AI projects in mid-sized companies rarely fail at the model. They stall at integration with existing systems and at adoption by the people who use them, two jobs no licence does. One engineer embedded part of the week, owning one use case, fits a firm of 50 to 1,000 people better than more software, a consultancy or a new hire.
Before AI, I worked in industrial automation, and nobody in that trade expects a control system to run a plant out of the box. The cabinet arrives in a crate. Someone has to come on site, wire it to machines installed twenty years apart, discover that a sensor reports in different units from the drawings, and stay until the shift trusts it. The trade calls this commissioning, and it is part of the purchase, not an afterthought.
Enterprise AI is sold as if it needed no commissioning. The evidence suggests it needs more.
Why do AI projects stall after the software is bought?
Because buying is the only step the software can do by itself. IBM asked 2,000 chief executives in 2025 how their AI initiatives had gone: only 25% had delivered the expected return, and only 16% had scaled across the company (IBM Institute for Business Value). A year later, in IBM's 2026 study, only 25% of the workforce was using AI regularly in their jobs (IBM).
BCG's 2025 survey of 1,250 executives shows where it sticks. 72% named integrating AI with their systems, tools and APIs as a challenge, and 77% named getting people to adapt and use AI every day (BCG). BCG sells the remedy, so read the figures as its clients' view. But integration and daily use are precisely the two places where a licence stops working for you.
Neither is a model problem. A model that answers well in a demo still has to read from the ERP, write to the CRM, respect who may see what, and fit into the day of someone whose day already has a shape.
Why are mid-sized companies further behind on AI?
In 2025, 20.0% of EU enterprises with ten or more employees used at least one AI technology. Romania was last, at 5.2% (Eurostat). The gap is widest in the middle: 30.4% of EU firms with 50 to 249 employees used AI, against 55.0% of large firms, and in Romania only 7.8% of medium-sized firms did (Eurostat, AI key results 2026). The breakdown by size and sector is in AI adoption in European SMEs.
Eurostat also asked the firms that considered AI and did not go ahead why. The most common answer, at 70.3%, was a lack of relevant expertise, ahead of unclear legal consequences (53.6%) and data-protection concerns (52.7%); among firms of 50 to 249 people it was 69.2% (Eurostat). Its own explanation of the size gap names the missing piece: large companies are better able to cope with the complexity of introducing AI and to "provide the necessary expertise".
Large firms also buy that expertise in. In 2024, 45.8% of large EU enterprises using AI used systems built or adapted by external providers, against 29.7% of medium-sized ones (Eurostat). The middle tries to do it alone, and mostly does not start. The blocker, in other words, is a person, not a product.
Why not more software, a large consultancy or a new hire?
More software
Another licence is another system to integrate. If the first tool is not in daily use, the second meets the same wall, and a company where a quarter of the staff use AI regularly does not get to half because the menu grew longer.
A large consultancy
A consultancy is built to produce a plan and hand it over, with teams sized for larger budgets than a mid-sized firm has. For that firm the plan is rarely the scarce part. What Eurostat's respondents said they lacked was expertise: someone who can make the system work on the company's own data, after the plan is written.
A new hire
Hiring one AI engineer feels like the permanent answer, but a single hire lands in a company that is not set up to support them. As I wrote in how to structure your company for AI, a lone AI engineer with no prepared data risks having nothing to build. Training points the same way: in 2024, 73% of large EU enterprises trained their staff in ICT skills, against 21% of small and medium ones (Eurostat). For that half of the problem there are AI workshops for teams.
What does one embedded engineer change in a mid-sized company?
The work the licence does not do gets an owner. One senior engineer sits in the company, on its systems, about three days a week, and owns one use case until it runs in production and people use it. It is the forward deployed engineer at the scale of a mid-sized firm. Four things change:
- Integration has an owner. The engineer writes the connections to the ERP, the document store or the CRM, on real data and with real permissions, not on an export.
- Success is written down first. A baseline and a target for one use case, agreed with a named sponsor before the build, so the pilot is judged on a number rather than an impression.
- Adoption is part of the job. The engineer is in the room when the people who use the system find what it gets wrong, and fixes it.
- The capability stays. One of the company's own people pairs with the engineer from the first week and keeps the code, the documentation and the evaluation set at handover.
It also changes what can honestly be promised. I do not quote savings before a pilot: the only return worth trusting is the one measured in your own pilot, against a baseline taken before it began. My first AI project was paid only if it worked, and I still measure our work by that test. How the companies that made the model famous run it is in the forward deployed engineering model.
At Sapio we run this as a sequence, and each step can be the last. It starts with the contact form on sapio.ro and a short call; then I send a custom offer, sized to the project: AI consulting, or the Tech Audit, one morning on site and a report in three working days. Next comes a transformation roadmap with a baseline and a target per use case, and then a senior engineer embedded about three days a week, owning one use case until it runs in production.
In a plant, nobody calls the commissioning engineer a luxury, because without one the machines visibly do not run. AI is harder to see. So the commissioning gets skipped, and the system sits in its crate with a login.
If you are deciding where AI should start in your company: vladtudor.com/consulting.