Percentages for AI automation are only comparable when the definition behind them is public. This is our entire method: the same rules that govern our product code, verbatim. When you see a number from Saazly, this is the method that produced it, and you can hold us to it.
The metrics we publish
Our home page shows two different metrics from the same rolling 30-day period at our reference customer Ernst P. They measure different things and use different denominators, which is why the numbers differ.
- Share of chats where the AI answers first. Of all chats that received at least one reply during the period, we count those where the first reply came from the AI rather than a human. The denominator is chats with at least one reply.
- Chats where the AI gave a complete answer. The number of chats during the period where the AI delivered at least one complete answer that passed the evidence requirement. The denominator is every chat started during the period, including those that never received a reply. A human may also have taken part in the same chat, so this metric does not mean the case was resolved without a human.
Both metrics are recalculated every night and always published with source and period next to the number.
The definition of AI-resolved
In customer reports, a conversation counts as AI-resolved only when all four conditions are met:
- The AI answered with a real answer, not a referral or an apology.
- No human needed to step in afterwards: no operator replied, no handover was requested and no assignment happened after the AI answer.
- The conversation ended: the visitor closed it, or it stayed silent for 24 hours and was closed automatically after 72.
- The customer did not reopen the case: a reopening resets the classification.
The denominator
Shares are calculated against every conversation in the period where the AI was enabled and had the chance to answer, never against a hand-picked selection. Bot-filtered traffic is excluded from every number: we classify known crawlers at the network level and only count humans.
Three levels we never mix up
- Fully automatic: the AI resolved the case on its own according to the definition above.
- AI-assisted: the AI wrote a draft that a human edited or approved before sending.
- Human: an operator handled the case.
A headline that merges the three levels into one number says nothing. We report them separately, always with period and channel.
What we never claim
- We never call a visit to an order confirmation page a verified purchase or revenue.
- We never publish a multiple without definition, denominator, channel and period.
- We never present correlation as causation: two things moving together does not prove one caused the other.
- Numbers in our AI answers to end customers require evidence in the customer's own material before they may be said.
Check us
The definition above is the same rule coded into the product, and every customer's own report shows its numbers against exactly this method. You can always see operational status on the status page, and how we handle data is described under GDPR and security.