The customer is a Belgian engineering consultancy with around a hundred employees, active in logistics, energy and construction technology. The company works for public and private clients on projects involving many parties. At the start of the project, it used Microsoft 365, without a defined AI strategy or its own AI tooling.

A project that was struggling provided the direct trigger. Communication with the customer thinned out, timing became unclear, and the escalation only came into view once it was already a fact. In project work involving many parties, the early warning signs usually sit in the communication itself, scattered across hundreds of emails and messages that nobody sees in full.

The company wanted to pick up on those signals earlier, without adding an extra layer of reporting on top.

The project started broadly. Through an AI agent, all employees were asked where their work ran into friction, which tasks ate up time, and which patterns kept recurring. That produced more than twelve candidate use cases.

In a half-day workshop, the team chose the priority together with management, weighed on impact and feasibility. Project escalation came out on top, given its direct link to customer satisfaction and to losses from earlier projects.

The build ran over ten weeks, with success criteria fixed in advance. One of them: at least seventy percent of structurally negative sentiment trends must lead to an intervention within a week.

Xylos built a first working prototype in two weeks. An AI agent reads along with incoming and outgoing project communication, analyzes the sentiment and alerts the project manager as soon as the trend turns structurally negative. From there, the agent grew iteratively with three functions.

If the sentiment stays negative, the agent automatically escalates to management. A weekly summary shows sentiment trends per customer, giving leadership proactive visibility into projects at risk of derailing. A third function supports the communication itself, suggesting how a message can best be tailored to its recipient.

Under the hood runs a deliberate mix of generative AI and classic machine learning. Azure OpenAI handles the sentiment analysis, a fine-tuned open source language model the style assessment. The agent connects with Microsoft Teams, the project dashboards and time tracking.

The pilot ran for six weeks on more than 2,500 emails and 830 internal messages. Three escalation peaks turned out, in hindsight, to have been visible in the data three to five weeks earlier than with the customer itself. The agent has been running in production since late 2024. An evaluation six months later shows the following.

• 8 out of 11 structurally negative trends led to an intervention within a week, above the targeted criterion of 70 percent.

• 3 escalations avoided that, according to management, would otherwise have landed with the customer.

• 78 percent of project managers use the weekly summary, 7 out of 9 also use the communication support.

• The adoption test came back unanimously positive: the agent would be missed if it disappeared.

The approach is now growing into resource planning and customer relationship management. Internally, an AI coordinator has been added.

Want to know where AI would deliver the most value for you? We'll figure it out together and build a working prototype within a few weeks.