Is AI actually saving small businesses money? What the research says

The short version

The honest answer is "some, for some tasks, if someone sets it up and checks it." Here are the measured numbers on hours, sales and labor cost, and what the winners do differently.

Every vendor says AI will save you money. So what does the research say, once you take the sales pitch out?

Short answer: it saves time on specific tasks. It does not reliably cut payroll or raise sales. And the owners who get a real return do a few boring things the others skip.

This post walks through the measured numbers, as of November 2026. Where a number comes from a survey instead of a measurement, we say so. It's the hub for a set of posts on what AI can and can't do, so we link out to the details as we go. If you want the bigger picture first, start with our guide to AI implementation for small business.

Time: real, modest, and uneven

The best evidence on time saved comes from a few large studies. None of them is about a Tampa shop, so read them as direction, not as a forecast for you.

Customer support, with AI suggesting replies. An NBER study followed 5,179 support agents as an AI assistant rolled out. Agents resolved about 14% more issues per hour. New and lower-skilled agents gained 34%. Experienced agents barely moved. Notice what was happening: the AI suggested replies and a person sent them. It wasn't answering customers alone.

Office-type work in general. The St. Louis Fed ran a national survey of workers in late 2024. Among people who used generative AI, time saved averaged 5.4% of their work hours, about 2.2 hours of a 40-hour week. Across all workers, including people who never touched it, it was 1.4%. Those are self-reported hours, and the tools have changed since. But look at who saved the most. Of daily users, 33.5% saved four or more hours. Of people who used it one day a week, 11.5% did. Using it more, and using it well, is what moves the number.

A reality check on hours turning into dollars. A large Danish study, NBER working paper 33777, linked worker surveys to official records. It found "precise null effects" on earnings and recorded hours, and ruled out effects larger than 2% two years after ChatGPT launched. People saved some time. It didn't show up in pay or in hours worked. That study is about workers, not owner-run firms, and it covers an early period. But it's a useful warning: saved time isn't the same as saved money.

Sales and labor cost: the weak spot

Now the numbers that matter for a cash register.

The New York Fed analyzed a survey of 5,248 small employer firms (fielded in late 2025, written up in October 2026). Among firms using AI, only 31% reported higher sales. And 77% reported no change in labor costs.

Read that again. Most small businesses using AI haven't seen it show up in sales or in payroll. That doesn't mean AI is useless. It means "we use AI" and "AI makes us money" are different statements.

A few more numbers that show the gap:

  • A JPMorganChase Institute report looked at Chase transaction data for 4.6 million small businesses. By December 2025, 17.7% had paid for an AI service. Those are paid tools only; free use isn't counted. Construction was at 8.9%.
  • The same report found median entry spending of $20 a month for firms that started in 2024, rising to $29 in 2025. So the tool itself is cheap. The money question is mostly about what you do with it.
  • A Federal Reserve note put adoption at about 18% of firms by the end of 2025, and said plainly that it doesn't measure productivity effects.
  • Goldman Sachs surveyed alumni of its 10,000 Small Businesses program. Only 14% said AI is fully embedded in their core operations, and 73% said they'd benefit from more training and implementation help. That group isn't a random sample of all small businesses, so take it as one signal.

Different surveys give different adoption rates because they ask different questions. The honest move is to quote the question, not just the percent.

Why results vary so much

Three things show up again and again in the research.

AI is good at some tasks and bad at others, and it won't tell you which

In a Harvard Business School study of 758 BCG consultants, people using AI on tasks inside its ability finished 25.1% faster and did more of them. On a task outside what it could do, they were 19 percentage points less likely to get the right answer. That was a 2023 model, but the lesson holds: the tool sounds equally sure either way. Someone has to know which tasks are which. Our post on where AI works and where it doesn't sorts them task by task.

People overestimate their own gains

A METR randomized trial in 2025 gave experienced software developers AI tools. They took 19% longer, and they believed they'd been about 20% faster. METR notes those results are out of date, and developers are a narrow group. The takeaway isn't about coding. It's that how it feels isn't a measurement. An NBER paper from March 2026 found something similar one level up: executives perceive bigger productivity gains than the measured data show.

Most projects don't reach the goal

IBM's 2025 survey of 2,000 CEOs found only 25% of AI initiatives delivered the expected return, and only 16% scaled across the company. Half said fast investment left them with disconnected, piecemeal technology. Those are big companies, not your shop. But the causes (scattered tools, not enough skill, no plan) are the same ones small owners run into. We dig into this in why AI projects fail.

What the winners do

The evidence doesn't prove a recipe. Nobody has run a trial on "start small" against "go big." But the same factors keep showing up, and they match what the failures lack.

They train people. A BCG survey of more than 10,600 workers found regular use is much higher with at least five hours of training plus coaching. Only about a third said they'd been properly trained. That's a link, not proof of cause. But it fits the NBER support study, where newer staff gained the most. If you want a starting point for your own team, see how to train your team on AI.

One person owns it. Someone is responsible for what the AI does, who can use it, and whether it's working. When it's "everybody's job," it's nobody's.

They start with one task. One repeated job, like after-hours calls or appointment reminders. Not "AI for the whole business."

They measure one number. Hours spent on the weekly report. Missed calls answered. No-show rate. Review response time. Pick one, write down the "before," and check it again in 60 days. Because people overestimate, the number is the only honest judge. Our post on measuring AI ROI shows how for a contracting company.

They keep a person in the loop. AI drafts, a person checks, especially for anything a customer will read or any price.

They keep it running. Tools change under you. A bot that was right in March can be wrong by June. In one Stanford and Berkeley study, the same model family scored 84% on one task in March 2023 and 51% on the same questions in June. That's a 2023 result on one task, and critics say part of the drop was how it was measured. The point still stands: hosted models change, so you recheck.

A simple way to judge your own numbers

Use plain math, with your own figures. Here's an illustration with made-up numbers.

Illustrative example, not a result. Say a front-desk person spends 5 hours a week writing reminder texts and follow-up emails. A tool cuts that to 2 hours with a quick check. That's 3 hours saved. At an $18 an hour wage, that's $54 a week. Now subtract the tool ($20 to $30 a month), setup time, and the time spent checking. What's left is your real saving. If those 3 hours just become 3 hours of scrolling, the saving on paper is zero.

The point of the example: time saved only becomes money if the hours go somewhere. More jobs booked. Less overtime. A task you used to pay someone else for. If you can't name where the hours go, don't count them.

So, does it save money?

For a small business, as of November 2026, here's the fair summary:

  • Likely: time saved on drafting, reminders, summaries and routine questions, especially for newer staff.
  • Possible, if set up and checked: better response to customers, fewer missed follow-ups, less after-hours loss.
  • Not shown by the research: lower payroll, higher sales, or any particular dollar figure. Anyone who promises a number is guessing. We won't, and you shouldn't trust a vendor who does.

Frequently asked questions

How much time does AI really save a small business?

Among workers who use generative AI, the St. Louis Fed survey found an average of 5.4% of work hours, about 2.2 hours a week. Across all workers it was 1.4%. Heavy users saved the most. These are self-reported hours from 2024, so treat them as a rough guide. See our guide to AI implementation costs for the other side of the math.

Will AI let me cut staff?

The research doesn't support planning on it. In the New York Fed survey, 77% of small firms using AI reported no change in labor costs. The more common result is the same people doing more, or doing the work faster. A person checking the output is still part of the job.

Why do some businesses get results and others don't?

The same few things keep showing up: training, one owner, a narrow first task, one number to measure, a person checking the output, and someone keeping it working as tools change. Skip those and you usually end up paying for a tool nobody uses. The research links these factors to success but doesn't prove they cause it.

What should I measure first?

One number tied to one task. For example, missed calls answered, hours spent on a weekly report, or your no-show rate. Write down where it is today, run the change for 60 days, and compare. If the number doesn't move, you've learned that cheaply.

How Di-Hy fits

Our job is the part the research says people skip. We audit where your time goes, set up the one or two things that are proven, train your team, and keep it running. Start with Cut the Busywork for repeated office tasks, or Never Miss a Lead for calls and follow-up. Keep It Running covers the upkeep. We're in Ybor City and work all over Tampa Bay.

Next step: tell us the one task that eats your week. A short call is enough to say whether AI fits it, and what number to track.

Most of these numbers come from surveys and a few large studies on bigger companies and workers. They show direction, not a guaranteed result for any one business. Read in November 2026.

Charles Holmes IV
Written by Charles Holmes IV
Founder of Di-Hy Digital, based in Ybor City, Tampa. He sets up and runs the systems he writes about.
More about us →

Want us to take a look?

Tell us what's not working. Charles reads every message and replies himself. The first call is free.

Thanks, it's in.

We'll be in touch soon. Prefer email? dihydigital@gmail.com

Want us to look at yours?

Book a free 30-minute call. We'll look at how customers find you, how fast they hear back, and what's worth fixing first.

Book a Free Strategy Call →