You already have AI in your business. You did not buy it, and you do not govern it.
MIT's Project NANDA measured this directly as part of its GenAI Divide study. Only 40% of companies had purchased an official AI subscription. Workers at more than 90% of companies reported regularly using personal AI tools for work anyway.
That is a fifty-point gap between what leadership believes is deployed and what is actually deployed. It is not an exception — it shows up in nearly every company that runs the check. The study also noted that the unsanctioned use is often producing better real returns than the sanctioned, budgeted projects sitting next to it.
In this business the gap has a specific shape
An agent pastes a client's financial circumstances into a consumer chatbot to draft a delicate email. A coordinator drops an executed contract into a free tool to summarise the contingencies. Someone asks a chatbot to rewrite listing copy without knowing that Fair Housing language is precisely the category these models get subtly, confidently wrong.
RPR's survey of agents backs this up from the other direction.
| Concern | Share naming it |
|---|---|
| Accuracy of outputs | 63% |
| Compliance or legal exposure | 49% |
| Misinterpretation of market data | 47% |
| Fair Housing specifically | 28% |
They are worried about exactly the right things. They are also doing it anyway — on personal accounts, on their phones, with your client data.
Ninety percent is what a ban produces, not what it prevents. It only moves the activity somewhere you cannot see, which is the state you are already in.
What works is unglamorous
Name the two or three tasks it is genuinely good at in your shop. Give people a sanctioned way to do exactly those. Say plainly, in one sentence people remember, what never goes into a chatbot — client financials, executed documents, anything with a protected-class characteristic in it.
That is most of the exposure closed, and it costs a meeting.
The second finding, for when you do decide to buy something
The same study found that tools bought from a vendor with a real implementation partnership accounted for 66% of successful deployments, against 33% for internal builds — twice as likely to succeed, in MIT's own summary. Against a backdrop where 95% of enterprise AI pilots produced no measurable return at all.
The most common instinct in a small shop — have the tech-savvy agent build something over a few weekends — is the single most reliable path into the 95%.
Not because that person is not capable. Because internal builds have no implementation discipline behind them, no defined outcome agreed before the build starts, and nobody whose job is to make sure it actually enters the workflow it was built to change.
The test. One afternoon.
Ask three people in your office what they used AI for last week. Not whether — what. Specifically. Which tool, which task, which information went into it.
You will get a list. That list is your real AI deployment, and it is almost certainly the first time anyone has written it down. Two columns: what should be sanctioned and made easier, and what needs to stop today.
That is a genuinely useful hour and you do not need me for it.
Sources
- MIT Project NANDA, The GenAI Divide: State of AI in Business 2025, July 2025. 52 organisations interviewed, 153 senior-leader survey responses, 300+ public initiatives reviewed. 40% purchased official subscriptions vs 90%+ reporting personal-tool use; strategic partnerships 66% of successful deployments vs 33% internal builds (“twice as likely to succeed”); the report cautions this correlation does not prove causation; 95% of pilots with no measurable return. Report (PDF), coverage.
- Realtors Property Resource (RPR), 82% of Real Estate Agents Use AI. The Real Gap Is Confidence, 12 February 2026. Survey of 225 US NAR members. Article.