Starting out of Milan with a €3.2 million check to rethink one of the least digitized industrial processes: that is the bet of Compri, a 2024 startup founded by Edoardo Arbizzi and Edoardo Gava, which has just closed a seed round led by Picus Capital with participation from Shapers, Italian Founders Fund, DFF Ventures, and private investors. The deal pushes total funding above €5 million and shines a light on a sector – industrial procurement – still trapped between emails, spreadsheets, and disconnected ERP systems.

The promise: AI agents working alongside procurement teams

Compri’s platform aims to act as a digital workforce: it collects data from ERPs, inboxes, PDFs, and external databases, normalizes it, and feeds it into large language models (LLMs) trained on procurement-specific domains. The output is software agents capable of chasing suppliers, gathering documentation, monitoring compliance, and verifying order confirmations, freeing human teams to focus on higher-value activities like strategic negotiations.

From an architectural standpoint, combining LLMs with a layer of proprietary, context-rich training is no trivial choice. Without public details on the specific model, it is reasonable to assume the use of fine-tuning or retrieval-augmented generation (RAG) to ground responses in actual documentation, minimizing hallucinations. This approach is especially critical in procurement, where a misinterpretation of an order or a certificate can lead to severe operational and legal fallout.

The real watershed: where does the intelligence reside?

The funding news, however, raises a question that goes beyond the product: AI deployment. Compri has not declared whether the platform runs on public cloud, private cloud, or a hybrid setup. Yet, for the targeted customers – European manufacturing companies, often reluctant to move supplier and pricing data outside their perimeters – the physical location of the model is far from a detail. Cloud-only services can clash with corporate data residency policies, GDPR requirements, and the need to integrate legacy systems that cannot talk to the outside world.

This is where the debate we follow at AI-RADAR comes in: for LLM workloads in sensitive contexts, self-hosted on-premise is not a technical whim but often the only viable path. Running inference on local GPUs, using quantization-optimized models, orchestrating containers on Kubernetes: solutions that allow total data control, avoid unpredictable recurring costs, and meet audit requirements. They do entail higher upfront investments and the need for specialized in-house skills, but for a growing number of industrial players, TCO must be assessed over a much longer horizon than a simple cloud bill.

Beyond the startup: what this investment signals

Compri’s round is a strong signal for the Italian and European ecosystem. On one hand, it confirms that AI applied to industrial processes – not just generic chatbots – can attract confident capital. On the other, it highlights the tension between the promise of a ready-to-use platform and the deployment demands that the manufacturing market imposes. Anyone choosing an AI procurement tool today must look past the demo: they need to understand which deployment model is on offer, whether on-premise or hybrid options exist, and how ownership of the data used for fine-tuning is handled.

Looking ahead, it is plausible that startups like Compri will soon have to articulate differentiated infrastructure roadmaps. The demand for “air-gapped” versions, installable inside the factory and disconnected from the network, may well become a decisive competitive factor. For those evaluating such solutions, transparency on the technology stack and architectural flexibility will count at least as much as application features. The procurement of the future will be agent-centric, yes, but probably also far more rooted in the server room than early seed rounds suggest.