What small businesses are actually doing about the squeeze, and where it goes wrong

Part 2 of a three-part series on the pressures facing small businesses in 2026. Part 1 examined the squeeze and its root causes. Part 3 lays out how to do it right, step by step.

The money has started to move.

A quarter of small business owners told the National Federation of Independent Business in July that they planned significant capital outlays in the second half of 2026, the highest share since late 2024. Gartner expects global software spending to rise 14.7 percent this year, to 1.47 trillion dollars, with artificial intelligence features now built into nearly every renewal. After two years in which Main Street mostly watched, it is now buying.

The question is what it is buying, whether it works, and why so much of it will not. The evidence on all three is unusually clear, and it does not match the marketing.

Where the money is going

The spending is not spread evenly. It concentrates in four places where the labor squeeze bites hardest.

Customer service is the proving ground. Owners face rising expectations for instant, around-the-clock support at the same time that support roles are hard to fill and expensive to keep. Autonomous agents now resolve, or “contain,” 40 to 65 percent of routine inquiries such as order status, return policies, and scheduling, without a human touching them. The economics are stark: an AI resolution on the leading platforms is priced around 50 cents, against 1.50 to 2.50 dollars in labor for a human agent handling the same interaction.

Sales prospecting is second. The cost of acquiring a customer is a survival metric for a small business, and the manual work behind it, researching accounts, drafting outreach, watching for buying signals, is exactly the kind of work that no longer justifies a junior hire. HubSpot now prices its prospecting agent at one dollar per recommended lead. Salesforce charges for its agents by the action, roughly ten cents each.

Finance is moving from looking backward to looking ahead. Bookkeeping platforms have automated invoice processing and transaction categorization for a while. The newer shift is forecasting: tools that read historical transactions against seasonality and warn an owner of a cash crunch weeks before it arrives, in time to adjust purchasing or draw on a credit line.

And employees are already using it, everywhere. The U.S. Chamber of Commerce Foundation’s 2026 Main Street AI Monitor found that half of all workers at small businesses already use AI on the job. Ninety percent of those use it for writing and editing, 88 percent for research, and 86 percent for technical work. Most are not automating themselves out of a job; 64 percent use it for personal productivity, and 59 percent reinvest the saved time into better output.

Software is no longer priced by the seat. It is increasingly priced by the outcome, and that changes how an owner should judge whether it is worth buying.

That last shift deserves attention on its own. Microsoft still sells its small-business Copilot the old way, at roughly 18 to 21 dollars per user per month, a predictable cost tied to headcount that relies on the human to extract the value. Salesforce charges by consumption, which lowers the entry price but makes the bill swing with volume. HubSpot’s move in April 2026 to charge only for a resolved conversation or a qualified lead turns software from a fixed overhead into a variable cost tied directly to a result. For an owner, this is good news and a trap at once. The alignment is real. So is the difficulty of forecasting a bill that rises with a successful month.

What the evidence says works

The strongest research on the question comes from a study by MIT and Stanford economists of more than 5,100 customer support agents at a Fortune 500 software company. Agents given AI-generated guidance resolved 13.8 to 14 percent more customer problems per hour.

The more important finding was who benefited. The gains went disproportionately to the least experienced workers. Agents with two months on the job, using AI, matched the output of colleagues with six months of experience who did not. Turnover fell, particularly among newer staff, and customer sentiment improved. For a small business that cannot win bidding wars for experienced talent, this is the single most useful fact in the literature: the technology is a capability equalizer. It lets a business hire the people it can find and close the gap faster than it could through experience alone.

Where it goes wrong

Against that evidence stands a number that should give every owner pause. Roughly 90 to 95 percent of enterprise AI initiatives never turn a profit, never accelerate revenue, or never make it out of the pilot stage.

The failures are not, for the most part, technological. They are the same three mistakes, made over and over.

The problem was never defined. Companies adopt AI in order to have AI, rather than to fix a specific, high-friction workflow. Leaders describe a desired outcome in language a technical team misreads, and the project drifts. The implementations that succeed start with one narrow, measurable use case and a number agreed in advance.

The data was not ready. As many as 85 percent of AI projects fail on poor, fragmented, or siloed data. A model is only a reflection of what it is fed, and in most small businesses the relevant data lives in disconnected systems and spreadsheets that do not agree with one another. No tool fixes that from the outside.

The spending went to the wrong place. The pattern behind successful adoption is a specific split: about 10 percent of the effort on the algorithm, 20 percent on technology and data infrastructure, and 70 percent on people and process. Organizations that invert that ratio, buying the software first and thinking about the workflow later, consistently fail.

There is also a dip that catches the well-prepared. Research from MIT Sloan documents a “productivity J-curve”: firms adopting AI typically see output fall by about 1.33 percentage points in the short term, because integrating it demands workflow redesign, training, and data cleanup before the gains appear. The businesses that push through the dip outperform their peers over a four-year horizon. The ones that judge the investment at month three abandon it just before it starts to pay.

The traps outside and inside

Two further hazards have less to do with execution and more with the market itself.

The first is being sold something that is not what it claims to be. “AI washing,” the practice of labeling ordinary rule-based software as artificial intelligence, has become common enough that the Federal Trade Commission launched a coordinated enforcement sweep, Operation AI Comply. It ordered the company behind DoNotPay, marketed as “the world’s first robot lawyer,” to pay 193,000 dollars over claims it never tested. It has pursued a string of “AI-powered” business-opportunity schemes that sold small business owners automated storefronts and delivered nothing. The consequences run past wasted money: a business that uses a vendor’s tool to generate fake reviews or misleading claims can itself become the target of enforcement.

Fraud has grown alongside it. In 2025 the FTC reported more than 8 billion dollars lost by consumers and small businesses to investment scams, a 38 percent increase in a year, much of it driven by AI-generated video and voice impersonating trusted figures.

The second hazard is already inside the building. The Chamber Foundation found that 19 percent of AI adoption in small businesses happens bottom-up, through employees using tools on their own with no oversight. When a well-meaning employee pastes a customer list or a financial statement into a free consumer chatbot to draft an email, that data can leave the company for good. For a business handling health, payment, or European customer data, it is also a compliance event. The instinct to ban the tools tends to backfire; it simply drives the use further out of sight.

The human variable

Underneath all of it sits a gap the surveys keep surfacing. A 2026 study by Harvard Business Review Analytic Services found that 76 percent of small and mid-sized businesses expect to increase their use of AI in the next twelve months, while only 19 percent feel prepared to recruit or develop the skills to manage it.

That gap matters because the technology is not reliable enough to run unattended. Stanford’s 2026 AI Index notes that even the most advanced models still fabricate facts, and their accuracy falls on hard cases. Someone has to check the work. Tellingly, 52 percent of business leaders now rank deep industry experience above technical AI experience when hiring, and 70 percent say the technology is raising the value of distinctly human judgment. An employee who does not understand the business cannot tell when the machine is wrong.

Read together, the evidence describes a clear pattern. The businesses getting value are not the ones that bought the most. They picked one expensive, repetitive problem, cleaned up the data underneath it, ran a short trial with a number attached and a person checking the output, and only then expanded. The ones losing money did the reverse.

How to do that deliberately, step by step, and how to vet a vendor before signing, is the subject of the next piece.

Sources: NFIB Small Business Economic Trends (July 2026); Gartner; U.S. Chamber of Commerce Foundation, 2026 Main Street AI Monitor; Brynjolfsson, Li, and Raymond (MIT and Stanford); MIT Sloan; Federal Trade Commission; Harvard Business Review Analytic Services with TriNet (2026); Stanford Institute for Human-Centered AI, 2026 AI Index; published pricing from Microsoft, Salesforce, and HubSpot.