Adoption Is Not Value (And the 95% Headline Proves It)

If you have heard one statistic about enterprise AI this year, it is that 95% of pilots fail. It gets quoted in board meetings as proof of either inevitability or futility, depending on who is talking. Almost everyone quoting it has misunderstood it, and the misunderstanding matters.

What the number actually says

The finding comes from MIT's Project NANDA, in a 2025 report on the state of AI in business. The precise claim is that only about 5% of integrated AI pilots are extracting significant, rapid value to the bottom line. That is a specific and demanding bar, rapid P&L acceleration, not a general statement that 95% of AI does not work. The evidence base is modest: a few hundred public deployments, a few dozen interviews, a couple of hundred survey responses. Credible critics pushed back hard on the framing.

Read carefully, the report says something far more useful than the headline. The barrier is not model quality, and it is not regulation. It is a learning and integration gap: pilots that were never wired into the actual workflow. The report also found that AI tools bought from vendors succeeded roughly twice as often as tools built in-house, for the unglamorous reason that the bought ones arrived already integrated. And it found a shadow AI economy: employees using personal AI tools at scale while their employers debate official ones.

None of that says stop. All of it says integration and adoption are where pilots die, not the model.

Adoption is necessary. It is not the finish line.

Here is the trap the 95% number sets, and the one your own dashboards will set for you. It is easy to measure adoption, things like seats activated, prompts entered, and weekly active users, and tempting to treat those numbers as success. They are not. They are necessary, and they are not sufficient.

The organisations with the best adoption numbers say so themselves. JPMorgan, with a quarter-million employees on its platform and about half using it daily, is candid that there is a gap between what the technology can do and what the enterprise has actually captured, and that closing it will take years. That is the most honest sentence in enterprise AI, and it comes from one of its biggest adopters.

McKinsey shows what drives adoption when it works: its internal assistant reached the large majority of staff not through a mandate but through leadership modelling it, with partners visibly using it and asking "have you asked Lilli?" Adoption is a change-management outcome, not a software feature. But even McKinsey notes the work of converting adoption into sustained engagement is not finished.

What over-automation costs

The sharpest lesson comes from getting adoption backwards: removing the humans instead of equipping them.

Klarna automated two-thirds of its customer service and spoke publicly about doing the work of 700 agents. Within roughly a year its CEO reversed course in public: too much focus on cost, the result was lower quality, and that was not sustainable. The company began rehiring. The failure was not the model's capability. It was the decision to take the human out of the loop entirely, with no path to escalate when the machine reached the edge of what it could do.

Contrast that with how regulated content work actually scaled. Where pharma cut medical-writing time dramatically, with clinical study report first drafts going from weeks to days, it did so with the medical writer firmly in the loop, reviewing and owning the output. The human in the loop is not the thing slowing the system down. It is the thing that makes the output trustworthy enough to use. In a patient-facing, regulated context, it is the difference between a deployable system and a liability. The most expensive healthcare AI failure on record generated treatment recommendations its own documents called unsafe, and they never reached patients, precisely because a human was still required between the model and the decision.

Value lives past the gate

This is the load-bearing idea, and it is why the AI Value System does not draw value as a layer.

In AIVS, Human Adoption is a gate. Everything upstream, meaning governed knowledge, governance, coordination, and orchestration, exists to get a useful output to a person. But nothing produces value until that person does something differently as a result. Value is behaviour change past the gate: the medical writer who ships faster and still owns the science, the analyst who finds the document in seconds instead of hours, the customer whose problem is actually solved.

If you draw value as a box you build, you will optimise for the box, for outputs, dashboards, and activity. If you treat value as an outcome that only appears past the adoption gate, you will optimise for the thing that matters, which is whether anyone's behaviour actually changed. Adoption is how you get to the gate. It is not what is on the other side.