Code open-ended survey responses with AI that learned your codebook

You can ask a general AI to code answers, and it will invent sensible themes. But your codebook isn't a generic list of themes:

• "Too expensive" and "not worth the price" may be one code for you, and two for someone else.
• "Couldn't find the setting" may belong under Ease of use in your codebook, not under Features.
• Your trackers depend on the same answer getting the same code in every wave, or the trend line moves for reasons that have nothing to do with your customers.

Your coded waves already hold every one of those decisions. A model trained on them learns where your team draws each line, and gives the same answer to the same text every time.

It reads the answer, as the respondent wrote it, and optionally the question when one model codes several:

I switched because the app kept logging me out and support never replied

It answers with your codes, one or several, and how sure it is:

Reliability · Customer support · 0.93

• Several codes per answer: one answer often touches two or three themes. Train on answers coded with several codes, and the model gives each new answer a list of codes.
• Net codes too: train a second model on your nets, or code nets and codes as an answer with two parts.
• Many languages: the AI Classifier Model reads meaning in many languages, so a multi-country study can share one codebook.

1. Export a wave you've already coded

One row per answer, with the codes your team gave it: from your survey tool, your coding tool or a spreadsheet. If an answer has several codes, keep them in one cell separated by commas, or in separate columns. One or two waves is usually enough to start; about 20 answers per code is a good minimum.

2. Upload and check

Upload the file (Excel, CSV or JSON), or train straight from Google Sheets with the StayCharted add-on. Choose the answer column and the code column. StayCharted checks the data first:

• the same answer coded two different ways, which is usually a coder disagreement worth settling;
• codes with too few examples;
• codes whose names look like the same code;
• personal information respondents typed in, such as email addresses and phone numbers, which it can mask before training.

3. Test it the way you'll use it

Choose Test on the newest rows with the column holding each answer's date, so the model learns from earlier answers and is tested on later ones, the way it will meet your next wave. The report shows how often it chose your coders' codes, code by code.

4. Code the new wave

• In a spreadsheet: fill the new wave's file, or use the Google Sheets add-on. Every answer gets its codes and a confidence.
• From your own tools: call the model through the API as answers come in (Essentials and Business).

5. Check the unclear ones, then retrain

Sort by confidence, or use the Review Queue (Essentials and Business), and have a researcher check the answers below your cutoff. Their corrections join the next training run, so the model keeps up as your codebook changes. Add a new code by coding a few dozen answers with it and retraining.

For a new question with no coded answers yet, upload the answers alone and choose Find them for me: StayCharted groups similar answers so you can name each group as a code, split or merge groups, and review the suggestions. The reviewed answers become the first training set. On Essentials and Business.

• It doesn't write your report or summarize themes. It codes answers; you count, chart and interpret them.
• It doesn't invent codes. It uses the codes it was trained on. Low confidence helps you prioritize review, but an answer outside your codebook can still receive a confident prediction. Check a sample of confident answers too, and add codes when needed.
• It doesn't run surveys. Keep your survey tool; this works on what it exports.