Before you buy your company an AI tool, check what your people already use today. On their own accounts, from their own phones, without asking IT.
MIT measured this last year on a sample of 52 interviews, 153 surveys and three hundred publicly described deployments. An official subscription to a language model had been bought by 40% of companies. Employees use personal tools for work tasks in over 90% of companies. The authors themselves caution that this is qualitative research, not a statistic. The direction is hard to dispute, though, because anyone who manages a team sees it at home.
The report describes a lawyer whose firm bought a contract analysis tool for 50,000 dollars. She still works in ChatGPT. Her explanation is one sentence: "it gives rigid summaries, in ChatGPT I can iterate".
Employees are not doing this against the company. They do it because the tool in their pocket answers immediately, remembers the context of the conversation and accepts corrections. The tool the company bought usually does not. In the MIT interviews the same complaint returns in different words: "it breaks on edge cases", "I have to paste the whole context every time", "useful for a week, then it repeats the same mistakes". The biggest barrier to scaling in the report's ranking is people's reluctance to adopt new tools, not cost and not technology.
There is a second side to this grey zone. When a salesperson pastes a customer enquiry into a private chat, and an accountant pastes a letter from the tax office, company and customer data land on an account the company has no control over. One leader in the report says it plainly: "I can't risk our client data ending up in someone else's model". An official tool solves this problem only if people want to use it.
Companies whose deployments work do one thing differently in the report. Before they buy anything, they look at what people already use private tools for, and they start there. The strongest deployments began with a few people who experimented on their own, and the ideas came from line managers, not from a central AI team. A mid-sized company goes from pilot to full deployment this way in 90 days. A corporation with a central programme needs nine months or more.
The same study shows where AI has already won and where it has not. For writing e-mails, 70% of respondents prefer AI; for simple analysis, 65%. For projects that run for weeks and for managing a client relationship, 90% prefer a human. The line does not run through the model's intelligence. It runs through memory and learning from corrections.
That gives a simple rule for the first official deployment. It has to beat private ChatGPT on the same pitch: answer immediately, know the company's context and learn from what people correct. That is why the agent we put on a client's website starts from the client's own knowledge, not from a general model, and every morning leaves a report of what it did not know, with a ready correction to approve. If the person in customer service cannot fix an answer in a minute, they will go back to ChatGPT, and they will be right. That applies to us too.
McKinsey, on a larger sample of 1,719 companies, shows the same divide from the other side: nine in ten companies use AI, and 6% see a clear effect on the bottom line. The difference between those groups is not in the models. It is in whether the tool entered daily work or stayed next to it.
At your next team meeting, ask one question: what do you use ChatGPT for at work, on your own accounts. Write the list down. You now have a map of your first deployments and the names of the people who will lead them. If the list includes "answering customers' questions about our offer", start there.
Want to see this on your own website rather than on slides? Enter your website address and e-mail in the demo form. In 20 minutes you get an agent trained on your offer, ready to be quizzed.
Sources
- MIT NANDA, The GenAI Divide: State of AI in Business 2025 (52 interviews, 153 surveys, 300 deployments; Fortune coverage, 18 Aug 2025): https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/
- McKinsey, The State of AI in 2026: on the road to ROI (n=1 719): https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai