The AI landscape doesn't move in one direction — it lurches. Some techniques leap from experiment to table stakes in a single quarter; others stall against regulatory walls, technical ceilings, or organisational inertia that no amount of hype can dislodge. Knowing which is which is the hard part. The State of Play cuts through the noise with a rigorously maintained index of AI techniques across every major business domain — classified by maturity, evidenced by real-world adoption, and updated daily so you always know where you stand relative to the field. Stop guessing. Start knowing.
A daily newsletter distilling the past two weeks of movement in a domain or two — delivered to your inbox while the index updates in the background.
Each dot marks the weighted maturity of practices within a domain — hover for a brief summary, click for more detail
AI that synthesises user research transcripts, survey data, and customer feedback into themes and feature signals. Includes automated affinity mapping and sentiment-driven feature prioritisation; distinct from product analytics which analyses behavioural data rather than qualitative feedback.
AI-powered synthesis of qualitative user feedback -- interviews, surveys, open-ended responses -- is a proven capability with mature tooling and documented enterprise ROI. The question is no longer whether it works but why it has stalled. Over half of UX researchers now use AI for synthesis, and vendor-commissioned studies report ROI figures from 236% to 665%. Yet adoption remains confined to large enterprises with established research operations, and the category shows clear signs of a maturity plateau. The binding constraints are organisational, not technical: integration complexity favours vendor-led deployments over internal builds, practitioners hide AI tool use from colleagues even while reporting productivity gains, and hallucination risks demand human oversight that erodes the speed advantage. Consensus has settled on AI as an efficiency multiplier for theme extraction and summarisation -- compressing weeks of affinity mapping into hours -- rather than a replacement for interpretive research judgment. The result is a good-practice capability whose rollout challenge is less about tooling maturity and more about embedding AI-assisted synthesis into research workflows that organisations actually trust.
Three platforms dominate the vendor landscape -- UserTesting, Dovetail, and Thematic -- each with GA AI features and named enterprise customers including Amazon, Canva, Meta, and Mayo Clinic. UserTesting's latest Forrester TEI study documents 665% ROI with measurable business outcomes: 60% conversion improvement and 140% lift in customer spend. Dovetail has pushed furthest on workflow integration, with 3.0 (Fall 2025) and April 2026 releases shipping AI Agents, Dashboards, and Figma integration that embeds synthesised insights directly into design workflows; quantified outcomes show product managers reducing workload from 100 to 10 hours per week and teams saving 38+ hours weekly through AI-native synthesis. In April 2026, UserTesting launched its Figma plugin (GA), auto-generating test plans and embedding analysis results directly into design tools with named early-adopter outcomes (CarMax refined form flows; AJ Bell streamlined workflows before development). Thematic continues demonstrating strong enterprise adoption with documented 92% time reduction in feedback analysis and $4.8M incremental revenue generation. Outset, a new entrant, launched a visual intelligence suite (April 2026) extending synthesis beyond text to multi-modal analysis (facial cues, physical interaction) and integrated with Dovetail to send AI-moderated interview transcripts directly into synthesis platforms. A peer-reviewed deployment (Muse, ACL 2026) validates that LLM-based synthesis achieves inter-rater reliability κ=0.71 with human researchers on structured theme identification. Market analysis projects the text analytics segment reaching $18 billion by 2028, driven by demand for AI-powered open-ended response analysis.
Beneath the vendor momentum, adoption data tells a more complicated story. Adoption among researchers has stabilized at 69% using AI in at least some synthesis projects (April 2026 survey), though 83% limit it to summarisation rather than deeper thematic analysis. Ecosystem analysis reveals a market-wide shift toward AI-first platforms (Looppanel, Condens, Marvin, Marvin) characterized by auto-tagging and low-friction workflows, with AI-assisted synthesis now table-stakes vendor capability rather than differentiator (April 2026 industry report). Yet a critical risk surfaces in deployment: 47% of enterprise AI users have made major business decisions based on hallucinated synthesis content, with documented examples of entire features built on fabricated user preference findings. Practitioner assessment identifies that AI works reliably for structured synthesis tasks (transcription, ideation, desk research) but requires professional review for complex synthesis (summarization often introduces bias, incorrect details, fabricated metrics) and fails on interpretation and pattern discovery. Independent rigor assessment (Nielsen Norman Group, March 2026) reveals critical limitations: AI tools hallucinate findings, fail to identify meaningful patterns in qualitative data, and cannot adequately consider nuanced research questions; they excel at semantic pattern finding in pre-coded datasets but cannot replace trained human researchers. Comprehensive hallucination benchmarking shows Gemini 2.0-Flash at 0.7% on summarization tasks but 18.7% on legal questions and 15.6% on medical queries; Claude models range from 4.4% to 10.1%; newer reasoning models paradoxically show higher hallucination rates (o3 at 33% on person-based questions). Peer-reviewed evidence from healthcare research shows LLMs perform similarly to humans on deductive coding with predefined codebooks but fail on inductive theme generation and hallucinate themes without evidence support. MIT research has documented systematic bias against lower-literacy and non-Western users, with refusal rates reaching 11% -- a direct threat to synthesis accuracy for diverse feedback pools. Heavy AI users report three times more hallucinations than casual users and require ten times longer verification time. Expert practitioners face unique risks: fluency trust and velocity pressure enable hallucinations in specialist workflows even among teams with stated AI literacy. New governance frameworks (April 2026 business school guidance) emphasize source verification, construct validity checks, and enterprise-grade data security as prerequisites for responsible synthesis deployment, positioning synthesis as "prediction not verification."
These reliability concerns reinforce the organisational friction that defines the category's stall. Critical research (April 2026) documents that AI-based research methodologies fail at adoption scale: systems designed from users' stated preferences achieve only 57.7% accuracy, underperforming naive baselines, and deployment variance is extreme (bottom-quartile teams reach 12-18% daily active users vs. top-quartile 82-88% within 90 days). The differentiator is understanding the problem before building—a research design issue, not a technology one. Vendor-led implementations succeed at roughly twice the rate of internal builds, and a 42-day average project cycle suggests the bottleneck is process and research methodology, not processing power. Emerging hallucination mitigation strategies--multi-model orchestration where frontier LLMs cross-examine each other for quality--show promise but add deployment complexity. Practitioners increasingly argue the real gap is not categorisation -- where AI hits about 80% accuracy -- but prioritisation and action, the stages where human judgment remains irreplaceable. Industry consensus has shifted from AI-as-replacement toward responsible AI augmentation: vendors explicitly position AI as effective for accelerating interpretation and synthesis automation while humans own meaning, impact, and decisions. This maturation signals the category has settled into a sustainable but bounded equilibrium: proven value for large enterprises with mature research operations and research discipline, persistent structural barriers preventing expansion to mid-market segments rooted in adoption methodology and organizational readiness rather than tooling capability, and training incentives within foundational models that reward confident guessing over truthfulness—challenges that tooling improvements alone cannot resolve.
— Critical risk assessment: qualitative synthesis especially vulnerable to hallucinations because research summaries lack verification numbers; 66% of employees trust LLM outputs without verification; documents persistent adoption barrier rooted in quality validation requirements.
— Independent analysis of AI research transformation: synthesis time collapsed from 16-26 days to 4 hours, async AI moderation dominates (80% of new studies), sample size scaling 50-100x, with Greenbook GRIT tracking AI as #1 emerging method for 3rd consecutive year.
— Production adoption analysis of 412+ enterprises shows 68% running AI interviews in production (up from 31%), synthesis time dropped 11→4 days, continuous discovery emerging as dominant use case (28%), demonstrating operational maturation beyond pilot stage.
— Enterprise market adoption snapshot: 51% of enterprises running AI research agents in production, synthesis timelines compressed 6-8 weeks→24-48 hours, continuous discovery moving from theoretical to practical, positioning AI-native research operationalization as mainstream.
— SAGE peer-reviewed edited volume with 10 chapters on AI-assisted QDA covering hybrid human-AI workflows, five-level QDA method adaptation, hallucination/bias/ethics treatment—scholarly validation of AI synthesis as mature field with established practice and pedagogy.
— Business school faculty implementation framework: GenAI compresses research timelines but requires verification of sources, construct validity checks, and enterprise-grade data security; synthesis outputs are predictions, not verifications—demand interrogation over acceptance.
— UserTesting Figma plugin GA (Apr 2026) auto-generates test plans and embeds synthesis results directly into design tool; named deployments (CarMax, AJ Bell) show ecosystem integration reducing research-design iteration delays.
— Market analysis: ecosystem shifted from manual taxonomy-driven repositories to AI-first platforms with auto-tagging and low-friction workflows; 10-tool comparison shows AI-assisted synthesis now table-stakes; segmentation by use case (speed, cross-functional, scale) signals ecosystem maturity.