88% of AI pilots never reach production in Spain

Spain leads Europe in generative AI adoption, with 60.5% of companies already folding it into their strategy, yet only 17.3% report a high return, and as many as 88% of proof-of-concept projects never make it to production, according to the second edition of SAS and IDC’s “Data and AI Impact Report” (SAS/IDC, reported by El Mundo El Salvador). At Evicron, an AI and custom software studio based in Barcelona founded in 2019 with over 200 projects delivered across 12 sectors, we run into the same gap in real conversations: companies that have already tried AI, have a pilot that works fine in a demo, and can’t get it to become a system their whole team actually uses. The same week this data surfaced, Anthropic announced a $100 million investment to train 10,000 engineers capable of solving exactly that problem. That’s no coincidence: the bottleneck in enterprise AI is no longer the model, it’s everything that sits between the model and the day-to-day of the business.

The numbers worrying executives

The SAS/IDC report, built on surveys of AI and data leaders across several countries, confirms Spain doesn’t have an adoption problem: most companies already use or are piloting generative AI. The problem sits in the next phase. Out of every ten pilots that launch, fewer than two end up generating a measurable return, and almost nine out of ten never even turn into a production system. The causes the report points to aren’t the model chosen or its cost: they’re a lack of adequate infrastructure, no real redesign of work processes, and data governance fragmented across departments.

Where an AI project actually gets stuck

In the projects we review at Evicron, this stall happens at three very specific points:

• The pilot is built on hand-curated, clean data, and breaks the moment it connects to real production data, scattered across several systems with inconsistent formats.
• Nobody redesigned the workflow around the AI: a chatbot or agent got bolted onto a process still built for a person to run, and the result is one more layer, not an improvement.
• No one owns the system once the demo is over: an external team or a one-off internal test built the pilot, and when an edge case needs fixing or volume needs to scale, nobody has the knowledge or the mandate to do it.

None of these three is a technical AI problem. They’re implementation problems, which is exactly why a better model doesn’t fix them.

Anthropic’s bet: paying for the talent that’s missing

On October 2, Anthropic announced the Claude Frontier Academy, a $100 million program to train 10,000 “Frontier Deployed Engineers” by the end of 2027 (Anthropic). The first cohorts come from consultancies and large enterprises such as Accenture, Bain, Capgemini, Deloitte, McKinsey and Morgan Stanley, who send engineers through intensive training followed by twelve weeks leading a real Claude use case inside their own organization. Anthropic’s own rationale is explicit: “the people with the skills to make it work inside a real business have become the hardest talent to find.” In other words: the company building some of the world’s best models is investing in training people, not in a better model, to fix the exact same problem the SAS/IDC report describes.

What this means for a Spanish SME

An SME isn’t sending anyone to the Claude Frontier Academy, but the diagnosis still applies. Buying access to a good AI model — Claude, GPT or Gemini — is the easy part, and any vendor already has it covered. What decides whether a project reaches production is who designs the integration with your systems, who redesigns the workflow around the AI, and who stays on to maintain the system after the first demo. That’s exactly what good AI consulting for business delivers: not selling access to a model, but carrying one concrete use case from pilot to full team adoption.

What makes an AI project actually reach production

Based on what we keep seeing in our own applied AI projects, the ones that scale share three traits the ones stuck in pilot don’t:

1. They’re validated against real data from week one, not against a curated sample that hides the edge cases production will throw at them.
2. The workflow gets redesigned, not decorated: AI changes who does which step, rather than bolting a new button onto a process that stays manual underneath.
3. There’s a clear internal owner from day one, with access to the code and the data, who isn’t permanently dependent on whoever built it.

In our own AI consulting engagements, that’s exactly where we start: before writing a line of code, we check whether the use case has the data, the process and the internal owner it needs to survive past the pilot stage. If any of the three is missing, we say so before we start, not after we bill.

The bottom line

Spain is adopting AI faster than the rest of Europe, but the SAS/IDC report confirms exactly what Anthropic just put $100 million on the table to fix: the question isn’t which model you use, it’s who carries that model into your actual operation and keeps it running. Before launching another pilot, it’s worth asking whether you’ve already solved the data, the process and the ownership that will decide if it reaches production or becomes one more demo.

Have an AI pilot that won’t quite get off the ground at your company? Get in touch: the first consultation is free, and we reply within 24 hours.