Is there a single average ROI figure for AI in small businesses?

No credible one exists. The figures that circulate, like $3.50 back per $1, either carry no source or point to unverified vendor-commissioned surveys of self-reported estimates. Rigorous sources measure narrower things: hours saved per user, task speed, project abandonment rates. Build your number from your own hours and costs instead of borrowing an average.

How fast should an AI tool pay for itself?

For a typical off-the-shelf tool under $50 a month, the math either works within the first month or two or it does not work at all. If two rounded-down hours a week at a modest delegation rate cannot cover the subscription, the workflow is probably the wrong target. Cancel and test the next candidate.

Do the '95% of AI pilots fail' headlines apply to a small business buying a $40 tool?

Not directly. That figure comes from MIT Project NANDA's preliminary 2025 research on enterprise GenAI pilots, based on 52 interviews and 153 leader surveys, and it describes organizations getting zero return from custom pilots. In the same self-reported sample, purchased external tools reached deployment about twice as often as internal builds. Buying a proven tool for one mapped workflow is a different bet.

Does time saved actually equal money saved?

Only if the hours get redeployed. An hour back becomes cash when it turns into billable work, a sales call, or overtime you no longer pay. It is worth nothing if it dissolves into the day. Conservative math values saved hours at the rate you would pay to delegate the task, not the owner's billing rate.

What should I measure before buying anything?

One honest week of hours. Track where repetitive time goes: drafting, scheduling, chasing invoices, re-typing data between systems. The Federal Reserve's 2025 small business survey found writing and marketing is the most common AI use, at 83% of AI-using firms, but your biggest sink may be elsewhere. Measure first, then match a tool to the worst one.

Why do published AI ROI statistics disagree so wildly?

They measure different things on different populations. Controlled experiments measure task speed. Surveys measure self-reported impressions, often of a vendor's own customers. Abandonment studies measure enterprise projects. Representative Census-style surveys show far lower adoption than industry surveys of self-selected firms. Before trusting any number, check who was asked and what unit was measured.

Is a paid assessment worth it before buying tools?

It depends on whether you can map your own workflows honestly. If you can log a week of hours and run the payback formula in this article, do it yourself. An assessment earns its fee when it finds time sinks you would miss, prices tools you do not know exist, and stops you from commissioning a build you do not need.

The Realistic ROI of AI for a Small Business (2026)

Quick answer (updated August 2026): There is no reliable average ROI for AI in a small business, and pages that claim one rarely cite a source. What verified research shows: workers who use generative AI report saving about 2.2 hours a week (St. Louis Fed), yet 42% of companies abandon most of their AI initiatives, up from 17% (S&P Global). Returns come from matching a cheap, proven tool to a measured workflow, then doing the payback math yourself. The formula is below.

What does the research actually show?

Two bodies of evidence exist, and they point in opposite directions. Individual-level research finds real, measurable time savings when a person uses generative AI on the right task. Organization-level research finds that AI projects fail, stall, or get abandoned at striking rates. Both are true, and any honest ROI page has to hold both. Here is the evidence at a glance, every figure from a named primary source:

Finding	Source	Year
GenAI users report saving ~2.2 hours per 40-hour week (5.4% of work hours)	St. Louis Fed	2025
Writing tasks done 40% faster, quality up 18%, in a controlled experiment	Noy & Zhang, Science	2023
71% of small-firm AI users say AI increased productivity; just 7% of users have fully integrated it	Federal Reserve SBCS	2026
24% of small businesses say AI made their workdays shorter	Intuit QuickBooks	2025
Companies abandoning most AI initiatives rose from 17% to 42%	S&P Global	2025
More than 80% of organizations see no tangible EBIT impact from genAI	McKinsey	2025
Only 26% of companies get past proofs of concept to tangible value	BCG	2024

The savings side

The most defensible time number in circulation comes from the Federal Reserve Bank of St. Louis. In a November 2024 national survey, researchers found generative AI users saved an average of 5.4% of their work hours, which works out to about 2.2 hours in a 40-hour week. That is savings reported by actual users, not a vendor projection. The experimental evidence is stronger still: in a preregistered experiment published in Science (Noy and Zhang, 2023), 453 college-educated professionals completed occupation-specific writing tasks, and those randomly given ChatGPT finished 40% faster with 18% higher-rated quality. And among small firms specifically, the Federal Reserve’s 2025 Small Business Credit Survey found 71% of AI-using firms said AI increased productivity. These numbers are modest. Two hours a week is not a revolution. It is, priced correctly, a very good return on a $40 tool.

The failure side

The failure numbers are just as well documented. S&P Global’s 451 Research surveyed 1,006 companies and found the share abandoning most of their AI initiatives jumped from 17% to 42% in a year, with the average organization scrapping 46% of proof-of-concept projects before production. McKinsey’s March 2025 State of AI survey found more than 80% of organizations report no tangible impact on enterprise-level earnings from generative AI. BCG put it at 74% of companies struggling to achieve and scale value, with only 26% getting beyond proofs of concept. Gartner predicted in 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, weak risk controls, rising costs, and unclear business value. Those failure modes are worth understanding before you spend anything; we break them down in why AI projects fail.

Why the savings studies and the failure studies are both right

The two literatures measure different things. The savings studies measure a person using a working tool on a task the tool is good at. The failure studies measure organizations running projects: custom builds, integrations, pilots with committees attached. A small business buying a proven $40-a-month tool for one mapped workflow is running the first experiment, not the second. MIT Project NANDA’s preliminary 2025 report, the source of the widely mangled “95%” headline, found the same split inside its own sample: externally purchased, customized tools reached deployment about 67% of the time versus about 33% for internally built ones. The report itself cautions that this is self-reported and correlation, not causation. The practical reading for an owner: your risk is not that AI cannot save time. The research says it can. Your risk is buying the wrong thing for an unmeasured workflow, which is precisely how projects join the abandonment statistics. Sequencing matters, and we lay out a full first-90-days path in the owner’s field guide to AI.

The honest payback formula

You do not need a consultant’s spreadsheet for this. Three lines:

Monthly value of hours back = hours saved per week × 4 × delegation rate ($/hr)
Monthly net                 = monthly value − monthly cost (rounded up)
Payback period (months)     = one-time setup cost ÷ monthly net

Three rules keep the math honest. First, round hours down: if you think a tool saves three hours a week, run the numbers on two. Second, round costs up: a $40 subscription becomes $50 in the model, to cover overages and the odd add-on. Third, price the hour at the delegation rate, what you would pay someone to do that task, not the owner’s billing rate. Drafting follow-up emails is a $25 to $35 an hour task even when a $150-an-hour owner is currently doing it. Count setup time as a real cost too: your hours configuring the tool, at the same rate. If the formula cannot clear zero under those rules, the tool fails the test regardless of what any vendor page claims.

A conservative worked example: one $40/month tool

This is an illustration built from public research averages, not a client result. Picture a 10-person services firm where the owner drafts proposals, follow-up emails, and meeting summaries. The St. Louis Fed’s user average is 2.2 hours a week saved; round down to 2. Price the hour at $30, a coordinator’s wage, not the owner’s rate. The tool costs $40 a month; model $50. Setup and learning take an afternoon; call it 6 hours, or $180 one-time.

Line	Conservative figure
Hours back	2/week → 8/month
Value of hours	8 × $30 = $240/month
Tool cost (rounded up)	$50/month
Monthly net	$190
One-time setup	$180 → paid back in month one
Year one	~$2,880 in time value vs ~$780 in cost

Roughly 3.7 to 1, from one tool, using rounded-down averages. Now the counter-illustration: the same tool bought on impulse for an unmapped workflow, opened twice, cancelled in month three. Total: about $150 spent, zero hours back. That is the 42% abandonment statistic operating at small-business scale. The tool did not fail; the selection did. Remember also that the $240 only becomes cash if the hours are redeployed into billable work, sales, or payroll you no longer need. Hours that dissolve into the day are worth exactly nothing.

Numbers you’ll see quoted that don’t check out

The pages ranking for AI ROI queries repeat the same figures, and most of them fall apart when you chase the citation. We traced each of these while building our fact base:

Circulating claim	What we found when we chased it
“AI returns $3.50 for every $1 invested”	Usually quoted with no source at all. The occasional attribution points to vendor-commissioned surveys of large organizations’ self-reported estimates, none of which we could verify at a primary source. No measured small-business return backs the figure.
“98% of small businesses use AI”	The U.S. Chamber’s 2024 release says 98% use AI-enabled tools, a category that includes spam filters and autocomplete. The verified generative-AI figure from the same Chamber/Teneo series is 58% in 2025, up from 40% in 2024.
“MIT proved 95% of AI pilots fail”	Triple overreach. The source is a preliminary, non-peer-reviewed report from Project NANDA (52 interviews, 153 leader surveys, 300+ public deployments), it says 95% of organizations got zero return from enterprise pilots, and its authors call the figures “directionally accurate,” not proven.
“Missed calls cost small businesses $126,000 a year”	Untraceable. It is vendor arithmetic layered on a 2016 study of 85 small businesses that modern blogs misdate as recent. No primary source publishes this number.
“85% of AI projects fail (Gartner)”	Traces to a 2018 Gartner prediction about analytics projects producing erroneous outcomes. Gartner’s actual verified figure is the 2024 prediction that at least 30% of genAI projects would be abandoned after proof of concept.
“AI agents save knowledge workers 6.4 hours a week”	Attributed to Slack across dozens of SEO pages, but it appears nowhere on slack.com or salesforce.com. Slack’s published Workforce Index reports adoption rates, not an hours-saved figure.

The vetting habit that catches most of these: find the named organization, open the primary document, and check the unit. “95% of organizations got zero return from enterprise pilots” and “95% of AI tools fail” are different claims, and only one of them was ever written down.

How to run the math on your own business

Five steps, one week, no software required. One: log an honest week of hours across the repetitive work: drafting, scheduling, invoicing, data re-entry, phone tag. Two: rank the sinks by hours, and flag the ones that are rule-based and text-heavy; that is where the Science experiment’s 40% speedup lives. Three: shortlist one proven off-the-shelf tool per sink and get real prices from the vendors’ own pricing pages; for what paid help costs at every tier, we publish full market rate tables in what an AI consultant costs. Four: run the payback formula with hours rounded down and costs rounded up. Five: buy one tool, not five, and re-measure after 30 days. If you would rather have the mapping and tool-matching done for you, that is exactly what an AI opportunity assessment is: a scored map of your hours and a short list of priced prescriptions. We sell no software and take no commissions, so the math has no thumb on the scale.

When AI ROI goes negative

The formula also tells you when to stop. ROI goes negative when setup hours balloon past the model (a tool that needs an integrator is not a $40 tool), when nobody on the team actually adopts the workflow (the Federal Reserve survey’s gap between 46% of small firms using AI and just 7% of those users fully integrating it is an adoption gap, not a technology gap), and when the “saved” hours never get redeployed into anything that earns or saves money. Gartner’s abandonment drivers, poor data quality, unclear business value, and escalating costs, all show up in miniature at small-business scale. The discipline is the same as any other purchase: a written number the tool must hit, a 30-day check against it, and a willingness to cancel. A cancelled $40 subscription is a cheap experiment. A half-adopted $15,000 custom build is how a small business ends up inside the S&P statistic.

Bottom line

There is no trustworthy average ROI for AI, but there is trustworthy math. The verified research says a person using generative AI on the right task saves real time, about 2.2 hours a week for the average user (St. Louis Fed) and 40% on writing tasks in controlled conditions (Science), while 42% of companies abandon most of their AI initiatives (S&P Global). The difference is what gets bought: projects, or proven tools matched to measured workflows. Run the three-line payback formula with hours rounded down and costs rounded up. If a $40 tool cannot pay for itself inside two months on those terms, skip it and test the next one.

Sources

Federal Reserve Bank of St. Louis, “The Impact of Generative AI on Work Productivity,” 2025: stlouisfed.org
Noy & Zhang, “Experimental evidence on the productivity effects of generative artificial intelligence,” Science, 2023: pubmed.ncbi.nlm.nih.gov
Federal Reserve Banks, 2026 Report on Employer Firms (2025 Small Business Credit Survey): fedsmallbusiness.org
Intuit QuickBooks Small Business Insights, April 2025 survey: quickbooks.intuit.com
S&P Global Market Intelligence, “Generative AI shows rapid growth but yields mixed results,” 2025: spglobal.com
McKinsey, “The State of AI: How organizations are rewiring to capture value,” 2025: mckinsey.com
BCG, “AI Adoption in 2024: 74% of Companies Struggle to Achieve and Scale Value,” 2024: bcg.com
Gartner, “Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025,” 2024: gartner.com
MIT Project NANDA, The GenAI Divide: State of AI in Business 2025 (preliminary), 2025: report PDF
U.S. Chamber of Commerce / Teneo Research, Empowering Small Business series, 2024–2025: uschamber.com