We use the number "~90%" a lot, and we think it deserves a precise definition instead of being a marketing round number. Here's exactly what we mean, and — just as importantly — what we don't mean.
What We Measure
For every AI-drafted test case, we track how much of it survives engineer review unchanged: correct steps, correct test data, correct assertions, and correctly identified edge cases. Across our engagements to date, that figure lands around 90% — meaning roughly one in ten elements of a typical test case needs a human correction or addition before it ships.
What "Accuracy" Doesn't Mean
It doesn't mean 90% of test runs pass, and it doesn't mean AI decides pass/fail on your production releases unsupervised. It specifically describes how much of the authoring work — the drafting — AI completes correctly on the first pass, in a workflow where a human still reviews every case before it's part of your suite.
Where the Other 10% Goes
Typically it's one of three things: an ambiguous business rule the AI interpreted reasonably but incorrectly, a compliance-sensitive path we deliberately require a human sign-off on regardless of AI confidence, or a genuinely new edge case with no historical pattern for the AI to draw from.
Why This Matters for Trust
Salesforce runs real business processes — quotes, contracts, approvals. We don't think "fully autonomous AI testing, no review" is a responsible claim for systems like that, and we'd rather be precise about where the human checkpoint sits than oversell the number.