interfood
AI agents that tell you exactly what is urgent in the export process
The results
01
The foundation: the logistical Control Tower
To understand why we were able to build this AI solution in 40 hours, we need to take a step back. At Interfood, dairy shipments depart daily all over the world, involving a logistical puzzle of customs requirements and tight deadlines.
To manage this effectively, we previously built Control Tower in OutSystems (O11). This was a substantial project in which we built, among other things, the data model and a streamlined process architecture for the shipment lifecycle.
Control Tower brought all the work together in one place, from planning to customer communication. That formed the indispensable foundation. Without Control Tower, this AI agent simply could not exist.
02
The problem
Control Tower provides a fantastic overview, but Interfood ran into a limit regarding prioritization. The list of shipments was simply sorted by delivery date. That proved to be too simplistic. After all, a delivery date ignores the process phase of a shipment and does not tell you whether a coordinator can actually take action at that moment.
The logistics coordinator therefore had to manually scan the list daily. This required human judgment each time to determine which phase a shipment was in and whether anything could be done. This manual sorting work quickly took 15 minutes per shipment.
03
The AI solution: Rule-based prioritization
We abandoned static dates and built a rule-based prioritization model using the OutSystems Agent Workbench. An AI agent continuously monitors a dynamic mix of factors:
Outstanding actions: What exactly needs to be done?
Status & Process Step: What stage is the shipment in?
Date-based baselines & offsets: The agent looks at base dates (such as the sales date or the moment an email was sent to the customer) and applies a weighting (offset) per action type to calculate the actual urgency.
Does something require attention now? Then it appears at the top of the list immediately.
Smart design: Two agents in a 'chaining' setup
A complex challenge was that a shipment often has multiple outstanding actions simultaneously. For example: a certificate needs to be added and the correct labeling still needs to be arranged. Each of these actions has a completely unique urgency profile. Which one then determines the position on the list?
We solved this with a two-agent approach (Agent Chaining):
Agent 1: Calculates the exact priority score for each individual outstanding action.
Agent 2: Removes duplicates and retains only the most important, most urgent action per shipment. That is the action the coordinator sees.
The AI reorders, humans retain control
Crucial for adoption in the workplace: the agent makes no decisions and automates no operational tasks. Selecting carriers, closing tasks, or sending emails remains with the coordinator.
The agent continuously evaluates and reorders the worklist. It is not an autonomous decision-maker, but an intelligent layer that guarantees the human focus always remains on the right shipment.
Why we chose the Agent Workbench
The choice for the OutSystems Agent Workbench was key to the short development time of 40 hours. It enabled our team to write out the business logic (the prioritization rules) in natural language, instead of having to struggle with complex technical AI infrastructure. Moreover, it supported the chaining of two agents out-of-the-box, which was indispensable for this case.
It also makes the solution scalable and future-proof. If new actions are added to the export process in six months, we can easily expand the prioritization rules without having to rebuild the AI architecture.
04
Conclusion: AI as a logical extension of your core system
The AI agent runs in Interfood's daily operations and permanently puts an end to unnecessary manual searching. The most important lesson from this case? Implementing AI does not have to be a complex, standalone experiment. If your core system (like Control Tower) has a solid foundation and the processes are right, you can add an AI-driven layer in just 40 hours that immediately transforms the way of working.
