MIT NANDA 95% AI Report: What It Says About Shadow AI
The MIT NANDA '95% zero return' headline covered custom AI projects, not chatbots. Here is what the report says about shadow AI and how adoption spreads.
For Executives, IT and Managers

What does the MIT NANDA 95% AI report actually say?
The 2025 MIT NANDA report is best known for one line: 95% of organizations are getting zero return on generative AI, despite an estimated $30 to $40 billion in enterprise investment. That figure describes custom AI implementations that companies built or bought as formal projects. It left out general chatbots like ChatGPT and Claude. The same report also found that about 90% of companies have employees regularly using personal AI tools for work, often without IT approval.
TL;DR
- The 95% figure measures formal, custom AI projects, not everyday chatbot use. Reading it as "AI doesn't work at work" is a misreading.
- Chatbots were specifically excluded from the count, so the headline says little about tools like ChatGPT or Claude.
- The report describes a shadow AI economy: employees paying for AI tools themselves and using them even when the company doesn't allow it.
- About 90% of companies report regular use of personal AI tools for work, so adoption is widespread but mostly unofficial.
- Adoption inside companies spreads unevenly. It tends to start with coding, then finance, then legal, with life sciences and manufacturing following more slowly.
- For most workers and managers, the practical question is not "does AI work here?" but "how do we use it on purpose, safely, and with results we can measure?"
What was the 95% figure actually measuring?
The study looked at custom AI implementations: the pilots and projects a company launches to automate a process or build an AI-powered capability. These are the efforts that get a budget line, a project manager and a steering committee. The 95% number says that almost none of those efforts produced a measurable return.
That is a real finding, and it is worth taking seriously. A lot of formal AI projects stall between demo and daily use. But it is a narrow finding. The study did not count the chatbot subscriptions people use to draft emails, summarize documents, write formulas or debug code.
When a headline says "95% of organizations get zero return," most readers hear "companies are spending billions and AI doesn't work." The report says something closer to "companies are struggling to turn formal AI projects into results, while individuals are quietly getting value from simple tools."
The headline circulated widely without most people reading the document itself. If your boss or a colleague quotes the 95% number, it helps to know what it covered. If you want the full picture, read the report directly instead of relying on summaries (including this one).
What is shadow AI?
Shadow AI is the use of AI tools at work that the company has not approved, bought or set up. The classic example: an employee pays $20 a month out of pocket for a chatbot subscription, then uses it to speed up reports, emails and analysis at their job.
The report calls this a shadow economy. Employees use ChatGPT and Claude even when the company does not allow it, and they pay for it with their own money. Two details from the report stand out:
- About 90% of companies report regular use of personal AI tools for work.
- Almost every person used large language models (LLMs, the technology behind chatbots) often, and often without IT's knowledge or approval.
So while the official AI program may be stuck, the unofficial one is running every day. The gap between those two is the story the 95% headline hides.
Why do employees use AI tools on their own?
The reasons are practical.
- The tools are easy to get. A personal account takes minutes. A formal procurement request can take months.
- They help with real tasks right away. A chatbot can draft, summarize, explain and calculate without any integration work.
- Nobody has to change a system. Custom AI projects often need data connections, process changes and training. A person using a chatbot needs none of that.
- The person closest to the work knows what to try. A recruiter, an accountant or a marketer can test an idea in an afternoon without waiting for a project to be scoped.
This is also why shadow AI is hard to stop with a policy memo. If the approved route is slower than the personal route, people take the personal one.
How does AI adoption spread inside a company?
According to the report, adoption is not even across departments. The pattern it describes:
| Stage | Where AI shows up | Why it spreads there |
|---|---|---|
| First | Coding | The fastest and earliest adoption inside enterprises |
| Next | Finance | Used to automate financial decisions or comb through large amounts of data |
| Then | Legal, life sciences | Document-heavy work where reading and summarizing matter |
| Slowly | Manufacturing | Starting to be used to make operations more efficient |
If you work in a function further down that list, expect more of the same: individual people trying tools first, followed by teams, followed by a formal program (or a formal ban). If you work in finance or coding, you are probably already past the experimentation stage.
Is shadow AI a problem?
It can be. The risks are real even though the benefits are too.
- Data exposure. Pasting customer records, contracts or financial data into a personal account may break company policy or legal obligations, depending on the tool's terms and your industry.
- No visibility. If IT does not know which tools are in use, it cannot assess them, secure them or support them.
- Uneven quality. One person's careful prompt routine is another person's unchecked copy-and-paste.
- Lost knowledge. When a workflow lives in one person's private account, it leaves with them.
The report's finding does not say shadow AI is good. It says it is happening at scale, which means a "no AI" rule is not the actual state of affairs in most companies. Pretending otherwise is the riskiest option.
What does this mean for you?
If you are an individual employee. Check your company's AI policy before putting anything sensitive into a personal tool. Keep customer data, financials and anything confidential out of consumer accounts unless you have explicit approval. Then document what is working: which tasks, how much time saved, what you checked by hand. That record is useful if you want to make the case for an approved tool.
If you are a manager. Assume your team is already using AI. Ask openly, without making it an interrogation. Find out which tasks people use it for, then decide what to allow, what to approve and what to prohibit. A clear "here is what you can use and for what" beats silence.
If you are in IT, security or leadership. The report suggests the demand is already there, so the job is to channel it. Offer approved tools that are as easy to get as the personal ones. Be specific about data rules. Start with the tasks people already do with chatbots, and measure those outcomes before funding large custom projects.
If you are evaluating an AI project. Ask what it will replace or speed up, who will use it daily, and how you will measure the result. The 95% figure is a reminder that a pilot with no owner and no metric is likely to stall.
Could you build the small tool yourself instead of buying another subscription?
Much shadow AI exists because there is no approved tool for a small, specific job: a request intake form, a tracker for contract renewals, a dashboard that pulls a few numbers together. If your team keeps working around that gap with spreadsheets or personal accounts, one option is to describe the tool in plain English and have it built. Remy is an AI product agent that turns that description into a deployed full-stack app with a database and sign-in. For companies, it aims to give employee-built software a governed home, with SSO and an audit log, instead of personal laptops and private accounts.
Frequently Asked Questions
Does the MIT NANDA report say AI doesn't work?
No. It says most custom, formal AI projects did not produce measurable returns. It excluded chatbots from that count, and it separately found that employees at about 90% of companies regularly use personal AI tools for work.
What does "95% zero return" refer to?
It refers to organizations' returns on custom AI implementations, set against an estimated $30 to $40 billion in enterprise generative AI investment. It does not measure everyday chatbot use, which the study left out of the count.
What is the shadow AI economy?
It is the informal use of AI tools that employees pay for and use themselves, often without their employer's knowledge or approval. The report describes it as widespread, with nearly every person using LLMs often.
Which departments adopt AI first?
The report describes adoption starting with coding, which is the fastest and earliest. Finance is the next big wave, followed by legal and life sciences. Manufacturing is beginning to follow, more slowly.
What should my company do about shadow AI?
Start by finding out what people already use and why. Set clear data rules, offer approved tools that are easy to get, and measure results on the tasks people already do before investing in large custom projects. A blanket ban tends to push use further out of sight.
How we source, write, and correct our articles. Editorial standards
Related articles

What Caused the Microsoft 365 Outage? Network Fault, AI Debate

What Is Shadow AI? Corporate Liability Risks and Who Is Accountable

What Is MCP Shadow IT? How Employees Bypass Enterprise AI Plans
