Choose a first AI project worth shipping

For a mid-market team, the useful question is specific: which workflow is worth improving first, and what would justify putting it into production?

Companies doing €100M to €500M in revenue often have valuable workflows and limited implementation capacity. For a team working in the EU, privacy, security and data-handling requirements need to be agreed before choosing the implementation.

That's who this blog is for.

The Companies Making the Leap

Across the EU—manufacturing, logistics, professional services, financial services—mid-market companies share a common starting point:

• Cost centers eating margin. Customer support backlogs. Manual data entry. Document processing that takes humans days when it could take AI minutes.
• Failed AI experiments. You tried ChatGPT. Maybe a vendor demo. Nothing stuck. The ROI case died somewhere between IT approval and procurement.
• Competitor pressure. You've seen what the big players are doing with AI. The question isn't whether to automate—it's how to start without betting the company.

If that sounds familiar, you're in the right place.

What's Actually Working

These are practical questions to resolve before committing to an implementation:

• AI strategy without the fluff. Where to start. What to automate first. How to run an AI audit that produces a real roadmap, not a 200-page PDF.
• LLMs and RAG for business. When GPT-4 is enough. When you need fine-tuning. How to build retrieval-augmented generation systems that work with your actual data.
• Process automation that ships. Document processing, email triage and invoice handling. Establish the current handling time and exception rate before estimating a saving.
• AI agents and agentic workflows. What's hype. What's real. When autonomous AI systems make sense for operations.
• A bounded first project. Scope, ownership, integration and acceptance checks. Use 90 days as a planning horizon, with timing confirmed after diagnosis.

Where to Look

The practitioners shaping how AI actually gets deployed—not the hype cycles:

• Andrew Ng on AI strategy and implementation at scale
• Ethan Mollick on practical AI adoption in organizations
• Allie Miller on enterprise AI and startup integration
• Andrej Karpathy on what's technically possible and what isn't
• Pascal Bornet on intelligent automation and hyperautomation

Use their work to frame questions, then test those questions against your own process, data and budget.

The Pattern

A useful starting rule: diagnose before you prescribe.

Identify one process, agree the baseline and acceptance checks, and decide whether automation is justified. A simpler process change or a hire may be the better answer. The first project should be small enough to test before committing to a wider programme.

The aim is a working improvement that the process owner can operate and assess. Expand only when the evidence supports it.