AI in companies, 2026: where it actually stalls
MIT, McKinsey, Microsoft, Gartner: what the big 2025–2026 studies say about AI in companies, and why most of it comes down to context, not models.
Every few months a new study on AI in companies makes the rounds, usually with one striking number that ends up on LinkedIn without the rest. We went back to the main ones from the last year, because they keep pointing at the same problem from different directions. Below is what they say and what we make of it for a mid-sized company.
The return hasn’t arrived for most
MIT, July 2025. The GenAI Divide report from MIT’s NANDA initiative made the biggest splash: “Despite $30–40 billion in enterprise investment into GenAI, this report uncovers a surprising result in that 95% of organizations are getting zero return.” The 95% got quoted everywhere. The explanation got quoted much less: “Most GenAI systems do not retain feedback, adapt to context, or improve over time.”
McKinsey, August 2026. The State of AI 2026 is less dramatic and says something similar. 37% of respondents report that AI has contributed positively to their organization’s EBIT, “essentially unchanged from 2025”, even though more organizations are scaling AI. The group McKinsey calls high performers, with at least 5% of EBIT from AI, stays at about 6%.
Gartner, June 2025. On the newer wave, agents, Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, “due to escalating costs, unclear business value or inadequate risk controls.”
So: lots of use, little measurable return, and the agent projects look set to repeat the pattern.
People use AI anyway, just not the official way
The same MIT report found that “only 40% of companies say they purchased an official LLM subscription”, while “workers from over 90% of the companies we surveyed reported regular use of personal AI tools for work tasks.”
A PagerDuty survey of 1,250 office workers in the US, UK, Australia and Japan, published in June 2026, puts numbers on what that looks like. Two-thirds have used AI tools at work even though they believed company policy didn’t allow it. 88% have shared work-related information with public tools like ChatGPT, Claude or Gemini.
The usual reaction is a ban. We think that misreads the situation. People use private accounts because the AI helps them, and the official setup either doesn’t exist or doesn’t know enough about the company to be useful.
It’s the organization, not the individual
Microsoft’s Work Trend Index 2026, based on a survey of 20,000 AI users in ten countries, has the most useful finding of the lot. Organizational factors like culture, manager support and talent practices account for more than twice the reported AI impact of individual factors like mindset and behavior.
And only 26% of the AI users surveyed say their leadership is “clearly and consistently aligned on AI.”
That matches what we see. Training people to write better prompts helps a little. Whether the company has decided how things work, and whether its AI tools know it, helps a lot more.
What forum discussions add
Studies measure. Forums show what it feels like. In discussions among admins, developers and people building AI note systems, the same complaints keep coming up:
- AI over internal documents gives confident answers from outdated pages.
- Two documents contradict each other and nobody knows which one the AI picked.
- Each tool has its own memory or instruction file, and they drift apart.
- When a long-time employee leaves, a surprising amount of knowledge leaves with them.
We wrote about the personal side of this, the AI second brain and where it breaks for teams, in a separate piece.
What we take from it
Read together, the studies don’t say the models are too weak. They say three other things:
- AI doesn’t know the company. MIT’s phrase is “do not retain feedback, adapt to context.” Every correction a person gives an assistant disappears into one chat.
- The organization is the bottleneck. Microsoft’s numbers say it directly. Value comes from shared decisions, not from individual skill.
- Unofficial use is a signal. If people bring their own AI, the more useful question is how every tool, official or not, gets the same rules.
For a company with 20 to 200 people, that boils down to a fairly practical job: write down how things work here, name who decides each part, and make every AI read that one version. That’s what a company brain is for, and what we build with CtxCore Brain.
You don’t need software for the first step. Pick the question where a wrong AI answer would hurt most and write the rule down once, with a reason and an Owner. Our ten practices for keeping it current cover the rest.