Why most AI pilots fail: the model is fine, the company context is missing

By now most companies have run at least one AI pilot. Most of them have little to show for it. The reports from 2025 and 2026 disagree on the exact number but agree on the shape: a clear majority of generative AI pilots never reach measurable value.

The reflex is to blame the model, the vendor or the users. We think the evidence points somewhere else. The AI was asked to do a job inside a company whose knowledge it had no reliable way to reach.

The numbers, with dates

Statistics about AI get repeated without their dates, so here are the ones we rely on and when they were published.

In 2025 MIT's NANDA initiative studied enterprise generative AI deployments and found that about 95 percent of pilots produced no measurable impact on profit and loss. The authors did not blame model quality. They pointed to a learning gap: most tools do not retain feedback, do not adapt to context and do not improve with use.

Gartner, also in 2025, estimated that 60 percent of AI projects unsupported by AI-ready data would be abandoned through 2026, and that only 37 percent of organisations were confident in their data practices. In 2026 Gartner went further and named the missing layer: context engineering, and a new category of AI context platforms, with context graphs expected to underpin more than half of agent systems by 2028.

On the adoption side, Microsoft 365 Copilot passed 30 million paid seats in mid 2026. That is a large number and a small share of a commercial base of several hundred million. Seats sold and seats used are not the same thing, and every analysis of the gap lands on the same cause: people tried it, it did not know their business, they stopped.

In Sweden, Statistics Sweden reported in 2025 that 31 percent of small companies and 50 percent of medium-sized ones used AI, against 72 percent of large firms. Tillväxtverket's 2025 review of SME competence found that smaller firms lack both a strategy and the resources to integrate AI.

What the failures have in common

Put the reports side by side and a pattern appears that has nothing to do with model benchmarks.

The pilots did not fail because the AI was stupid. They failed because the company was, to the AI, unknown.

A sales assistant that does not know your pricing rules writes confident, wrong proposals. A support bot that cannot see the latest process document gives last year's answer. A Copilot that can reach every file in the tenant quotes the draft instead of the approved version, or, worse, a file the asker should never have seen. In every case the model performed exactly as designed. It was given no reliable memory of how this particular business works, so it filled the gap with fluent guesses.

This is the learning gap MIT describes, seen from the inside. The fix is not a better model. It is a better memory.

What a context layer is, in plain words

Analysts call it a context layer, a context graph or an AI context platform. We call it a company second brain, because that is what it does: a shared, governed memory of decisions, processes, ownership and customer knowledge, connected to the systems you already run, kept current by a few short rituals, and readable by people and AI agents alike.

Three properties separate it from the pile of documents most companies already have:

1. It knows what is current. One version of each truth, with a date and an owner.
2. It knows who may see what. Permissions are enforced when a question is answered, not assumed.
3. It can say no. When the answer is not there, it says so instead of inventing one.

Give an ordinary model that memory and the pilot that failed last year often works. Deny it, and the next model will fail the same way, slightly more eloquently.

Why this is an opening for Swedish SMBs

The gap in adoption between small and large Swedish companies is real, and it is tempting to read it as small companies being behind. We read it differently. Large companies have spent two years discovering that licences without context do not pay off. Smaller companies get to skip that lesson.

A company of fifty people can map where its knowledge lives in a week, write the fifty questions it should be able to answer in another, and have a working, governed memory for one team inside two months. The infrastructure can be modest. The discipline is the hard part, and small companies are better at discipline than they think, because everyone knows who owns what.

What to do before the next pilot

• Write the fifty questions your company should be able to answer. Test how many it can today.
• Find out where the answers live, and who holds the ones that are not written down.
• Check who could see what if you connected an AI tomorrow. Fix that first.
• Pick one team and one use case, build the memory for it, and measure the answers weekly.

Do that, and the pilot stops being a pilot. It becomes the first slice of a second brain.