How ClientClue AI works
Code counts. ClientClue AI reads.
Reading is what ClientClue AI is good at, so it reads. Counting is where it can slip, so code counts. That split is the reason every number in your report adds up.
Below is one real file followed through every step: the public Google Play reviews of MTN Nigeria, the first case study. MTN Nigeria is not a ClientClue customer and did not take part. Every number is what ClientClue counted, and every message is the reviewer's own words with names hidden.
Read the file
CodeCode opens the CSV or Excel file, finds the header row and reads every row as it is.
In this file
597 rows, 8 columns
Match the columns
CodeCode works out what each column holds from its name and its values. You confirm before anything is read.
In this file
8 of 8 columns matched
Clean the rows
CodeCode drops what is not a customer's own words: empty rows, repeats, staff and automatic replies, greetings on their own. One-word complaints such as "Slow" stay.
In this file
590 kept, 7 dropped
Hide personal details
CodeBefore ClientClue AI sees a single word, code replaces names, phone numbers, emails, account, card and ID numbers with a label.
In this file
2 details hidden
Dear [name] developer, I would like to be able to purchase an MTN esim and complete my KYC on-boarding through this app without having to...
Find the areas and issues
ClientClue AIClientClue AI reads a spread of messages from every week and names the areas and the issues inside them, in plain words.
In this file
9 areas, 49 issues
Read every message
ClientClue AIClientClue AI reads each message in batches of about 50 and records its language, feeling, kind, issue, whether money was taken with nothing given, any leaving words, and the short quote that says it best.
In this file
590 read in 12 batches
“mtn you guys are doing rubbish with your network even stealing people's data why this why this period”
“useless network, God punish una and ur poor network”
“Login issues as never changed... Back to [another network] 🙂”
Check every quote
CodeCode looks for each quote, word for word, in the message it came from. A quote that is not there is dropped and never shown. The marked words above passed this check.
In this file
582 of 582 quotes found word for word
Count by week
CodeCode counts every issue in every week. An issue is Rising when it has at least 5 messages and at least one and a half times its average over the 4 weeks before.
In this file
3 Rising in Week 7
Bars show messages each week; amber is the negative share. Week 1 and Week 8 are part weeks.
Write the weekly report
ClientClue AI, checked by codeCode picks the issues, the numbers and the quotes. ClientClue AI writes one plain line for each, and code checks every number in those lines against the counts.
In this file
The report for Week 7
The team received 95 messages this week with 66 negative reports. Rising issues include app crashes along with missing airtime and data balances.
Customers report the app closes unexpectedly or refuses to open when they try to purchase data.
Check this first: Inspect crash logs triggered when users tap on specific service logos on the home screen.
Users cannot view their airtime and data balances on the dashboard following the latest update.
Check this first: Verify balance display services and dashboard data fetching endpoints for recent update regressions.
Users experience extreme delays and freezing when attempting to log in or use the app.
Check this first: Review server response times and app loading performance during the authentication flow.
Ask ClientClue
ClientClue AI, checked by codeAsk a question in your own words. ClientClue AI turns it into filters, code counts the messages, and the answer cites only what was counted.
In this file
“What are customers complaining about most this week?”
Customers are complaining most about the app crashing or closing unexpectedly, with 9 messages. Another 8 messages report issues with airtime and data balances not updating. A further 8 messages mention app lagging and slow response times.
The whole file: 590 messages in 94 seconds, about 376 a minute, 0 failed batches.
Languages
It reads Pidgin, slang and one-word reviews
English
Read as written, including short and one-word messages such as Slow.No English line needed
Nigerian Pidgin
Read in its own words. The original is kept, with an English line beside it.Original words kept
Hausa
Read in its own words. The original is kept, with an English line beside it.Original words kept
Yoruba
Read in its own words. The original is kept, with an English line beside it.Original words kept
Igbo
Read in its own words. The original is kept, with an English line beside it.Original words kept
Mixed
A message that moves between English and Pidgin or another language is read as one message.Original words kept
The quote check
Every quote is checked word for word
ClientClue AI picks the short part of each message that says it best. Then code looks for those exact words in the message the quote came from, letter by letter, in any language and any script.
If the words are not there, the quote is dropped. You never see a quote that was reworded, shortened in the middle or made up.
Numbers get the same treatment. Each line ClientClue AI writes for the report is checked against the counts, and a line with a wrong number is not used.
What ClientClue AI never does
- It never invents a number. Every count comes from code
- It never sees a name, a phone number or an account number. Those are hidden first
- It never replies to your customers
- It never keeps training on your messages
- It never answers from outside your data. Ask ClientClue only uses your own messages
Measured accuracy
How well it reads each language
Each figure below is the share of messages where ClientClue AI read the feeling the same way as the people who labelled them, with the number of messages tested.
- Nigerian Pidgin400 tweets tested60%Between 55% and 65%
- Hausa200 tweets tested75%Between 69% and 80%
- Igbo200 tweets tested71%Between 64% and 76%
- Yoruba200 tweets tested82%Between 76% and 86%
The test uses NaijaSenti, tweets labelled by people, licensed CC BY 4.0. Credit: Muhammad et al., LREC 2022. See the dataset.
Tweets are not customer messages, so these figures show how well ClientClue AI reads each language, not how it will do on your file.