Peter Cappelli sat in his Wharton office, reading through the Ricoh deployment files. Ricoh had done what the AI pitch told it to do: hired consultants, built the tools, launched. The first cost report came back at three times the price of human labor. Cappelli's verdict was blunt. The industry sold what was possible and ignored what was practical.
That gap, between the demo and the dollar, is now visible in three federal and institutional data sets. The evidence suggests something most people find uncomfortable. The problem with AI adoption in small business is not speed. It is the aim.
The Platitude
"AI will transform your business." You have heard it from the vendor running the demo. You have heard it from the peer who just signed a new tool contract. You have heard it at the conference where the speaker showed a slide deck full of promise and no profit-and-loss sheet behind it.
It sounds right. It is not.
The Mechanism
The research is clear on this. Cappelli, a management professor at Wharton, spent months studying how firms deploy AI versus how they describe deploying it. The Harris Poll put a number on the gap in early 2025. CEOs estimated that 35% of their own AI efforts amounted to what the poll called "AI washing." Adoption for optics. Not for output. The admission came from the people signing the checks.
The Federal Reserve's 2026 Small Business Credit Survey made the same pattern visible from the ground. The Fed surveyed more than 6,500 small employer firms across the country. Of those using AI, 83% pointed it at writing and marketing. Just seven percent had fully woven AI into how the business runs. Nearly half were still in the experiment stage, testing tools with no clear path to measured return.
Goldman Sachs surveyed its 10,000 Small Businesses network in early 2026 and found the same shape from a second angle. Ninety-three percent of AI users said the impact was positive. Only 14% had put AI into core operations. The gap between reported satisfaction and actual integration was enormous. The U.S. Chamber of Commerce added a third line: less than a quarter of the firms it surveyed used AI for tasks that drive revenue. Supply chain work, lead finding, product development. The work that touches the money stayed manual.
Three institutions. One finding. The tool went to the most visible layer, not the most valuable one.
The Structural Flaw
The problem is not that firms adopted AI. The problem is that content-layer work replaced process-layer work. A blog post is easy to demo. A broken invoicing cycle is not. The vendor shows the easy win. The buyer calls it progress. The back office stays the same.
Most people treat this as a knowledge problem. It is a system flaw. The deployment order is backwards, and no amount of better content fixes a leaking process.
The Price Tag
Ricoh is the case that puts a dollar figure on this pattern. Cappelli walked through the numbers in a Fortune piece earlier this year. Ricoh set out to use AI for a specific operational task. It brought in outside consultants. The consulting fees alone ran close to $500,000.
When the system launched, it worked. It also cost three times what human labor cost for the same task. After months of tuning, the ongoing AI fees settled at roughly $200,000 a month. That number was higher than the total payroll for the task it replaced. Headcount went from 44 to 39. Five positions. Not a transformation. A line item that moved in the wrong direction and kept moving.
Goldman Sachs economist Ronnie Walker sealed the wider view. His team wrote that they still found "no meaningful relationship between productivity and AI adoption at the economy-wide level." The case data and the macro data land in the same place. The spend is real. The return, at scale, is not showing up.
The Replacement
Deploy AI against rule-based process bottlenecks first. Content comes second. Once that order is clear, three moves follow from it.
Move 1: Audit the Time Sink
Spend one week tracking every task that follows the same steps each time it runs. Invoice prep. Appointment scheduling. Lead follow-up sequences. Data entry from one tool to the next. Write the list in one place. Note how many hours each task burns per week.
This is where most operators stop, because the list feels too small to matter. It is not small. It is the layer where margin leaks live, and it demands honesty about where your hours actually go.
Move 2: Run One Tool Against One Task
Pick the item on the list with the most hours. Find one AI tool built for that task. Run it for 30 days. Track the cost of the tool against what you were paying in labor or your own time priced at your billing rate. If the tool costs more than the manual way, cut it. If it costs less, keep it and measure again at 60 days.
Move 3: Hold Content Until Process Pays
Do not point AI at writing, social posts, or marketing until the process-layer tool has a measured return. Content is where every vendor starts the demo because it is the easiest win to show. That is the vendor's interest. It is not yours. A polished newsletter draft is worth nothing if the pipeline feeding it has no measured conversion rate.
What Becomes Visible
Running this for 90 days does something the platitude never did. It shows where the real friction sits. Which tasks are rule-based enough to hand off. Which "AI results" were polished drafts with no revenue trace. Where the actual margin leaks hid behind busy work. And whether the tool earned back its cost or just looked like it did.
The Check
At the end of 90 days, ask three things.
→ What moved the number?
→ What looked like progress but left no trace in revenue?
→ What friction showed up more than once?
That is the difference between advice that sounds right and a system that proves itself.
Where You Stand
Cappelli warned against listening too closely to the firms selling the tools. "They're telling you what's possible," he told Fortune. "They're not thinking about what is practical."
If the AI spend sits on content and the process layer is still manual, what the business has is not a strategy. It is a subscription.
