Skip to content
Guides5 min read

What we learned from 66,000 messages between people and AI agents

Keiki customer data on reply times, conversation depth, agent handoffs and speed to launch, with time windows and sample sizes.

Soft, grainy painting of a rider on a white horse leading a second horse across a marble terrace beneath towering pink clouds, framed by dark curtains and columns, a helmeted guard watching from the corner

Keiki customer data, Aug 21 to Nov 16, 2026. Every number below lists its window and sample size. Methodology is at the bottom.

It's part of our guide, Customer-Facing AI Agents: What They Are and How to Launch One.

The dataset

66,169 messages across 6,576 conversations from 76 companies. That's 12 weeks of customer traffic on Keiki across iMessage, WhatsApp, Telegram, Slack, email and API. 57,766 of those messages (3,527 conversations, 63 companies) went over a real messaging channel. (Aug 21–Nov 16)

People start the conversation

97.5–100% of conversations were started by the user, not the agent. On WhatsApp, Slack and email it was 100%; on iMessage it was 99.5% (1,335 conversations). When an agent has a phone number, people message it first. (Aug 21–Nov 16, 6,576 conversations)

Agents answer in seconds

Median agent reply time: 7.3 seconds. That's across 23,197 replies. Nine in ten arrived within 29 seconds. The median was 7.3s on WhatsApp, 7.3s on iMessage and 8.4s on Telegram. (Aug 21–Nov 16, 23,197 replies)

People reply in under a minute

Users reply to an agent in a median of 42 seconds. That's across 19,452 user replies. On WhatsApp the median was 39 seconds, and on iMessage and Telegram it was under a minute. Messaging gets the kind of attention that apps and email chase. (Aug 21–Nov 16, 19,452 replies)

Conversations keep going

72% of WhatsApp conversations continue after the agent's first answer. On customer-supplied iMessage lines it's 87%. Telegram came in at 70% (273 conversations), Slack at 60% (157) and iMessage on Keiki lines at 52% (1,290). "Continue" means the user writes again after the agent's first reply. (Aug 21–Nov 16, WhatsApp n=1,684)

Messaging threads run long

Median iMessage conversation: 7 messages (mean 22). WhatsApp: 6 (mean 12). On Telegram the median was 5 messages, and on Slack it was 4. People don't text an agent once and leave. They keep talking. (Aug 21–Nov 16, 6,576 conversations)

Agents handle 97% of conversations alone

96.8% of conversations never needed a human. That's out of 6,461 conversations with at least one user message. On iMessage, Telegram and Slack, 98.8% were handled end to end by the agent. (Aug 21–Nov 16, 6,461 conversations)

Agents hand off when people ask for a person

62% of handoffs happen because the user asked for a human or for contact details. Of 681 handoff and escalation reasons written by agents, 421 were a direct request for a person, a phone number or an office. The rest were mostly case reviews and billing questions, which go to the team when that's the right call. (Aug 26–Nov 16, 681 reasons)

Signup to first live customer conversation: under a day

Median time from creating a Keiki org to the first real user message on a live channel: 23.9 hours. The fastest quarter of teams went live within 6.1 hours. (Orgs created Aug 14–Nov 15; 62 orgs live on iMessage, WhatsApp, Telegram, Slack or email)

Photos and threaded replies mean deeper conversations

Conversations where the user sent a photo ran a median of 21.5 messages, against 4 for the rest. That compares 58 photo conversations with 5,384 without photos. Conversations where users replied inline to a specific message ran a median of 57 messages (45 conversations). Rich messaging features go with the most engaged threads. (Aug 21–Nov 16)

Customer snapshots

  • A travel-visa company (WhatsApp): 1,646 conversations and 20,095 messages in 8 weeks, with 72% continuing after the first reply. (Sep 17–Nov 14)
  • A veterinary clinic platform (iMessage): a median of 13 messages per conversation and a mean of 39, across 87 conversations. (Sep 17–Nov 9)
  • An insurance company (iMessage): 151 conversations, 69% continuing after the first reply, and 96.7% resolved without a human. (Sep 15–Nov 16)
  • A consumer app (iMessage): a mean of 127 messages per conversation, with 76% of 49 conversations continuing. (Oct 31–Nov 16)

Sources

Keiki's read-only production message store, Aug 21–Nov 16, 2026. Queries and exports are kept privately.

Methodology

Source: Keiki's production message store (read-only), Aug 21–Nov 16, 2026 (UTC).

  • Included: customer organizations.
  • Excluded: Mail-0's own consumer app, test, demo and internal organizations, playground sessions, status probes, internal rows and tool-call records.
  • Conversation: one agent and one user address.
  • Channel: taken from the address type.
  • Reply rate: share of conversations where the user wrote again after the agent's first message.
  • Response times: medians between consecutive user→agent and agent→user turns.
  • Handoff: a recorded human handoff or escalation.
  • Speed to launch: organization creation to the first inbound user message on a non-API channel.
  • Anonymization: customer snapshots are anonymized.

Published · Guides