A customer case study lives on exact words and exact numbers. So somebody on the marketing team replays the whole call to find them, usually the afternoon before it's due.
That's the slow cost. The bigger one is public. A misspelled customer name or a number the customer never said goes out under your logo, and now you're emailing a customer to apologize for the story that was supposed to make them happy.
I run Stepthrough, free hands-on AI and software training for white-collar workers, and I built a lesson on this with Transgate, a transcription tool with an AI chat over your transcript. (Disclosure: Transgate is listed in our free AI tools directory. No payment was involved, and I used my own free trial account.)
The example: you're Dana, in marketing at a made-up scheduling software company. You just interviewed Marcus, a customer at a made-up physical therapy group. The recording is about two minutes. Swap in your own.
The most useful finding: what it got right and wrong
I ran it twice. Here's the scorecard:
| Item | Result |
|---|---|
| 11 hours, 9 no-shows, 3 a week, $90 | Correct in both runs |
| The customer's surname, Oyelaran | Wrong both times: "Oialaren" |
| A "Sure." the customer said | Put on the interviewer's line, both times |
| Our company name, Lumen Desk | Written as one word |
| Hours saved | Chat refused the claim. One analysis run stated it anyway |
Numbers came through and the name didn't. That makes sense: speech to text writes what a name sounds like. No AI can know the spelling from sound alone, so get names from your own records.
Part 1: transcribe
- In Transgate, add your recording to the drop area. The file shows Ready.
- Next to Audio language, click Choose the language and pick yours.
- Tick Speaker Recognition. A case study needs the customer's words, so you need to know who said what.
- Click Start Transcribing. My 2-minute file took about 10 seconds, and each run took 2 minutes off the balance. The pricing page listed a free trial with 20 minutes when I tested it. Check the current plans.
Part 2: fix it before you do anything else
Order matters here. Fix the transcript first, so everything downstream uses the correct text. The analysis I ran afterward did use the corrected spelling.
- Click Speaker 1, then Edit speaker..., type the real name and click the blue check mark. One rename changes every line for that speaker.
- Read the transcript against the recording. Point at a line and click the pencil to correct it, then click the check mark.
- Check names first, then numbers, then company names.
Part 3: ask for what the case study needs
In Ask AI, ask for the quotes and numbers, and limit it to the transcript:
We are writing a customer case study. List the 3 strongest quotes from
Marcus, word for word. Then list every number he gave and what it measures.
Use only what is in the transcript.
"What it measures" is doing real work in that prompt. In my transcript the number 3 appeared twice: once for locations and once for no-shows a week. A list of bare numbers would have invited a mix-up.
I compared all three quotes against the transcript, and they matched word for word. Yours may differ, so compare them. Tables came back as raw text with | marks, so plan to tidy those.
Part 4: test the claim you're tempted to write
This is the step I'd never skip. Marcus said the front desk spent about 11 hours a week before switching. He never said how many hours it takes now.
Marketing copy wants to say "saved 11 hours a week." So ask the chat directly whether the transcript supports it.
In both runs, the chat said the transcript doesn't state the hours saved, only the 11 hours before. It didn't take the bait.
But the automatic AI Analysis in my first run listed "11 hrs/week saved." The second run didn't. Same recording, different wording, and one of them is the sentence that would have gone into your case study.
So your copy can say no-shows fell from about 9 a week to about 3. It can't say how many hours were saved until you ask Marcus for the after figure and do the math yourself.
The analysis can also use the wrong template
Transgate read my customer interview as an interview and added a Candidate Assessment, rating the customer 5 out of 5 stars. That template suits hiring, not customers. Don't forward the analysis as it stands.
The summary paragraph and suggested next steps were still useful. Both runs suggested getting approval to use his name, which is a real step.
The routine I'd keep
- Fix names from your own records before running anything else.
- Ask for every number with what it measures.
- Ask where the transcript supports each claim you plan to publish.
- Treat the analysis as a draft. Its wording changes between runs.
- Click Download, confirm your fixes are in the file, and get your customer's OK before you publish.
Practice it step by step
The free lesson has 27 steps in 5 parts. You click through a practice copy with a made-up recording, it checks your work as you go, and nothing touches a real account.
Top comments (0)