AI-Assisted Research: What Works and What Doesn't

The State of AI in Research (2026)

AI capabilities in user research have advanced significantly. 88% of researchers now identify AI-assisted analysis as a major development. But adoption is uneven, and results are mixed.

Here's what actually works—and what doesn't.

What AI Does Well

1. Transcription and Initial Processing

AI excels at:

• Transcribing interviews with high accuracy
• Speaker identification and labeling
• Timestamp generation
• Initial cleanup of filler words

Benefit: Researchers spend zero time on transcription, which used to take 3-4x the interview length.

2. Summarization and Synthesis

AI can:

• Summarize long transcripts into key points
• Extract quotes organized by theme
• Generate initial interview summaries
• Compare across multiple interviews

Benefit: First-pass analysis that used to take hours happens in minutes.

3. Pattern Identification

AI helps with:

• Clustering similar feedback
• Identifying frequently mentioned topics
• Detecting sentiment patterns
• Flagging unusual responses

Benefit: Patterns emerge from large datasets that humans would miss.

4. Question Generation

AI can:

• Suggest follow-up questions based on discussion guides
• Generate variations of interview questions
• Adapt questions based on previous responses
• Create survey questions from research objectives

Benefit: Researchers focus on research design, not question wording.

What AI Doesn't Do Well (Yet)

1. Understanding Context

AI struggles with:

• Reading between the lines
• Understanding organizational politics mentioned implicitly
• Recognizing when customers don't know what they need
• Interpreting body language and tone

Human needed: Researchers catch what customers mean, not just what they say.

2. Making Judgment Calls

AI can't determine:

• Whether feedback represents a real pattern or vocal minority
• How feedback connects to business strategy
• Which insights are actionable vs. interesting
• When to probe deeper vs. move on

Human needed: Research judgment about significance and next steps.

3. Building Rapport

AI can't:

• Make participants feel comfortable
• Adapt conversation flow to participant energy
• Create trust that surfaces honest feedback
• Navigate sensitive topics with empathy

Human needed: The human connection that makes research valuable.

4. Detecting Its Own Errors

AI doesn't know:

• When it's hallucinated a conclusion
• When the training data biases its analysis
• When the pattern it found is spurious
• When the summary missed the most important point

Human needed: Quality control and verification.

The Right Model: AI-Assisted, Human-Led

The most effective approach:

AI handles:

• Transcription
• Initial organization and tagging
• First-pass summarization
• Pattern detection at scale
• Routine categorization

Humans handle:

• Research design
• Interview conduct
• Interpretation and judgment
• Connection to strategy
• Quality verification
• Final synthesis

The handoff: AI provides a starting point. Humans refine, correct, and contextualize.

Practical Implementation

For interview research:

1. AI transcribes → Human reviews for accuracy
2. AI summarizes → Human corrects and adds context
3. AI tags themes → Human validates and adjusts
4. AI clusters insights → Human interprets patterns
5. Human writes final synthesis with AI-assisted components

For feedback analysis:

1. AI categorizes incoming feedback → Human spot-checks
2. AI detects anomalies → Human investigates
3. AI generates reports → Human edits and contextualizes
4. AI tracks trends → Human interprets significance

Tools and Approaches

Transcription:

• Otter, Rev, Descript (good accuracy, requires review)

Analysis assistance:

• Dovetail, Condens (research-specific AI)
• General LLMs with custom prompts (flexible but requires more work)

Pattern detection:

• Text analytics platforms
• Custom clustering pipelines

Key criteria:

• Can you verify AI conclusions?
• Does AI show its sources?
• Can you correct AI mistakes easily?
• Does AI integrate with your workflow?

Risks and Mitigations

Risk: Over-reliance on AI conclusions Mitigation: Always verify with source data. Never cite AI summary without checking original.

Risk: Bias amplification Mitigation: Diverse research participants. Review AI outputs for bias patterns.

Risk: Missing nuance Mitigation: Human review of all final outputs. Don't ship AI-only analysis.

Risk: Privacy concerns Mitigation: Understand where data goes. Use privacy-compliant tools. Anonymize sensitive content.

The Future (What's Coming)

Emerging capabilities:

• Real-time analysis during interviews
• Suggested follow-up questions mid-conversation
• Automatic connection to prior research
• Predictive pattern detection (what's likely to emerge)

Human researchers won't be replaced. But researchers who use AI will outperform those who don't.

---