B2B International cuts interview quality checks from an hour to five minutes with HappyScribe

B2B International’s business-to-business market research depends on interviews run across dozens of languages and every one of those interviews needs to meet high quality and compliance standards. The problem was scale: a single audio could take up to an hour to review, making quality and compliance a time-consuming part of the interview workflow. Then they tested HappyScribe's AI Chat on a live project.

Checking interviews one language at a time

For Simon Brett, Senior Quality & Compliance Manager at B2B International, the old process was slow and partial by design. English interviews meant listening to the audio end to end. Other languages would inflate costs, as external resources and QC training had to be accounted for. Some languages, including French, German and Brazilian Portuguese were translated to English using HappyScribe and following along as best he could.

"Reading the translation while listening to it was the only way that I could check some of these interviews," Simon said. The team brought in native speakers, "but we don't always have the resource to do that in every language."

The bigger constraint was coverage across a project. The Interviewer Quality Control Scheme [IQCS] and Market Research Society [MRS] standards call for checking at least 10% of interviews; B2B International always checks more. "I'd only ever listen to between 10 and 15% of the total number of interviews," Simon said. "So if they completed 100 interviews, I'd only ever listened to 10 or 15 of them." Some languages were effectively difficult for him to monitor. Team ‘catch-ups’ 1-2-1 chats and time consuming data checks were employed.

And the gaps had a cost.

When an interviewer accidentally skipped a question that should have been asked or the survey question routing failed, the team had to investigate why and sometimes re-contact the respondent. "To maintain quality and ensure all questions were asked, the rising cost to fieldwork can be significant, so spotting issues in any language early on is paramount." Simon explained.The pressure was clear. Catch issues early or pay for them later.

Turning a transcript into a structured QC report

To test the approach, Simon performed initial testing on a 5-minute project. After reassuring himself on how the AI works and its effectiveness in English, he then worked on a live project that had just gone into field. He ran HappyScribe's AI Chat across roughly 30 to 40 of its interviews. While still following IQCS and MRS standards in monitoring, HappyScribe was used on all languages equally, so there was no bias towards English being monitored more than say Japanese or Arabic.

The method was a single, carefully written prompt. He uploaded the approved question set as a PDF, then asked the AI to list every question the interviewer actually asked, compare each one to the approved question set, and mark it as a perfect match, a partial match, or no match. The output came back as a table including the recording name, the question asked, the match status, any discrepancies, and a timestamp pointing straight to the moment in the audio.

"The AI is only as good as the prompt you use," Simon said. The timestamp was a later addition. "I thought it'd be really good if I could get this to take me directly to the question in the audio for further checks. So I just added it, and it did. It was perfect."

The results held up against his own manual checks, and then some. "It probably picked up more than I would listen to it," Simon said. On one question, the research team had substituted a single word, "established" for "stable," across a batch of interviews. "In all six audios, the AI actually spotted a single word had been replaced with another" and flagged it for investigation. It turned out the interviewers were using a newer, approved version of the question set. The AI had caught a deliberate change before it was communicated to the quality team.

Simon was clear-eyed about the limits. "AI is a tool rather than a replacement," he said, and "I'd rather it flag something than not flag it." “You do spend time checking AI flags, as it turned out to be much stricter than a human reviewer, but the time was well worth it and the flags were a benefit rather than a nuisance.

Compliance checks that go the extra mile

One thing surprised him. The AI model recognized the legal nature of the content and began checking the informed consent steps at the front and back of each interview, treating them as a legal script. It verified that interviewers read out the privacy policy, gave their name, and got a clear yes or no on consent, and it listed each step with its own timestamp, exactly as it did for the questions.

"It checks the compliance step by step without me actually asking it. It seems to recognize that it's a legal script," Simon said. "No idea why it knows how to do that. Didn't ask it to, but it does."

From a 15% sample to almost every interview

The time saved is the headline. A single audio that took 45 to 60 minutes to review manually now takes about five minutes for HappyScribe’s AI to analyze, a reduction of roughly 90%. Simon can queue 10 or more at once and work on other compliance tasks while it processes in the background.

That speed changes what is possible. Instead of a 10 to 15% sample, Simon can now run a quick check of almost every interview submitted. "It allowed me to check far more interviews than I would normally," he said, "instead of just checking around 10 or 15%, which is what we normally do." It also opened up the languages he found difficulty monitoring before. "It allowed me to pick up with more interviewers and give them some really good feedback and guidance" across more languages than before.

The freed-up time went to other impactful and collaborative work. In one week alone, Simon was able to pull together security and certification material that helped a colleague win a new contract, the kind of task he says he "probably wouldn't have had time to do" before.

"It's allowed me to tighten up quality control across all of the languages, and really get a much more consistent and accurately presented questionnaire to the respondents," Simon said. "It's a great tool that helps speed up the whole QC process, where I can actually feel more confident about the whole project delivery."

Results in a nutshell

Review time cut from up to an hour to five minutes: A single interview audio that took 45 to 60 minutes to check manually is now analyzed by the AI in about five minutes, a reduction of roughly 90%, with batches of 10 or more processed at once.

Coverage jumped from a 10 to 15% sample to almost every interview: Quality checks now span the full project surpassing the IQCS and MRS recommended standards of 10% monitoring.

Quality control extended to all languages: All languages can now be monitored in-house and checked, providing better data consistency.

Caught discrepancies a human reviewer would miss: The AI flagged a single-word change across all six audios in a batch and surfaced partial matches stricter than a manual review would.

Compliance checked automatically: Informed consent steps, privacy policy, interviewer name, and recorded consent are verified step by step and timestamped without being prompted.

More time for impactful and collaborative work: With routine QC handled, the compliance team can dig deeper into client queries and support new business, including material that helped win a recent contract.