AI Will Diffuse Slowly, and Reliability Is Why

The most common mistake in forecasting AI's effect on the economy is to confuse capability with adoption. Models that can do a task are not the same as businesses that let them, and the gap between the two is not ignorance, regulation, or fear of change. It is reliability. For AI to be usable in a business process, it has to be more reliable than the human it replaces at that task, not merely as reliable, and that is a higher bar than most people building the technology have noticed. I expect diffusion to be slow for that reason, and I think the slowness is rational.

Why the bar is above human

A business does not evaluate a worker's error rate in isolation; it evaluates the error rate together with everything that surrounds the error. A person who makes a mistake can be asked what happened, can be retrained, can be held responsible, and can notice on their own that something looks wrong before the mistake leaves the building. A model that makes the same mistake offers none of that, so the same error rate carries more cost. Add the asymmetry of trust, in which one machine error is remembered longer than ten human ones, and the liability question of who exactly is accountable when the software was wrong, and the practical bar for deployment lands well above human performance. Parity is not enough. The machine has to be good enough that the surrounding apparatus of supervision can be removed, because if it cannot be removed, the economics rarely work.

Reliability compounds

There is an arithmetic to this that a demo hides. A step that is right ninety-five percent of the time looks excellent in a single interaction and is nearly useless in a workflow. Chain ten such steps and the whole is right about six times in ten; chain twenty and it works about one time in three. Real business processes are long chains, and most of the value in automating them sits at the end of the chain, where all the earlier errors have accumulated. This is why the impressive demo and the abandoned pilot are so often the same product a few months apart, and why the honest question for any AI system is not how good it is at a task but how many tasks it can string together before a human has to look.

A worked example

Take a form with twenty fields that a model reads so a person does not have to type them. A model that gets each field right ninety-nine percent of the time sounds excellent and produces a fully correct form only about eighty-two percent of the time, which means nearly one form in five still needs a human to find the error, and finding one wrong field among twenty takes about as long as typing the form. A careful clerk runs at perhaps ninety-nine and a half percent per field and delivers a clean form roughly ninety percent of the time. To remove the clerk, the model needs something like ninety-nine point nine percent per field, which yields a clean form ninety-eight percent of the time. The bar is not human parity. It is a tenth of a percent per field better than a good human, which is a very different engineering target from "as good as a person."

The two ways of being wrong are not equal, and this is where most threshold-setting goes astray. A field the model leaves blank is a recall failure, and it is cheap: the person sees the gap and fills it. A field the model fills confidently with the wrong value is a precision failure, and it is expensive, because it looks finished and sails through to a filing, a loan decision, or a ledger. So the threshold that decides whether a document goes through untouched is set on precision, and set by the cost of the mistake on the other side of it, while recall governs how much of the typing the model saves. A system can be well short of perfect on recall and still remove the reviewer; it cannot be short on precision by much at all. Reliability, in practice, means being wrong in the cheap direction.

What it looks like from the inside

My own experience running an AI-native data platform has taught me the shape of the curve. We put multimodal document understanding into production for financial documents, and the interesting work was never the model. It was the threshold: the line, set by evaluation rather than by enthusiasm, above which a document goes through without a person and below which it does not. We automate exactly where the expected cost of a wrong answer falls below the friction removed, and nowhere else. Over time the share of documents needing a human review fell by an order of magnitude, not because the model became brilliant overnight but because a weekly habit of examining its failures and repairing their causes moved the threshold a little at a time. That is what diffusion actually looks like from inside a company: not a switch but a threshold that creeps, task by task, as reliability is earned.

The same pattern held when we replaced a general-purpose model with a smaller one built for a single extraction job. The specialized model was faster and cheaper, which is the usual story, but the reason it shipped was that it was more consistent. Businesses will trade a great deal of capability for predictability, and the vendors who understand that will be the ones whose products survive their pilots.

What follows

If reliability is the constraint, then evaluation is the product, and the companies that treat their evals as a core asset rather than a testing chore will diffuse first. Narrow beats general in the near term, because a narrow system can clear the human bar on one task while a general one clears it on none. And the economically meaningful milestone is not a benchmark score but the day a business removes a reviewer from a process and does not put them back.

I hold this view loosely in one respect. It may be that the bar sits above human only while accountability for machine error remains unsettled, and that once the institutions adjust, as they eventually did for automated trading, parity will do. But institutions adjust on their own clock, which is another way of saying that even then diffusion will take the decade I expect it to, and that the useful work in the meantime is earning reliability one threshold at a time.