%202.png)
How KLA ran WhatsApp AI interviews for a major South African bank's board presentation, set them beside human qualitative interviews, and couldn't tell them apart.
How KLA ran WhatsApp AI interviews for a major South African bank's board presentation, set them beside human qualitative interviews, and couldn't tell them apart.
In short
KLA, a South African research agency, ran 250 WhatsApp AI interviews on Yazi for one of the country's major banks in August and September 2025, beside 50 human qualitative interviews and about 10,000 coded customer comments. The themes in the AI interviews matched the human ones, and every AI interview gave KLA a usable verbatim for the bank's board presentation.
KLA is a South African research agency and one of Yazi's longest-standing clients. In August 2025 one of the country's major banks asked KLA to show how its customers felt about their banking experience, in time for a presentation to the bank's board.
KLA designed three strands: about 10,000 existing customer comments coded for themes, 50 qualitative interviews, and a new strand of AI interviews on WhatsApp with the bank's customers, run on Yazi. The target for the WhatsApp strand was 150 to 200 interviews.
The bank was presenting to its board, so fieldwork couldn't slip.
The bank's customer list couldn't be passed around by email or FTP.
The fear with WhatsApp interviews is short, SMS-style answers.
AI interview findings had to hold up next to human interviews on the same topics.
KLA's lead consultant uploaded the bank's customer list straight into Yazi, so the data never travelled by email. Yazi set up the WhatsApp invitation and the AI Interviewer, which asked each customer about their experience and followed up on what they said, by text or voice. Invitations went out in batches, and KLA watched interviews start within minutes of the first send.
Customers responded warmly. Several replied to thank the bank for selecting them.
Did the AI interviews tell the same story as the human ones?
Illustrative screen showing how the conversation worked. Not a real participant. One customer sent a single voice note four and a half minutes long.
Three strands, three kinds of evidence
How many responses each strand produced, and how much each response said
The WhatsApp strand reached 250 completed AI interviews, drawing on the bank's customers and KLA's own panel. KLA's lead consultant found "one or two nuggets per person", with answers far longer than the SMS-style replies he had feared. One customer sent a single voice note four and a half minutes long.
About 1 in 3 invited customers started the interview
Two batches sent a few days into fieldwork, as read out at the time
KLA then set the AI transcripts next to the human interviews on the same topics. The themes matched, which gave the team confidence in the method's validity. KLA shared the transcripts with the bank's senior team, and the bank was happy with the result.
“The fear with a WhatsApp interview is that you'll get short, SMS-style answers, but people expressed themselves well.”
“Side by side, what was being said was quite similar. So in terms of validity, it's quite powerful.”
How the work grew
“I'd like them to start thinking about it more, because it gives us a bit of an edge in how we do research.”
“I want our proposal team to be there, because I want them to be throwing it into their proposals.”
| Human depth interviews | AI interviews on WhatsApp | |
|---|---|---|
| Sample | 50 interviews | 250 interviews |
| Customer data | Lists shared by email or FTP | Uploaded straight into the platform |
| Answer depth | Depth from a skilled interviewer | Matching themes and verbatims, by text or voice |
| Scheduling | Booked in working hours | Customers reply when it suits them |
| Analysis | Recordings to transcribe | Transcripts ready as interviews finish |
Evidence from a live client project that AI interviews hold up against human qualitative work. KLA has since trained its teams on Yazi and started building it into its proposals.
Customer voices from 250 conversations, ready for the board, alongside the coded comments and human interviews.
From the 250 people that responded, I could get out 250 quality verbatims. If I put them side by side and asked you which one is qual and which one is AI, you wouldn't be able to tell the difference.
Agencies asking whether AI interviews count as real qual rarely get to test it on a live project. KLA did, with a board deadline and a bank's customers. The AI interviews gave it the reach of a survey with themes that matched its human interviews, and KLA has kept using them since.
In KLA's study for a South African bank, the themes in 250 AI interviews matched 50 human interviews on the same topics. KLA's senior consultant said that side by side, you wouldn't be able to tell which was which. KLA used the AI interviews beside its human interviews, not instead of them.
Longer than many expect. KLA's lead consultant found one or two nuggets per person rather than short, SMS-style replies, and one customer sent a single voice note four and a half minutes long.
KLA uploaded the bank's customer list straight into Yazi, so the data never travelled by email or FTP.
Want to test AI interviews against your own qual?
Book a 30-minute demo →