AI for Southwest Michigan Small Businesses: What Actually Works

If you own a business in Kalamazoo, Battle Creek, St. Joseph or anywhere across the seven counties of Southwest Michigan, you have almost certainly been pitched an AI product in the last six months. Maybe a dozen. The pitches share a structure: a dramatic claim about hours saved, a monthly subscription, and very little detail about what happens in week three.

It’s worth separating the marketing from the measured evidence, because the measured evidence is actually pretty useful. It says three things at once: adoption among small firms has stalled, most formal AI projects produce nothing, and a narrow set of everyday uses reliably saves people real time. Those findings aren’t in tension. They describe the same problem from different angles.

The adoption picture is not what the ads suggest

The U.S. Census Bureau runs a biweekly, nationally representative survey called the Business Trends and Outlook Survey that asks firms whether they used AI in any business function in the past two weeks. It is the closest thing we have to a real adoption number, and it is far lower than the figures circulating in vendor decks.

Between December 2025 and May 2026, Census found that AI use rose among firms with at least 20 employees but did not change significantly among firms with fewer than 20. That is the number that should get a small owner’s attention. It isn’t that small businesses are behind on a curve everyone is climbing. It’s that the curve is bending upward for larger firms and sitting flat for the smallest ones. The gap is widening.

Sector matters too. Census recorded 39.7% adoption in Information and 33.9% in Finance and Insurance, but only about 14% in Retail Trade. If you run a storefront on the Kalamazoo Mall or a shop in downtown Three Rivers, most of your direct competitors are not using this yet. That is a genuine opening, and it will not stay open forever.

Why most AI projects produce nothing

Here is the uncomfortable counterweight. A 2025 study out of MIT’s Project NANDA, The GenAI Divide: State of AI in Business, found that roughly 95% of generative AI pilots fail — meaning they never produce a return the business can point to. (Forbes has a clear summary of the findings.)

The researchers’ explanation is the useful part. The failing projects weren’t failing on model quality. They were failing on integration. Companies bought generic tools that demoed beautifully, dropped them next to an existing workflow without changing the workflow, and got nothing that showed up in the accounts. The minority that succeeded picked one high-value process, accepted the friction of redesigning it, and built in a way for the tool to learn from corrections.

The failure mode is almost never “the AI wasn’t smart enough.” It’s “nobody changed how the work actually gets done.”

The same study noted something that will ring true to anyone who manages people: in over 90% of the firms surveyed, employees were already using personal AI tools on their own, whether or not the official project worked. Your team is probably ahead of your policy. That’s a governance question worth getting in front of, and we’ll come back to it.

Where the returns are actually documented

Set aside the enterprise transformation story and look at task-level evidence, which is much stronger.

In March 2026, the Census Bureau’s Household Trends and Outlook Pulse Survey asked workers about AI use across 11 specific job tasks. About 55% said they had used AI for at least one. Among those who used it in the prior week, a quarter said it saved them less than an hour, 31% saved one to two hours, 15% saved three to four hours, and 15% saved more than four. Ten percent said it saved no time at all, and 3% said it actually cost them time. Read the whole distribution, not just the top of it: the typical gain is modest, and roughly one in eight people got nothing or worse.

The tasks people actually used it for are unglamorous and directly transferable to a small business:

There’s one more finding that matters more for small employers than for anyone else. A landmark study of 5,179 customer support agents by economists at Stanford and MIT, published by the National Bureau of Economic Research and later in the Quarterly Journal of Economics, found that giving agents an AI assistant raised issues resolved per hour by 14% on average — but the gain was 34% for novice and lower-skilled workers and close to zero for the most experienced ones.

Read that in the context of a nine-person shop in Portage. The tool did not make your best person better. It moved your newest person up the learning curve faster. If your constraint is that a new hire takes six months to answer customer questions the way your veteran does, that finding is aimed squarely at you.

A low-risk way to start

The pattern in the research points to a specific method: one process, one owner, a real before-number, and a fixed date to decide.

Three guardrails worth setting now

1. Decide what never goes into a chatbot

Write one page, today, listing what employees may not paste into an AI tool: customer payment details, employee health or HR records, anything under a signed NDA, patient information if you touch healthcare. Given how many teams are already using personal AI accounts unofficially, a policy written after an incident is a policy written too late. One page is enough to start — and it works far better when your team helps write it, because the people doing the pasting know where the gray areas actually are.

2. Treat vendor claims as claims, not facts

The Federal Trade Commission has been explicit that, in its words, there is no AI exemption from the laws on the books. Through an enforcement sweep it calls Operation AI Comply, the agency has taken action against companies making unsupported AI claims, including DoNotPay, whose “robot lawyer” was marketed in part as a way for small businesses to check their websites for legal violations. The FTC’s order in that case bars the company from claiming its product can substitute for a professional service without evidence to back it up — a reasonable standard to hold your own vendors to. If one won’t quantify a savings claim for a business your size, that’s your answer.

3. Read the terms on data reuse

Check whether your customer data can be used to train the vendor’s models, and whether that can change later. The FTC has warned that quietly amending terms of service to permit new data uses may itself be unfair or deceptive. As a rule of thumb, paid business and enterprise tiers from the major providers exclude customer data from model training while free consumer tiers frequently do not — but confirm it in writing for the specific plan you’re buying rather than assuming.

The bottom line

The honest summary of the current evidence is narrower than the hype and more encouraging than the backlash. Broad AI transformation projects mostly fail. Specific, repetitive, language-heavy tasks — drafting, summarizing, looking things up, getting a new employee to competence — produce measurable time savings for most people who try them, and the largest gains go to the least experienced workers on your team.

For a small business in Southwest Michigan, that translates to a modest and testable move: pick one task, measure it, spend under $50 a month, and give yourself thirty days to find out. Meanwhile, the firms with 250 employees are not waiting.