Adoption Rate Is a Dashboard Metric

Eighteen percent of firms formally use AI in at least one business function, per the Census Bureau. I keep seeing that number cited like it settles something.

It reminds me of the years I spent watching people celebrate 99.9% uptime. Nobody ever asked what was happening during the 99.9%, or which customer was living inside the 0.1% every Tuesday morning. The number was true. It was also, operationally, worthless — an average that made everyone feel informed while hiding the only questions worth asking.

So pull the adoption number apart. Among adopters, 57% use AI in three or fewer business functions. Two-thirds describe augmentation — AI helping with work that already existed — rather than substituting for tasks or creating new ones. Nearly 7% of firms have workers using tools the company never formally adopted. Another 2.5% report formal adoption with no detectable use at the task level. "We adopted AI" and "someone is doing their job differently because of this tool" turn out to be two different facts that the headline number mashes into one.

The comfortable reading is that this is early. Integration deepens, the shallow deployments get less shallow, the gains compound. Give it a few years.

There's a less comfortable reading, and the evidence keeps getting friendlier to it: that local speedups don't travel upward, because speed was never the thing in short supply. Every automated action creates a claim on somebody's future attention. Somebody has to decide whether the output is right, whether it belongs in the system, whether the system can survive it. If that's the constraint, making the upstream task faster just means more work arriving at the same narrow gate, slightly sooner.

The experimental record on human-AI collaboration points the same way. Across 106 experiments, Vaccaro, Almaatouq, and Malone found that human-plus-AI beat the human working alone — and on average did worse than whichever of the two was already better at the task. When the person was the stronger performer, the combination beat both. When the model was stronger, adding a person made it worse. Productive combination has conditions, and they're specific, and most deployments were not designed with any of them in mind.

The macro data doesn't rescue the story either. BEA looked at industries where early adopters said they were using AI to automate labor or upgrade processes. What showed up clearly was more R&D spending. Broad productivity effects were mostly absent or inconclusive; the authors called the path from motive to outcome "murky." And in the one domain where measurement should be easiest, researchers studying coding productivity went out of their way to disclaim their own output proxy — merged pull requests — as a stand-in for delivered value. The task got faster. Whether the organization got more productive is a separate question, and almost nobody has an instrument pointed at it.

That absence matters. A timing gap closes on its own. A design gap compounds quietly while you wait for it to close. The adoption rate can't tell you which one you're in, and if you guess wrong you spend the next two years funding acceleration while the judgment capacity around it stays exactly the size it was, falling further behind every quarter.