. A detailed view of the Fiege case is available in the t3n use case report.

Set the threshold first

Measure the current cost per case, compare personnel costs with expected reimbursements and define a measurable threshold. Only then start the technical implementation.

Lesson 2: The last 30% are the real work

After around nine months of development, the automation rate was above 70%, sometimes exceeded. Fiege reports a resolution rate of 98%. But the journey was uneven: the last 30% of cases, mostly multilingual edge cases from carriers, required the most fine-tuning. When you plan an AI agent, reserve budget for edge cases, not for the basic pattern.

The decisive moment came when the most experienced claims clerk switched sides and translated tacit knowledge into explicit rules, based on around 2,000 historical cases. This is the difference between a generic chat tool and a real AI agent: the process knowledge lives in the system, not just in the model.

Lesson 3: Human-in-the-loop is not a fallback

The agent only answers a case on its own when it is confident enough. In every other case, it hands the claim to a human. That is not a safety net added later; it is a fundamental design decision. This decision keeps the error rate low and builds trust among employees.

What the Fiege case means for your business

The case is solid evidence that AI agents in the back office are not a toy, but a measurable business model. The production numbers are consistent with other European deployments. If you want to understand what this means for your company, the Swiss AI podcast on AI as a management topic is a good starting point.

Cases per month
up to 11,000
Parcel carriers
20 to 50 European carriers
Effort before
up to 14 employees, 15 to 25 minutes per case
Automation rate
over 70%, sometimes exceeded
Resolution rate
98%
Employees reassigned
10 moved to other roles

The technology: not a chatbot, but an orchestrated process

Carrier Claims AI is built on Azure Functions and Services, plus a GDPR-compliant OpenAI model operated in a sandbox with templated responses. The architecture matters more than the model: this is not a single model that magically handles everything, but a clearly scoped business process with rules, interfaces and handovers to humans.

For Swiss companies, the data question is central. Fiege's GDPR-compliant sandbox translates directly into revDSG requirements: data transfers to third countries need contractual and technical safeguards. That is not a blocker, but it belongs on the agenda from day one.

Your first step

Pick one recurring process, measure the current effort and define the financial threshold. The question is not whether an AI agent can do the job; the question is whether it pays off for your operation. If you want to explore how to make AI a management topic rather than an IT project, the Swiss AI podcast offers practical guidance.

Frequently asked questions

What is a carrier claim?
Carrier claims are damage, loss and delay notifications exchanged between shippers, freight forwarders and recipients. At Fiege, they arrive as emails with scans, photos and PDF attachments, in several languages and from 20 to 50 parcel carriers.
How much effort did Fiege save with the AI agent?
More than 70% of processing. The agent handles up to 11,000 cases a month, with a resolution rate of 98%. Ten employees moved to other tasks.
Why do the last 30% of cases cost the most?
They are multilingual edge cases from carriers that cannot be covered by standard rules. Fiege translated the most experienced clerk's tacit knowledge into explicit rules, based on around 2,000 historical cases.
Do you need a data science team to build such an agent?
No. Fiege shows that a clearly scoped process with a GDPR-compliant model is enough. The work lies in process definition and in translating domain knowledge, not in model training.
Does human-in-the-loop contradict automation?
No. The agent only solves a case when it is confident. Everything else goes to a human. This keeps the error rate low and responsibilities clear.

Sources

Chris Jon Graf
About the Author & Your Direct Contact

Chris Jon Graf · AI Strategist & CEO

For over 15 years, Chris Jon Graf has built productive Swiss software platforms and advised executive teams on the strategic deployment of artificial intelligence. Would you like to evaluate this topic for your company?

Check AvailabilityWould you like to explore this topic for your company?

More articles