I’ll start with an uncomfortable number: in a randomized study by the research institute METR, experienced developers using AI tools were 19 percent slower than without them. The remarkable part is not the slowdown. It is that the same developers were firmly convinced afterwards that they had been 20 percent faster.
Let that sink in for a moment. Almost 40 percentage points between perceived and measured productivity, among highly qualified professionals who should know their own work better than anyone.
I am seeing this gap in many executive teams right now. Everyone uses AI. Everyone feels more productive. And at the same time, the International Labour Organization reports an “aggregation paradox”: individual efficiency gains simply do not show up at company level. First drafts are produced faster, and the time saved moves into checking, reworking and correcting. Writing turns into editing. The net effect is a shift, not a gain.
Four disconnects
When I put the research of the past twelve months together, I don’t see a productivity problem. I see four disconnects.
First: the disconnect from one’s own result. In its study “Your Brain on ChatGPT”, the MIT Media Lab used EEG to measure what happens in the brain when people write with AI. The LLM group showed the weakest neural connectivity and the lowest sense of ownership of their own texts, and could barely quote their own work afterwards. The researchers call this “cognitive debt”: debt you take on without knowing the interest rate. If you produce results you no longer recognize as your own, you have lost your connection to your work. That is not an efficiency issue. It is an issue of substance.
Second: the disconnect from one’s own judgment. Language models are demonstrably sycophantic: they agree. Studies show that simply phrasing “I think the answer is X” leads models to adopt a wrong user opinion in about two thirds of cases. After being contradicted, they switch from correct to incorrect answers. The cause is not peer pressure as with humans but a training artifact: models are rewarded when people like their answers. The outcome, however, is the same as in any badly run management meeting: conformity instead of correction. If you use AI to confirm your views, you don’t have a sparring partner. You have an eloquent mirror.
Third: the disconnect from the company itself. A generic model let loose on your own databases does not deliver useful answers. It knows neither the strategy nor the language nor the context of the organization. AI is moving into companies, but it does not understand them. And tools that don’t understand produce exactly what the statistics show: lots of output, little impact.
Fourth: the disconnect from one’s own uniqueness. Strategically, this is the most dangerous one. When everyone points the same copilots with the same standard prompts at the same models, everyone produces the same thing. Research now confirms this empirically: LLM outputs are considerably more similar to each other across model families than human outputs are. Using AI as a thinking partner makes you converge with all other AI users on the same results. At the same time, a Danish study of 25,000 employees in 7,000 firms shows that despite adoption rates of 64 to 90 percent, the effects on wages and hours were practically zero after two years. Mass, unqualified use of standard AI produces exactly what economists would expect: what everyone has differentiates no one. Me-too tools produce me-too products. And worse: what makes your company unique, the experience of your people and your way of solving problems, flows into none of these copilots. It scratches the surface and calls it progress.
One could put it bluntly: we are building organizations in which people are accountable for results they have not fully understood, backed by an AI that mainly agrees with them, based on data that neither of them really understands. And in the end, all competitors look the same. It should not surprise us that none of this shows up on a balance sheet.
The constructive turn: it is not the technology
This article could end cynically here. It doesn’t, because the same research also shows the way out.
The METR study does not say “AI makes you slower”. It says that unstructured use of AI by experts in complex contexts makes them slower, and nobody notices because nobody measures. Research in customer support shows the opposite: for routine tasks and less experienced employees, resolution rates rise by more than 30 percent. And the MIT study shows that people who think for themselves first and use AI afterwards keep their engagement and ownership. The sequence makes the difference, not the tool.
So the problem can be solved. But it is not a tool problem. It is a problem of architecture and leadership. That is exactly why we at BRIDGE developed our approach: building capabilities, with people and AI. Three of its principles address the four disconnects directly.
1. Grounding instead of generic answers. AI needs a model of the company, not just access to its data. Our knowledge graph in BRAIN uses an ontology to build a living picture of the organization: strategy, knowledge, context, responsibilities. Every answer is anchored in your own sources and can be checked. That this works is measurable: at industrial customers, our technical service agent answers specialist questions across more than 140,000 documents with around 96 percent accuracy, precisely because it only says what it can back up. An AI that shows its sources can be checked. An AI that can be checked does not disconnect, it connects. And it is also the answer to the me-too trap: the model of your organization, its knowledge, its language and its particularities, is the only part of your AI stack that no competitor can buy. The language model is a commodity. Understanding is not.
2. People in the loop, structurally rather than symbolically. “Human in the loop” must not be a fig leaf where someone wearily nods through what the machine proposes. In our 6D logic (Discover, Define, Develop, Decide, Design, Deploy), the decision explicitly stays with people, and as a collective decision: AI consolidates perspectives, people wrestle over the assessment and commit to implementation. That is the most effective antidote to sycophancy I know. An AI can tell you what you want to hear. Twenty key people who are accountable for implementation will not. Ownership does not come from delegating to the machine but from taking part in the decision.
3. Measure instead of feel. When professionals misjudge themselves by 40 percentage points, satisfaction surveys about AI use are worthless. For us, this means impact is measured at the level of the organization, in lead times, quality and how far initiatives are implemented, not at the level of perceived individual productivity. Only then can you see where AI carries weight (routine work, access to knowledge, consolidation) and where it does harm (experts’ core work, forming judgment, accountability).
What I advise managing directors today
One more word on the sprawl of agents emerging in many companies right now: if you cover your company with mediocre standard agents, you need two things. First, hygiene: every agent with its own owner, its own identity, a defined end of life and measured accuracy before it gets responsibility for a process. Non-human identities already outnumber human ones in companies by 45 to 1, and only about one in five organizations governs them properly. Second, a two-layer architecture: allow standard copilots for personal productivity but keep them away from core systems, and direct investment into the layer that really understands the company. The good systems then replace the mediocre ones not by decree but because nobody voluntarily goes back to an unsupported answer.
Stopping the use of AI would be the wrong conclusion; the competitive disadvantage would be greater than today’s paradox. But “more licenses” is not a strategy. The question for 2026 is not whether your people use AI. They have long been doing so, with or without approval. The question is whether your organization has an architecture in which AI understands your company, your people decide and your numbers tell the truth.
Individual productivity has gone up. Now it is up to leadership to turn it into the productivity of the whole.
How do you measure in your company whether AI really has an effect, or only feels like it does? I look forward to the conversation.