A third of our “talked” calls were machines. What that did to every number downstream.

The short answer

In our audit of AI calls placed for a travel company between 19 and 22 Sep 2026, 226 of the 739 pre-sales calls logged as “talked” had no person on them. The only words from the lead's side were a voicemail greeting or a call screener asking the caller to stay on the line. Another 53 calls were mixed, and only 460 calls, 62%, were real conversations. Because a call counted as talked whenever its transcript held one line from the lead's side, every machine inflated the pick-up rate, added a callback nobody had asked for and reached the CRM as a conversation. The fix is to decide what a machine sounds like, count those calls as no answer, and keep them out of the CRM and the callback queue. If you run AI or dialler calls, check how your own system defines a conversation before you trust any pick-up rate it shows you.

How did we count a “talked” call?

We audited every outbound call placed for one customer, a travel company that sells group trips from Meta lead forms, between 19 and 22 Sep 2026. At the time the rule was simple: a call counted as talked when its transcript had at least one line from the lead's side. That one rule fed three things. It set the pick-up rate on the dashboard, it decided the grade each call received, and it decided which stage the lead moved to in the customer's CRM.

The audit read the lead-side lines of every talked pre-sales call and matched them against phrases that only machines say. Voicemail greetings were the obvious group. Call screeners were the less obvious one: the automated assistant that answers for the person and says something like “thanks, please stay on the line” or “when you have finished recording”. If every lead-side line on a call was one of those phrases, the call was marked as a machine. If some lines were machine phrases and some were not, it was marked as mixed.

How many “talked” calls were really machines?

Almost a third of what the dashboard called conversations were recordings. The shares above are our own division of the counts in the audit; the counts themselves come straight from the call log.

What did the machine calls do to the numbers downstream?

Every figure that started from “talked” was wrong in the same direction. Pick-up was inflated by about a third. The grades were worse. 183 of the 357 calls graded as follow-up were machines, which left roughly 69 real follow-ups where a person had named a time to call back. 85 of the 176 calls graded as hung up were machines too, so real hang-ups were 91, not 176.

The CRM carried the error forward. The machine calls were sent to it as conversations, so those leads sat in the follow-up stage with a next-call time that no person had asked for. Then the retry rule did what it was told: it called the voicemail box back at “its” time. A machine that gets answered and talked to also costs money, because the voice engine is billed for the minutes it spends talking to a recording.

Why do voicemail greetings and call screeners slip through?

Most voicemail greetings were already caught. The calling bridge hung up on them before the agent said much. Call screeners are harder, and the reason is deliberate. A screener asks the caller to stay on the line, and a real person often picks up a few seconds later. So the agent answered the screener once and waited. When a person came on, that was the right call. When nobody did, the call still had a line from the lead's side in its transcript, and under the old rule it counted as talked.

That is the general trap. Any system that defines a connected call by “the line opened” or “something was said” will count machines as people. The more your leads use call screening, the bigger the error.

What did we change in Revvox?

A call whose lead-side lines are all machine phrases now counts as no answer. It does not reach your CRM as a conversation, it does not get a follow-up grade, and it does not trigger a callback at a time nobody chose. The lead stays in the normal attempt schedule instead, which is capped at two attempts inside the calling window.

What are the limits of this audit?

• One customer, one industry and four days of calls. This is not a general voicemail rate.
• The machine check was a first pass, matched on phrases. A real person who says “please stay on the line” could be misclassified. A second pass using a fresh transcript and the call's length would be tighter.
• The counts cover pre-sales qualifying calls only, not the later closing calls.
• We report the counts as the audit found them and have not rounded them up.

How can your team check its own calls for voicemail?

• Write down what your system counts as a connected or talked call. If it is “the line was answered” or “any speech was heard”, assume machines are in the number.
• Pull 50 recent calls marked as connected and read or listen to the lead's side only. Tally how many are greetings, screeners or silence.
• List the phrases you heard. Greetings and screeners repeat across operators, so a short list catches most of them.
• Count machine-only calls as no answer, and remove any callback they created.
• Recompute pick-up, hang-ups and follow-ups after the correction. Decide staffing and retry rules from the corrected numbers, not the dashboard's first answer.
• If a vendor dashboard reports “connected” or “answered”, ask whether that means a person spoke or only that the line opened.

Where can you read the rest of the audit?

The same audit produced six other rules for calling paid leads, from the hour you call to how you open the call. Each has the number behind it.

• Seven rules for calling paid leads, from 2,739 calls
• How many times to call a lead: the attempt-by-attempt data
• How Revvox qualifies paid leads on the call