95% of AI pilots deliver nothing. The mid-market can do it differently.

Companies together poured 30 to 40 billion dollars into generative AI. For 95% of the pilots, that produced no measurable return, according to MIT in the report The GenAI Divide (2025). Five percent capture real value. The rest got stuck in a demo that briefly convinced the board and then slowly withered away.

Still, that is no reason to leave AI aside. The problem lies in how most companies start with it. A large project, an outside party that builds it, an expensive tool from a big vendor, and the promise that it will pay for itself later. The technology underneath is rarely the weak link. The setup around it is.

The graveyard is full of proofs of concept

Gartner predicted in July 2024 that at least 30% of generative AI projects would be abandoned after the proof of concept by the end of 2025. The causes cited: poor data quality, too little risk control, rising costs and unclear business value. S&P Global Market Intelligence saw this climb sharply in 2025. 42% of companies scrapped most of their AI initiatives, up from 17% a year earlier. On average, 46% of proofs of concept died before they reached production.

The obstacles that come up most often here are cost, data privacy and security. Anyone who lived through the previous wave of large software projects recognises the mechanics immediately. One big block, one promise, and no intermediate step where anyone can see whether it works. That is why we build in waves with a go/no-go at every step. This lets you stop the moment a pilot delivers nothing, instead of only when the budget is already gone.

The mid-market has an advantage it rarely uses

That same MIT study contains a figure the headlines rarely reach. Mid-sized companies go from pilot to production in about 90 days. Large enterprises take nine months or more. That difference is not down to better technology. It comes from the distance between the person who decides and the person who does the work being smaller.

In a company of two hundred people, the person who knows the problem sits at the same table as the person who frees up budget. There is no steering committee that spends three quarters debating a single use case. That is a real head start, and most mid-sized companies squander it by copying an approach designed for a large corporation.

What the 5% do differently: small, and under their own control

AI is only as good as the data underneath it. If that data is spread across a hundred SaaS contracts you do not control, you build your model on sand, and you pay an outside party to clean it up afterwards. The 5% that do get something out of it start on the other side. First find out where the data and the spend leak away, something a vendor-neutral audit maps out, and then bring the data together in one platform under your own control with the code in your own repository. AI comes after that, at the point where it solves a concrete problem.

Open source here is above all a practical choice. In the study The State of Enterprise Open Source by Red Hat (2022), companies name better security and higher software quality as the most cited advantages. For AI, something is added to that. The models and tools that matter are largely open. You run them on your own infrastructure, on your own data, without sending that data to an outside model vendor. That is at once your answer to the data privacy and security that were the biggest stumbling blocks in the figures above.

• Your data in one data model that you own, not smeared across a hundred separate contracts.
• An open-source stack that you run yourself, so your data stays within your own environment.
• AI added in one place where it solves a measurable problem, not as a layer over everything at once.
• A go/no-go after every step, so a pilot that delivers nothing fails early instead of after a year.

A stack that stays small can be maintained by a team of 1 to 2 people. That sounds like a detail, but it is the heart of the cost story. Where a large corporation hires three firms that each build a piece and each send an invoice, a mid-sized company keeps things running with a handful of people who genuinely know the system. This is how we work ourselves too, and it is the reason the bill moves with the infrastructure you run, not with the number of users.

AI does not belong at the front of the brochure, but at the back of the build order. You put it to work on a task you can measure, where it makes something faster or cheaper, and only once the data underneath is right. Such an addition makes no impression in a steering committee. It does keep working long after the attention has moved on.

The 30 to 40 billion dollars was largely spent on the question of whether AI works. For your company, that is the wrong question. The right one is narrower. Which task, on which data, with which number behind it, and who maintains it a year from now. Whoever starts small on a stack they own knows within 90 days whether they are among the 5% that keep something, or pulls the plug in time. Both are better than nine months waiting for a demo no one opens afterwards.