Derek Waldron runs data and analytics for JPMorgan Chase, a bank that spends $18 billion a year on tech and runs more than 450 AI use cases. In late 2025, he sat in front of CNBC cameras and named a problem most executives will not say out loud: there is a gap between what the tools can do and what the business actually captures from them. If JPMorgan cannot close that gap, a 15-person firm running five AI subscriptions has a problem it has not looked at yet.
The SBE Council's 2026 Tech Use Survey puts the headline number at 82%. That is the share of small business employers now using AI tools. The median firm runs five. The number reads like proof of progress. It is not. It is a count of subscriptions.
The Numbers Behind the Number
McKinsey surveyed nearly 2,000 organizations in mid-2025. The research is clear on this. Eighty-eight percent said they use AI. Only 7% had scaled it across their business. Roughly 6% qualified as high performers, the firms that could point to more than 5% of operating income coming from AI. The adoption number and the performance number are not the same number. They are not close.
Pacific AI brought the lens tighter. Their 2025 survey looked at how firms watch what their AI tools produce. Among small companies, 9% monitor production AI for accuracy, drift, or misuse.
Nine percent. The other 91% run tools that no one checks.
Stanford's 2025 AI Index filled in what happens next. AI incidents rose 56.4% in one year. Researchers logged 233 in 2024, a record. When no one watches the tools, the tools still produce output. Some of that output is wrong. Some carries risk the owner finds out about late.
The Structural Flaw
That is not an adoption problem. It is a measurement problem.
The advice to "adopt or get left behind" skipped the step that makes adoption count. Tool count went up. Oversight did not. Pacific AI found that pressure to ship fast is the top barrier to oversight for 45% of firms. Speed replaced the work of knowing what the tools do.
Most people treat this as a tech problem. It is a system flaw. McKinsey found that the small group of high performers are 2.8 times more likely to have rebuilt their workflows than everyone else. Fifty-five percent of that group changed how work moves through the business. Only 20% of the rest did. The gap was not the tools. It was the structure around the tools.
The Replacement Principle
Measure what you run before you add what you do not. Once that is clear, three moves follow from it.
Move 1: Run a 30-Day Tool Audit
List every AI tool your business pays for. Next to each one, write the output it produces and the last time someone checked that output for accuracy. If the answer is "never," mark it. If the tool has no clear tie to revenue or cost savings, mark that too.
This is where most owners find out they are paying for tools no one uses and tools no one checks. That finding alone changes the math on what "adoption" has cost.
Move 2: Write a One-Page AI Policy
It does not need to be long. State what data can go into an AI tool and what cannot. Name who reviews AI output before it reaches a client. Set a rule for how often someone spot-checks what the tools produce. Most small firms skip this step. The firms that do it catch errors before those errors cost money.
Move 3: Pick One Function and Measure for 90 Days
The pattern that works, drawn from firms that moved past the testing phase, is to start in one area. Pick the function where AI has the clearest output. Set a number to track. Run it for 90 days.
Then compare the result to what that function produced before. Firms that try to roll out AI across every function at once stall. The ones that start small and measure first scale what works and cut what does not.
What the System Shows
Running this for 90 days does something the advice never did:
You find out which tools produce a return you can put a number on. You find out which are dead weight, still pulling a monthly fee with no one using them. You find out which carry risk no one has looked at, tools that touch client data with no rules for how that data gets handled. And you find out whether your team knows how to check what the tools produce or just trusts the output by default.
The Feedback Loop
At the end of 90 days, ask three things:
→ What moved the number?
→ What looked like progress but left no trace?
→ 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
Waldron named the value gap from inside the most AI-forward bank on the planet. The gap does not shrink because you own more tools. It shrinks because someone watches what the tools produce, measures what they cost, and cuts what does not earn.
You are not behind because you own too few subscriptions. You are behind if no one is watching the ones you already run.
