AI-enabled Qualitative Research: Faster Analysis

Treat artifacts as first-class qualitative units because Anthropic finds 93% of conversations produce artifacts and many of those are explanations (17%) or documents (15%).

For qualitative research, codebooks and thematic coding should include artifact type as a primary field so teams can compare not just topics but deliverable forms across user cohorts.

Use cadence to triangulate intent because hourly patterns sharpen interpretation of motivations, for example news at 7 a.m., recipes at 6 p.m., and sleep advice around 5 a.m.

When you see a spike in a cluster, link timestamps to qualitative notes or open-text survey answers to capture context around the observed behavior.

Link behavior to perception at the respondent level because Anthropic shows people who delegate more are more optimistic about pay and job prospects.

Your qualitative survey should capture both usage logs and open-ended rationale to explain why delegation aligns with optimism or worry.

Report and adjust for sample bias because the survey respondents skew toward Computer & Mathematical (≈30%) and Management (23%).

Always report respondent composition and weight or stratify analyses when generalizing beyond heavy users.

Ingest transcripts and survey CSVs together to preserve timestamps and session IDs and run combined thematic and frequency analyses so artifacts, cadences, and survey responses are linkable in one project.

Solution in Evidano: upload files; preserve timestamps and session IDs; and run a combined thematic and frequency analysis so artifacts, cadences, and survey responses are linkable in one project.

Train or import a classifier to standardize artifact labels so you can compare artifact distributions across cohorts.

Solution in Evidano: import Anthropic-style codebook or train a custom classifier, then run AI-assisted coding to label artifact types and purposes and export disagreements for quick human review.

Run cross-segment analyses and hierarchical code trees to surface frequency tables, co-occurrence networks, and side-by-side quote buckets for decision-makers.

Solution in Evidano: run cross-segment analyses and hierarchical code trees; export frequency tables, co-occurrence networks, and quote buckets per segment.

Filter by timestamp to generate hourly heatmaps and to tag event windows so you can surface representative quotes and artifacts from those windows, for example the April 14 tax spike.

Solution in Evidano: filter by local timestamp to build hourly/daily heatmaps and to tag event windows and surface representative quotes and artifacts from those windows.

Use end-to-end encryption, PII redaction, and data-use guarantees when working with sensitive qualitative research to match privacy-preserving approaches.

Evidano encrypts data, supports PII redaction, and does not train third-party models on customer data, making it suitable for sensitive qualitative research similar to Anthropic’s privacy-preserving approach.

Normalize by session counts and present both proportions and bootstrapped confidence intervals to compare segments reliably.

A: Use Evidano’s cross-segment analysis to normalize by session counts and to present both proportions and bootstrapped confidence intervals for key themes.

Import a codebook or train a classifier on a seed set and then hand-audit edge cases to reproduce Anthropic’s artifact taxonomy.

A: Yes, import a codebook (or the report’s Appendix), train a classifier on a seed set, then apply and hand-audit edge cases.

Use E2E encryption, PII redaction, and guarantees that data is not used to train external models for sensitive research.

A: Evidano offers end-to-end encryption, PII redaction, and explicit guarantees that your data is never used to train external models, appropriate for research needing privacy-preserving linkage similar to Anthropic’s methods.

The core insights are that hourly-sampled conversations linked to a survey reveal temporal patterns, artifact prevalence, and shifting expectations about AI.

Anthropic reports that 93% of conversations produce artifacts, a linked survey of ~9, 700 respondents shows rising expectations about AI, and event-driven spikes such as the 8× tax spike on April 14, 2026, illustrate strong temporal effects.

Researchers can reproduce artifact and cadence analyses by ingesting transcripts and survey rows with timestamps and respondent IDs, running classifiers, and filtering by time windows.

Follow the seven-step workflow: collect transcripts and survey CSVs, ingest to Evidano, auto-label artifacts and purposes, run thematic coding, build hourly/daily heatmaps, compare segments, and generate report visuals.

Necessary data types include chat transcripts, metadata with timestamps and surface, classifier labels, and linked survey rows with respondent-level identifiers.

Anthropic used chat transcripts, metadata, classifier labels for purpose and artifact type, and a privacy-preserving linked survey to enable behavior-to-attitude mapping.

Teams should report respondent composition and apply weighting or stratification when generalizing beyond heavy users to handle sample bias.

The Anthropic survey skews toward Computer & Mathematical (≈30%) and Management (23%), so report composition and apply weights or stratify analyses when necessary.