Introduction
How a fast-growing services company automated its fleet administration and built a reusable foundation for internal AI agents
Like any growing organization, the' services company had to scale its internal operations alongside the business. Much of that growth showed up as repetitive, high-volume work, and a disproportionate share of it landed on the Office Management team.
Client
Services company
Client since
Services
Solutions
Technologies
The problem
As the company grew, the workload of its office managers started to outpace the size of the team. With many new employees joining, the same operational processes had to be repeated again and again. These processes often involved manual coordination, follow-up, and data entry across different systems, creating a growing administrative burden for the team.
One of those processes was fleet administration. Every new employee who needed a company car triggered a repetitive workflow: collecting the required information, such as their driver’s license and preferred car type, submitting the car request, checking the available options, coordinating the next steps with the relevant parties, and making sure each step was completed correctly.
That made fleet administration a strong first candidate to test a bigger question: "Could a custom AI agent take on a recurring internal process, reducing manual workload and supporting the office team, while keeping human oversight where it matters? "
How we solved it
Understanding the process first
We started by interviewing the office management team to understand the existing fleet administration process in detail. This let us map the full workflow, identify the repetitive steps, pinpoint where human judgment was still needed, and surface potential risks.
Designing for safety before building
Before writing any code, we ran a risk assessment. This defined where guardrails were needed, which steps required human validation, and which parts of the process could be automated responsibly.
Building in-house on Python and Azure
We tested several ways to build and host the agent. After running into limitations with Power Automate, we built the solution fully in Python and hosted it on Azure. The Fleet Agent uses custom connectors for Monday.com, Outlook, and other tools, so it works directly with the systems the office team already relies on.
Rolling out in parallel, not all at once
Rather than replacing the manual process from day one, we introduced the agent alongside it. The office team could compare the agent's output against their existing way of working, validate it, and build confidence before moving toward a higher level of automation.
The results
The Fleet Agent delivered a working proof of concept showing that custom AI agents can support real internal business processes, beyond simple chatbot use cases. It demonstrated that parts of the fleet administration workflow can be interpreted, structured, and automated by combining process knowledge with integrations into the tools already in use.
The project revealed something about making agentic automation reliable in practice. Running the agent in parallel with the office team validated its output against the manual process, confirmed where human approval remains essential, and refined the guardrails needed before wider adoption.
The build also created a practical foundation for future internal agents. Working in Python on Azure delivered hands-on experience with custom connectors, controlled execution, monitoring, and human-in-the-loop workflows. The Fleet Agent isn't just a fix for one process; it's a first step toward automating similar repetitive workflows across the organization.
3 days
to a first working proof of concept
1 modular framework
reused across all future agents
Shared resources
across agents to keep running costs down
Key Learnings
Start small, keep humans in the loop, and build for reuse.
Start narrow.
Pick a well-defined workflow where the expected steps and decision points are clear. A tightly scoped first use case is far easier to automate reliably than a sprawling one.
Keep a human in the loop.
Human validation is essential before you automate an operational process end to end, especially one that touches employee data.
Build for reuse.
Reusable agent infrastructure matters. Getting the framework right once means future agents can be built faster and more consistently.
More case studies
We deliver impact where it matters most.
Unlocking Viewer Insights with a Scalable Data Platform
Streamz is the Flemish video-on-demand platform offering local and international series and films through a subscription-based model. Streamz focuses on delivering high-quality content tailored to local audiences, supported by a strategic partnership with Paramount.
Predicting Platelet Demand: Transforming Healthcare Logistics with Time Series Forecasting
Improving platelet demand planning with time series forecasting to support reliable healthcare logistics.
Building Practical AI Governance at KBC with Collibra
KBC, one of Belgium’s leading financial institutions, has a long-standing focus on responsible innovation and efficient risk management. With the European AI Act introducing new expectations around transparency, risk management, and accountability, the organization saw an opportunity to strengthen and streamline its approach to AI governance. For KBC, this meant bringing different teams’ perspectives on AI together into one coherent framework—without disrupting existing processes.
