Two headlines about AI went around this year, and they can't both feel true.
The first says companies are getting $1.49 back for every dollar they put into AI. The second says 95% of companies get zero return. If you own a small business and you're trying to decide whether AI is worth your time and money, that's not much help.
Both numbers are real. They come from different studies that asked different people different questions. Once you see what each one measured, the picture gets a lot clearer.
The short answer: AI pays off when it's pointed at one real job, fed good information and measured honestly. Most AI projects that fail, fail on those three things, not on the AI. If you only want the action steps, they're the five at the end.
ROI just means return on investment: what you get back for what you put in. Here's what the 2026 numbers say, where each one comes from, and who paid for it. Every source is linked at the bottom.
The good news: AI is paying off for those who measure it
The biggest 2026 study is "The ROI of Gen AI and Agents," done by the research firm Omdia and paid for by Snowflake, a company that sells data software. They surveyed 2,050 business and IT leaders at companies already running generative AI.
Their headline number: companies that measured their return reported 49% ROI. That's $1.49 back for every dollar spent. A year earlier the same research found 41%. And 92% said they were seeing some kind of return.
Two things to keep in mind. First, these are companies with 500 or more employees. They were picked because they already had AI working. Second, the study was paid for by a company that benefits when businesses spend more on data. That doesn't make it wrong. Read it as one strong data point, not the final word.
The same report found that 48% of the code at these companies is now written with AI. That's one place the return is easy to see.
The bad news: most AI projects stall before they pay
The other headline came from a research project at MIT in 2025. Its report said that despite $30 to 40 billion spent by businesses on generative AI, 95% of organizations are getting zero return.
That number spread fast, so it's worth knowing what it measured. A "pilot" is a small test run of an AI tool. MIT counted a pilot as a success only if it moved past testing and showed a measurable profit impact within about six months. The research drew on 52 interviews, 153 survey responses and a review of public projects. The authors themselves call it "directionally accurate," and critics pointed out that a pilot that helped but didn't show up on the profit line still counted as a failure.
A UC Berkeley Executive Education piece pushed back. Its argument: this isn't an AI failure, it's "a measurement failure." A lot of companies are measuring the wrong things at the wrong time.
So who's right? Probably both. Plenty of AI projects really do stall. Plenty that work never get measured. The real question for you isn't whether AI works in general. It's whether you'll pick the right job for it. And measure it honestly.
Why AI projects fail: it's usually the data, not the AI
Here's the most useful finding in the whole report. When companies hit trouble, it usually isn't the AI. It's everything around it.
- 79% reported at least one technical or data problem.
- Only 20% of their everyday information (emails, documents, notes) was in good enough shape for AI to use. For organized data like spreadsheets, it was 32%.
- 60% said storage and computing costs pushed their AI projects over budget.
The report put it plainly: the problems "were not with the gen AI and agentic technologies themselves." They were talent and data.
A small business has a smaller version of the same job. You probably don't have twenty systems and a decade of tangled databases. You have an inbox, a calendar, a spreadsheet or two, and maybe a customer list in a program like QuickBooks or a CRM. That's still work to tidy up. But it's a lot less of it.
The skills gap hits smaller companies harder
Omdia found that companies with 500 to 999 employees were more likely to say they lacked the skills to make AI work: 43%, compared with 34% of the biggest companies. The gap shrank as companies got more practice.
The survey didn't include anyone under 500 employees, so it can't tell us about a 12-person shop. What it does show is that skill comes from practice, not size. A small business doesn't need an AI department. It needs one or two people who use a few tools well.
Shadow AI: the boss is the biggest rule breaker
This one surprised people. 57% of respondents admitted to using AI tools their company hadn't approved. It's called "shadow AI." And the worst offenders weren't junior staff. 66% of top executives on the business side said they did it, more than middle managers (56%).
Microsoft and LinkedIn found the same thing in 2024: 78% of people who use AI at work bring their own tools, and 80% at small and medium-sized companies.
Why does it happen? Omdia asked. 60% said the approved tools couldn't do what they needed, and 53% said getting approval took too long.
In a small business, shadow AI usually looks like someone on the team, owner included, pasting customer details into a free chatbot because it's quick. Nobody means harm by it. The fix isn't a ban. Pick the tools you trust. Decide what never goes into them. Then tell your team.
The honest counterweight on AI agents
"AI agents," AI that does whole jobs on its own, are the hot topic of 2026. The Omdia report found 32% of these companies already have agents running.
But the same people were refreshingly blunt about them:
- They expect about 41% of the agent projects they sponsor to fail over the next three years.
- 49% rated today's agent tools somewhere between "early but unproven" and "mostly vaporware," meaning more promise than product. Only half called them functional and maturing.
That's not a reason to avoid agents. It's a reason to start small. Pick jobs where a mistake is cheap and easy to catch.
What AI is doing to jobs
The numbers are mixed. Stanford researchers found that young workers (22 to 25) in the most AI-exposed office jobs are employed at levels 19% below where they'd otherwise be, up from 13% in their first version. Older workers in the same jobs show no gap. Meanwhile, 47% of small businesses told the US Chamber of Commerce that AI is creating jobs for them, and 6% said it's letting them cut staff.
AI statistics for small business in 2026
None of the big studies above surveyed small businesses. Here's what the small business surveys say:
- US Census Bureau: about 20% of all US businesses use AI, and less than 20% of firms with four or fewer employees. Larger firms (250+ employees) are at 37%. (May 2026)
- NFIB: 24% of small business owners used AI in its 2025 survey.
- US Chamber of Commerce: 66% of small businesses use AI, up 8 points in a year. (2026)
- Goldman Sachs 10,000 Small Businesses: 76% of the owners in its program use AI, but only 14% say it's fully part of how they run the business. (2026)
- Intuit QuickBooks: 77% of US small businesses use AI regularly, up from 48% in mid-2024. (2026)
- SBE Council: 82% of owners with 2 to 99 employees use at least one AI tool, and they report saving a median of 5 hours a week for the owner. (2026)
Why the huge spread, from 20% to 82%? The Census is a large official survey of every kind of US business, down to one-person shops. The others mostly survey their own members, program graduates or customers, who tend to be more engaged with technology to begin with. Both are honest. They just measure different groups.
Most of these surveys agree on one thing: trying AI is common, but building it into the business isn't. Goldman found 76% using it, but only 14% with it fully part of how they operate.
So, is AI worth the investment for a small business?
Here's how we'd read all of this, plainly:
- Start with one job, not a strategy. The big companies seeing returns didn't start with everything. Pick one repeat task that eats your week, like quotes, follow-ups or answering the same questions.
- Get your information in order first. The number one reason AI projects stall is messy data. For a small business, that means a clean customer list, your prices written down, and your common answers in one place.
- Measure it from day one. Write down how long the job takes now. Check again in a month. Most "failed" AI projects were never measured.
- Pick your tools on purpose. Choose one or two you trust, decide what never goes into them (card numbers, Social Security numbers, private customer details), and tell your team. That's your answer to shadow AI.
- Be skeptical of anything that does it all for you. Even the companies running AI agents expect four in ten of those projects to fail. Start small, keep a person approving the work, and grow from there.
Frequently asked questions
What is the average ROI of AI in 2026? In the largest 2026 study, companies that measured their return reported 49%, about $1.49 back for every dollar. That study covered companies with 500 or more employees.
Is it true that 95% of AI projects fail? That's from a 2025 MIT report that counted a project as a failure if it showed no measurable profit within about six months. Critics say it overstates the problem.
Why do AI projects fail? Mostly data and skills, not the AI itself.
How many small businesses use AI? Between about 20% (US Census Bureau) and 82% (SBE Council), depending on who's asked.
The bottom line
The 2026 AI ROI statistics don't say AI is magic, and they don't say it's a waste. They say it pays when it's pointed at a real job, fed good information and measured honestly. None of that takes a big budget. It takes one good choice and the discipline to check whether it worked.
If you'd like help picking that first job and setting it up right, that's what our AI training is for. Book a call, and we'll look at how you work today and where AI would actually earn its keep.
Sources: Omdia, "The ROI of Gen AI and Agents 2026" (commissioned by Snowflake) · Snowflake summary of the report · MIT NANDA, "The GenAI Divide: State of AI in Business 2025" · UC Berkeley Executive Education, "Beyond ROI" · Stanford Digital Economy Lab, "Canaries in the Coal Mine?" · US Census Bureau, AI use by businesses · US Chamber of Commerce, 2026 Empowering Small Business · Goldman Sachs 10,000 Small Businesses · NFIB Small Business and Technology survey · Intuit QuickBooks AI Impact Report · SBE Council Technology Use Survey · Microsoft and LinkedIn Work Trend Index 2024