Most small businesses use AI. Very few have put it to work.

What the research says

The tools work. The way businesses adopt them usually doesn't.

In July 2025, a team at MIT's Project NANDA published a study built on more than 300 public AI projects, interviews with 52 organizations and a survey of 153 senior leaders. Its headline finding traveled fast.

The bar MIT used was strict. A project counted only if it moved past the trial stage, had measurable goals, and still showed an impact on productivity or profit six months later. Most didn't clear it.

What matters for you is why. MIT says the cause was not the quality of the AI, and not regulation. It was approach. Most AI tools don't keep what they learn, don't adapt to the business they're used in, and don't get better over time. The researchers call this the learning gap. Users described it in plain terms: they were tired of re-explaining their business every time, and the tool kept making the same mistakes. One interviewee summed up a typical tool this way:

"It's useful the first week, but then it just repeats the same mistakes."

Small businesses are living a version of this right now. A 2026 survey of 1,256 owners in Goldman Sachs' 10,000 Small Businesses program found that 76% use AI, but only 14% have fully built it into their core operations. Nearly three in four said they'd benefit from more help implementing it.

Bigger companies are starting to close the gap, and the difference is discipline. Wharton's 2025 survey of enterprise leaders found three in four now report positive returns, and 72% formally measure what AI is doing for them. The ones getting results decided what AI was for, wrote it down, and kept score.

A note on scale: MIT's sample skews toward larger organizations and Wharton's covers enterprises only. The patterns hold for a 10-person business. The exact percentages may not.

The context gap

Generic AI knows everything except your business.

An AI tool out of the box has read most of the internet. It has never read your price sheet, your service area, your guarantee, your rebooking rules, or the way your best employee talks to a nervous customer. So when it answers a question about your business, it guesses. The guess sounds confident, which makes it worse.

Here is the same customer question answered two ways.

The second answer isn't smarter AI. It's the same kind of AI working from the right information. That information is what we call a Source of Truth: your services, prices and pricing rules, service area, policies, guarantees, hours, and voice, written down once, checked by you, and kept current. Every AI job in the business reads from it.

Without it, every tool you add is another place for wrong answers to start.

How it shows up

Five signs AI is in your business but not working for it.

What working AI has in common

Foundation first. Then one job. Then a number.

The businesses that get results from AI don't start with the flashiest tool. They get three things in place, in this order.

MIT's research points the same direction on who builds it. Projects done with an outside partner reached full use about twice as often as ones built entirely in-house. The researchers are careful to say that's a pattern, not proof. For a small business without a technical team, it usually comes down to time: building, testing and maintaining an AI system becomes a second job that nobody was hired to do.

Where to start

A good first job in businesses like yours.

The best first job is rarely the most exciting one. It's the one that repeats, follows rules, and quietly leaks time or customers.

The first 90 days

What a measured first quarter looks like.

MIT found the faster-moving mid-sized companies went from trial to full use in about 90 days, while large enterprises took nine months or more. A small business has even fewer approvals in the way. Here is the shape of a quarter that earns that pace.

1. Weeks 1 to 2

   Build the foundation

   Name the owner. Agree on one page of AI rules. Move everyone to company accounts. Write and verify the Source of Truth. Pick the first job and record its baseline.

2. Weeks 3 to 6

   Put AI inside the job

   Connect it to the tools the work already happens in, including the hand-offs and the exceptions. Decide what a person approves before it goes out. Train the two or three people who run the job.

3. Weeks 7 to 12

   Measure and decide

   Compare against the baseline every two weeks and update the Source of Truth when something is wrong. At week twelve, the owner makes the call: keep it, fix it, or stop it. That rule is agreed in week one.

4. Day 91

   Start the second job

   The foundation is already built, so the second job goes faster than the first. That is where the returns start to stack.

Run your own number

What is one job worth?

Pick one recurring job in your business. Enter roughly how many hours it takes each week, what an hour of that person's time costs you with taxes and benefits, and how much of it AI could reasonably take over. Treat the result as a starting estimate. The real number is the one you measure.

This counts time only. For jobs like after-hours replies or estimate follow-up, the bigger number is often the customers you stop losing, which is why a baseline matters.

Scorecard

Is AI working in your business?

Seven yes-or-no questions. Check the ones you can honestly answer yes to today.

Free starter

One page of AI rules for your team.

If your people are using AI, they're already making decisions about what to put into it. This page makes those decisions once, for everyone. Fill in the blanks, talk it through with your managers, and post it where people will see it.

[Business name] AI Use Rules
Effective: ________     Owner: ________

1. Approved tools. Use only the company AI accounts listed here: ________. Do not use personal AI accounts for company work.

2. Never put these into any AI tool: customer card or bank details, Social Security or ID numbers, passwords, patient or client health information (unless the tool is approved for it in writing), or anything a customer asked us to keep private.

3. Good uses: first drafts of emails, texts, posts and estimates; summarizing notes and calls; answering questions from our approved business information.

4. A person checks before it goes out: prices, quotes, guarantees, health or safety guidance, and anything sent under the company's name.

5. If something goes wrong (a wrong answer reached a customer, or information went somewhere it shouldn't), tell ________ the same day.

6. New AI tools need approval from ________ before anyone signs up.

We review these rules every 90 days.

A starting point, not legal advice. Businesses that handle health or financial information should have these rules reviewed against the regulations that apply to them.