Most enterprise AI projects don't fail because of bad technology. They fail because someone bought a solution before understanding the problem.

Here's what we see repeatedly: a company reads about GPT-5 or some new model, gets excited, hires a consulting firm to "implement AI," and six months later has a nice demo that nobody actually uses. The pilot runs great on clean test data. Then it hits real-world invoices with coffee stains, handwritten notes in the margins, and three different date formats — and it falls apart.

The gap between demo and production is where most projects die.

We learned this the hard way on our early projects. Now we start every engagement the same way: we sit with the people who actually do the work. Not the executives who approved the budget — the person who spends Tuesday afternoons manually copying data between two systems. They know where the real problems are.

What actually matters

1. Start with the bottleneck, not the technology

A manufacturing client came to us wanting "an AI chatbot." After spending a day on their factory floor, we realized their real problem was that quality inspectors were spending 40% of their time writing reports. We built a voice-to-report system instead. No chatbot. Way more impact.

2. Get the data pipeline right first

The sexiest part of an AI project is the model. The most important part is how data gets in and out. We've seen projects with state-of-the-art models that are completely useless because they can't connect to the company's actual data sources. We spend more time on integration than on model selection — and that's by design.

3. Build for the 15%, not the 85%

Any decent model can handle the easy cases. The value is in how you handle the exceptions. Our document processing systems don't just process clean invoices — they flag the messy ones, route them to the right person, and learn from the corrections. That's the difference between a demo and a production system.

4. Measure in euros, not accuracy percentages

We had a client whose AI model had 97% accuracy. Sounds great, right? But that 3% error rate on 10,000 daily transactions meant 300 errors per day that humans had to fix. Each fix took 15 minutes. That's 75 hours of daily cleanup — more work than before the AI. We restructured the system to route uncertain cases differently, and the effective cost dropped by 60%.

The uncomfortable truth

Not every process needs AI. Sometimes a simple script, a better spreadsheet, or fixing the upstream data quality problem is the right answer. We tell clients this, even though it means smaller projects for us. It builds trust, and those clients always come back when they have a real AI-sized problem.

If you're considering an AI project, start by asking: "What would a 10x improvement in this process actually look like in euros per month?" If you can't answer that concretely, you're not ready yet. And that's fine.