Your numbers
Free to share and embed. Inputs live only in the URL; nothing is tracked or gated.
Methodology, fully shown
The calculator prices two scenarios for the same yearly send volume (monthly sends × 12). The unreviewed scenario assumes some percentage of AI-drafted sends go out with a material error: a wrong amount, a wrong recipient, a wrong commitment. Each error costs fixing time (default 45 minutes at your hourly rate; finding an error after it reached a customer means the apology email, the correction, the credit note, not just an edit). A share of errors (default 20%) also carries direct financial impact: the underbilled invoice nobody notices, the disputed charge, the discount honored because the wrong price was quoted. That impact is priced as a fraction (default 30%) of the average send value, not the whole value, because most financial errors cost part of a transaction, not all of it.
The reviewed scenario prices the alternative honestly: every send read by a person before it goes, at a default of 20 seconds per send, valued at the same hourly rate. Twenty seconds assumes the draft arrives with its source data cited, the way an approval queue presents it; reviewing against evidence is fast, re-deriving the answer is not. If your review reality is slower, raise the slider and let the math say what it says.
What the defaults deliberately do NOT include, which means the unreviewed total is understated: customer-trust damage that outlasts the incident (the wholesale buyer who now re-checks every invoice), compounding errors that trigger further errors downstream, and the tail risk of a single catastrophic send, the documented incidents where one unreviewed action became a tribunal ruling or a viral screenshot. Tail risk resists averaging; we left it out rather than invent a number, and the result is a floor, not an estimate of the worst case.
There is no universal "AI error rate" and this page does not claim one; the default of 2% is a conservative placeholder you should override with your own observed rate. The honest structural point survives any input you choose: error costs scale with volume and value, review costs scale only with volume, and the gap between them widens exactly where your business grows. When the calculator tells you review is not worth it, believe it; that is what the copilot vs autopilot sorting framework is for.
Common questions
It is a conservative, adjustable default, not a measured universal constant; no honest one exists, because error rates vary enormously by task, model, and how much context the AI has. Published evaluations of LLMs on real-world business tasks routinely show single-digit error rates even on strong models. Set the slider to your own experience; the point of the calculator is that even 1% is expensive at volume.
The math is symmetrical: it prices the review time honestly too (default 20 seconds per send at your hourly rate), and if you set the error rate low enough, review costs more than errors. That crossover is the honest answer to "should everything be reviewed": no, only the sends where the math and the stakes say so.
A send that is wrong in a way that matters: wrong amount, wrong recipient, wrong commitment, wrong tone at a sensitive moment. Typos that change nothing are not material. The costly-share input separates errors that only cost fixing time from those with direct financial impact.
Yes, freely, with the embed snippet below the results. No email, no license, no tracking of your inputs; the URL parameters exist only so results can be shared.
The review step, without the review cost
flo.space prepares each send with its evidence attached, so reviewing takes the seconds this calculator assumes. Connect your tools read only and time yourself.