Enterprise Precision Threshold

The line

Robert F. Smith, chief executive of Vista Equity Partners, states the reliability gap between consumer and enterprise AI with an unusually concrete number: "in the world of consumer AI, 93% precision is okay. You and I can say, oh, let's go out to dinner, let's find a Thai restaurant...93% of the time it'll find the restaurant and we're fine and it'll book it at the time. 7% of the time we're going to show up at a restaurant that actually has a Thai dish right on the menu. And we might be okay with that in a consumer environment. That does not work in banking. It does not work in insurance. It does not work in automotive."¹

The number itself is illustrative, not measured. No benchmark, no task definition, and no source attaches to 93%; it is a rhetorical figure and should be read as one, since what carries the argument is its structure rather than its digits. "Precision" is used loosely too: in its technical sense it trades against recall, and an agentic workflow has no single such number at all, only a task success rate, an error distribution, and a cost-weighted expectation of what the errors cost.

Three things make this more than a truism. The failure mode in the example is not an error, it is a near-miss: the 7% case does not crash or invent a nonexistent restaurant, it returns something defensibly adjacent, a place that happens to serve one Thai dish. That is the characteristic failure of probabilistic systems in production, and it is the one that survives casual testing, because a demo asks whether the system did something sensible rather than whether it did the right thing. The threshold is set by the cost of the tail rather than the average case: consumer tolerance is high because the failure costs a mildly annoying dinner, while the three named industries are ones where a single wrong output is expensive, regulated, or both. And the gap itself is the product: closing the last several points of precision through constrained workflows, verification steps, deterministic guardrails, and human checkpoints is expensive, general, and portable across companies, which is the actual argument for doing it at scale rather than solving it separately ninety times over.

Why it matters

The usual story about slow enterprise AI adoption is that large organizations are conservative and bureaucratic. This account says they are correct: the same model that delights a consumer is genuinely unshippable in a regulated workflow, and the missing work is real engineering, not risk aversion. By Eric Glyman's own account, his early screen agent at Ramp ran at 60 to 90% task success and he deliberately held it back from unsupervised release in a financial product, which is the same judgment call made two years earlier by a founder rather than a private-equity owner, and it is independent corroboration of the same threshold from a different seat.

The banker's version, and the architectural consequence

Brian Moynihan gives the same threshold from inside a bank, and adds a refinement Smith's framing leaves out: the bar is not one number per industry, it is one per surface.² Inside a single consumer bank, a trivia answer and an account balance sit at opposite ends: getting whether one historical event predates another wrong is a mistake, but getting a customer's balance wrong is not survivable. On a record-keeping surface the acceptable error rate is not 99%, it is zero, and the failure is not proportional to the size of the error: a small discrepancy in a balance is not a small failure. That is why Bank of America shipped its assistant Erica in 2018 on a small, deterministic model rather than waiting for a better generative one: the design answer to a zero-tolerance surface is not a more precise probabilistic system, it is keeping the probabilistic component out of the causal path of the fact entirely. Above some threshold, in other words, the right move is a category change rather than an accuracy improvement, and the reason the threshold binds at all is liability rather than user patience: whoever deploys the tool answers for the outcome regardless of which tool it was.