AI in Diary Studies: Three Ways to Eliminate Compromises for Full Automation
A team of four processed data from 32 respondents in a week-long field study without traditional compromises: all voice messages were transcribed, dialogue with each respondent was deepened, and all diaries were analyzed. AI handled routine processing, monitoring, and scoring, saving 403 hours—equivalent to 2–3 additional staff members. The project was completed 5 days ahead of schedule while maintaining analysis quality.
Time Savings: Manual vs. AI Processing Comparison
AI integration drastically reduced labor at key stages. Here’s a breakdown by task:
| Task | Manual (hours) | With AI (hours) | Savings (hours) |
|------|----------------|-----------------|-----------------|
| Monitoring 32 respondents (7 days) | 224 | ~2 | ~225 |
| Diary status (4–5 runs) | 72 | 1.5 | ~72 |
| Candidate scoring (208 surveys + 60 detailed) | 37 | 0.25 | ~37 |
| Respondent profiles (24 people) | 48 | 12 | ~36 |
| Interview guides (24 unique) | 19.5 | 2 | ~17.5 |
| Source of truth (quota audit, feature map) | 14 | - | ~14 |
| Quota control | 1.7 | 0.017 | ~1.7 |
| Total | ~403 | | ~403 |
This arithmetic allowed analysis of all 32 diaries instead of the standard 12–15, uncovering additional hypotheses from the data.
Quality Arguments: Why AI Did Not Compromise Analysis
- Consistent Classification: AI applied a unified taxonomy of insights without drift across all 32 diaries. Manual analysis suffers from researcher fatigue after the 20th case.
- Mutual Verification: AI detected 4 errors in recruiter quotas; humans corrected 20% of AI detections (80% accuracy).
- Verification via Raw Data: Each insight card was linked to the original entry with fact and interpretation checks.
- Depth of Analysis: 4 out of 10 hypotheses emerged from the data; 300+ insight cards and 20 query clusters—double the standard.
- Full Coverage: Buffer respondents provided the richest diaries.
Mechanism 1: Telegram Bot for Automated Data Collection
A manager with no programming experience created a bot in a day using Python + FastAPI with Claude Code, deploying it to a cloud server. The bot processes messages from 32 chats in real-time via webhook, saving to Google Drive:
- Text—in Markdown.
- Photos—as is.
- Voice/video—with automatic transcription alongside the original.
Key technical decisions:
- Webhook Instead of Polling: Telegram API doesn’t provide chat history; failure risks minimized (incident on day 6 due to OAuth—fixed in 20 min).
- OAuth with Refresh Token: Service account couldn’t see “Computers” in Drive; OAuth enabled access to synced files.
- Markdown Format: After 3 iterations (JSON → text → Markdown), convenient for both humans and AI.
- Privacy Mode Off: Disabled in BotFather to read all messages.
Freed-up time was directed to clarifying questions: “Why was this time different?”, “What was inconvenient?” Respondents answered by voice—automatic transcription added nuances, slips, and emotions.
A trip tracking table with taxonomy (barriers, drivers, unawareness, issues) allowed filtering gaps and generating targeted questions in real-time.
Mechanism 2: Tracking Table and Scoring with AI
AI built a dynamic trip table: detecting new entries vs. responses to questions, organizing by taxonomy with context and influence strength. The researcher filtered by respondent, saw coverage areas, and suggested questions.
Scoring 208 surveys + 60 detailed took 15 minutes instead of 37 hours. AI classified cities and quotas; humans verified.
Profiles of 24 respondents—12 hours instead of 48: AI aggregated data by template.
Interview guides (24 unique)—2 hours: AI generated based on diaries and business goals.
Mechanism 3: Insight Clustering and Audit
300+ insight cards clustered into 20 functional query groups. Each linked to raw sources.
Quota control—1 minute: AI cross-checked with the source of truth (quota audit, feature map).
Key Takeaways
- AI saved 403 hours, enabling analysis of all diaries and deeper dialogue.
- Consistency and mutual verification ensured quality on par with manual analysis.
- Automating routine tasks freed up resources for data-driven hypotheses.
- Voice messages became an advantage thanks to transcription.
- The pipeline is scalable: from bot to insight clustering.
— Editorial Team
No comments yet.