How we count: why we refuse to report hours saved.

We sell AI software, which means we have every incentive to publish impressive effect figures. This page is the set of rules that stops us — written down, so you can hold us to it.

No “hours saved”

“Saves your team 30 hours a month” is the most common claim in this market and the least verifiable. It relies on self-reporting, it assumes the counterfactual (what the team would have done), and it quietly ignores the new work AI creates: reviewing, correcting, redoing. Nobody audits it, which is precisely why it spreads. We do not print it — about ourselves or anyone else.

Counts, not percentages

“41 of 77 articles needed corrections” survives scrutiny; “53% error rate” invites it. A count carries its own sample size, keeps its date, and cannot be silently rebased. Where we have a number, you get the count — with the denominator and the date.

Ten observations before we publish

Our own effect figure — what the gate catches, how often, at what cost — is not on this site yet. Not because it looks bad, but because it has not reached ten observations. When it does, we publish it as a count with its interval, and we keep publishing it when the next ten make it look worse.

Why not outcome-based pricing?

If you can measure the effect, why not charge for it? Because outcome billing has been tried across this industry and it historically collapses: the definition of “outcome” gets negotiated, attribution gets contested, and within a few quarters both sides retreat to something countable. We price per brand, fixed to what you run, and keep the measurement separate from the invoice — measurement you pay for is measurement you should not trust.

Three labels on every claim

• Measured — we have the count, the date and the source. Quote it.
• Widely advised, thinly evidenced — the industry agrees, the data is soft. We say so.
• Folklore — travels from deck to deck without a primary source. We name it folklore, even when it flatters us.

And our own interest

When a source sells the thing it is measuring, we flag it. That includes us: everything on this page is written by a company that benefits when you believe governed AI works. Which is exactly why the rules above exist — check them against the dossier and the benchmark, where they are applied.