The Cost of Almost-Right

There is a piece of intuition that runs through almost every small-business conversation about AI, and it is wrong in a way that costs real money.

The intuition: if the AI gets it right ninety percent of the time, that's ninety percent of the value. We'll just check the work, fix the ten percent, and we're ahead. It sounds reasonable. It sounds, in fact, like a sober and pragmatic way to think about AI quality. It is also, in most real workflows, dead wrong.

The cost of almost-right is not a discount on the value of right. In many cases it is a tax. Sometimes it is a tax larger than the value of getting it right in the first place.

The mental model behind ninety percent is ninety percent of the value assumes that the cost of finding and fixing the bad ten percent is small. It assumes you can distinguish good from bad quickly. It assumes the bad outputs are obviously bad — typos, missing fields, glaring errors.

In practice, none of these assumptions hold for current AI systems. Bad outputs from a competent model are confident. They look right. They use the right vocabulary, the right structure, the right tone. Finding them requires the reviewer to know the answer already, which raises the question of why the AI was being used in the first place.

And the time cost of reviewing every output to find the ten percent that's wrong is often equal to — or larger than — the time saved by having the AI do the work in the first place. The math, when you do it honestly, frequently lands at zero or worse.

I'll give you a concrete example because the abstract version doesn't land.

Imagine an AI that reads incoming customer emails and drafts replies. It gets ninety percent right. The team's plan is to review every draft before sending. Sounds reasonable.

What actually happens: the reviewer reads the email, then reads the AI draft. They mostly agree with the draft. They send it. They do this for a hundred emails in a day. Somewhere in those hundred, the AI has hallucinated a refund policy, misread a delivery date, or confidently agreed to something the company doesn't do. The reviewer was on email number sixty-three. They were tired. They sent it.

The cost of that one email — a chargeback, a complaint, a customer churn, a reputational moment on social media — can easily exceed the time-saving from the other ninety-nine. The system appears to be ninety percent right and is, in business outcome terms, net negative because the cost of the wrong ten percent is asymmetric and concentrated.

There is a second cost of almost-right that is even less visible, and that affects the entire business once it sets in: the erosion of standards.

When the team is reviewing AI output rapidly, the bar for acceptable starts to drift. The first week, the team rejects anything that's not perfect. The fourth week, they accept things they would have rejected on day one — because rejecting them means re-drafting, and re-drafting is the thing they were trying to avoid by using AI in the first place. The standard, quietly, drops to match the AI's output rather than the AI being pushed up to match the standard.

This is the most expensive consequence and the one that nobody puts on the ROI spreadsheet, because by the time it shows up — six months later, in customer feedback or revenue — the link back to the AI rollout has gone invisible.

The fix is not to abandon AI. It is to be clear-eyed about which workflows tolerate almost-right and which do not.

Workflows where almost-right is acceptable: anything where the output is a draft, the human reviewer is fresh, the volume is low enough that real review happens, and the cost of a bad output slipping through is low and recoverable.

Workflows where almost-right is catastrophic: anything customer-facing without a real review step, anything financial, anything legal, anything where the cost of a single wrong output exceeds the time saved across a thousand right ones. For these, you do not want a ninety-percent system. You want a system that is right or refuses to answer, and you build the second behaviour deliberately.

The first question to ask of any AI deployment is therefore not how good is it? but what is the cost of one wrong output in this workflow, and what is the cost of refusing to answer? Until you have answered that, accuracy benchmarks tell you almost nothing about whether the system will create or destroy value.