Reaching for some AI tool is now within nearly every mid-sized company's grasp. According to a survey of US companies with 200 to 5,000 employees, 82 percent have artificial intelligence in production somewhere. But only 26 percent have it scaled and governed across the whole company. So most companies use AI in some way — they just get almost nothing out of it yet.

And this is exactly where the most interesting rift in the whole field opens up. There's plenty of experimentation, precious few tangible results.

The problem isn't the model, it's the connection to processes

This was best described by an extensive MIT study called The GenAI Divide. Its authors reviewed some three hundred publicly documented deployments and found that despite tens of billions of dollars invested, 95 percent of generative AI projects delivered no measurable impact on the bottom line. Only 5 percent of companies extracted real value.

You'd expect immature technology to be the culprit. But the report argues the opposite. The models are good enough. What fails is the connection to real company processes. A general tool like ChatGPT serves an individual brilliantly, but in company operations it doesn't learn from its own work and doesn't adapt to how things actually run. The authors call this the learning gap.

It's worth adding one inconspicuous point. A large part of that 95 percent didn't fail because it technically didn't work, but because companies never measured the baseline before deployment. Without it, they then have nothing to prove anything improved. That alone is a lesson.

The mid-sized company has it worst

What's more, the rift doesn't hit everyone equally. Adoption is driven mainly by large corporations. According to Eurostat, 20 percent of all EU companies with ten or more employees use AI, but among large firms it's 55 percent. Czechia is slightly above the European average in this respect. Smaller and mid-sized players lag behind, and once they do deploy a tool, they find it that much harder to carry it through into real operation.

The mid-sized company is meanwhile in the most thankless position of all. It has no in-house AI team or budget for governance and security like a big player. At the same time, it can no longer make do with someone occasionally opening ChatGPT, the way a sole trader can. It's too big to improvise and too small for its own department. That's precisely why it matters where it directs its limited energy.

First the problem, then AI

The first and most common mistake is the order. AI is deployed today as the first choice, because it's visible and much talked about, not because it's the best fit for the task. The order should be reversed. First the problem, then the question of whether AI will solve it best, or whether a simple rule, a function in a system the company already has, or plain automation would do. Those tend to be cheaper, more reliable and easier to check. AI earns its place where the task is so variable and unstructured that fixed rules can't cover it.

The second filter is the size of the benefit. Putting AI on reading emails looks like progress, but it won't move any important number. Interestingly, successful companies aim elsewhere than you'd expect. According to MIT, the highest returns don't come from flashy things in marketing and sales, but from dull back-office processes — document processing, procurement or risk review. Where a large volume of work repeats and the result can be measured, AI pays off most. It pays off least where it's most visible.

A membership isn't a result yet

The whole thing is best explained by the gym. Buying a membership isn't the same as losing weight. The result doesn't come from access itself, but from a regular routine, measuring from a starting weight, and focusing on a few exercises that actually work rather than ten random ones. Deploying AI works the same way. Having it isn't enough. It has to be plugged into everyday operation, you have to be able to tell whether it's working, and someone has to be accountable for it.

Responsibility is what's most underrated. Without a clear owner, a solution belongs to no one and fades away over time. And one more parallel fits. People go to the gym without a trainer too, on their own and without a plan. It's the same in a company. Even where the official tool fails, people still reach for their own AI, just without any oversight at all. That's a risk, but also a clue. It shows where the benefit could be captured in a managed way.

Conclusion

The question of whether to use AI is already settled. Almost everyone uses it. The interesting question is whether you actually get any benefit from it. And that isn't decided by the smartest model, but by choosing the right place, integrating it into the process, and having someone who's accountable. Whoever manages this gains an edge that's hard to copy, because it doesn't rest on a tool the competition also has.

My take? The cheapest AI strategy for a mid-sized company isn't buying another tool. It's picking one process with a really big benefit. Getting measurement and ownership right there. Only then thinking about the next one. And above all, don't be reassured by the line that you already use AI. On its own it means nothing at all.