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 analyses learning data to predict outcomes, identify at-risk students, and measure engagement patterns. Includes early warning systems and engagement scoring; distinct from skills assessment which evaluates competency rather than predicting trajectories.
Learning analytics can identify at-risk students with increasing precision. Whether those identifications translate into better outcomes remains the practice's defining tension. Forward-leaning universities and K-12 districts have deployed predictive models at meaningful scale, and state-level mandates are now accelerating adoption. The technical capability is proven: meta-analysis of 15 studies (199K participants) confirms 91% accuracy in AI-based dropout prediction; cross-national PISA 2022 study of 26,969 students using XGBoost achieved 57% variance explanation in mathematics performance; ML forecasters reduce effort/progress prediction error by 22-33% versus heuristics; and emerging systems achieve 91.7% classification accuracy with privacy-preserving federated learning design. New measurement instruments (UDIFP-29 for dropout intention; Canvas engagement analytics) enable earlier, more nuanced identification. Production systems at named institutions deliver measurable retention gains (8pp reading improvement, 6pp math improvement, 60% intervention success rates). But the field has not yet crossed into mainstream practice. Documented racial bias in deployed models, persistent gaps between identification and effective intervention, and a research literature that overwhelmingly neglects learning outcome measurement—a 2020 systematic review of 46 studies found "rigorous, large-scale evidence of effectiveness is still lacking"—all constrain broader adoption. Research teams are advancing fairness-aware algorithms with demonstrable progress (e.g., 0.35→0.08 reduction in bias severity and 15.3%→4.2% improvement in demographic parity), and the ecosystem is maturing with IES-funded research on fair prediction and open-source toolkits. Yet adoption barriers remain structural and persistent: only 23% of administrators actively assess for bias, intervention effectiveness remains uncertain, and regulatory constraints (COPPA 2026 effective April 22, FERPA loopholes enabling 1,449 EdTech tools per district affecting 55M students) continue reshaping deployment constraints. Fundamental statistical limits on rare-event prediction (the "Likelihood Ratio Wall") limit achievable fairness independent of algorithm design. The vanguard is getting value; most institutions have not started.
Two US states — Utah and Iowa — now mandate early warning systems across all local education agencies, with Panorama Education serving as the primary vendor. That policy momentum, combined with Panorama's reach across 2,000+ K-12 districts and 15M+ students, marks real expansion in deployment footprint. Independent procurement data from 79K+ school agencies (Civic IQ, June 2026) confirms Panorama among top-5 K-12 EdTech vendors with 58+ active spend records and $19K-$27K contract values. State-level policy integration is accelerating: Illinois designated Panorama as approved alternative survey provider for 2025-2026, integrating it into state MTSS accountability frameworks. A critical July 2026 shift: Instructure (Canvas/LMS platform) released native predictive dropout risk identification, moving learning analytics from third-party add-on to platform-native capability—signalling the ecosystem has reached platform maturity at the LMS level. August 2026 validation extends the evidence base: South African University of Technology deployed early-warning models achieving ROC-AUC 0.835 with 78.9% detection of future non-graduates before Year 2; Achieving the Dream's Digital Holistic Student Supports Initiative committed $2.5M across five named US institutions for unified data systems and predictive dashboards; global market projections forecast learning analytics growing 18% CAGR to $31.2B by 2034, with predictive analytics segment representing 57% of education analytics market. In higher education, Civitas Learning serves 400+ institutions and reports retention gains of 3-11% across its client base; EAB's Navigate360 platform serves 850+ institutions and 10M students with documented 2-12% retention and 3-15% graduation improvement. SEAtS ONE platform is in general availability across 200+ higher education institutions. New deployments confirm continued adoption: University of Utah deployed dual analytics dashboards April 2026 for engagement and retention analysis; Broward County Public Schools (one of nation's largest districts) expanded Panorama Student Success for identifying early warning signals tied to attendance, academics, and behavior; Florida International University and Georgia State University deployed ML models achieving 7% graduation rate improvement with stronger gains for underserved populations; Penn State University deployed Panorama analytics platform (launched June 2025) earning 2026 1EdTech award for measurably improving course design and accessibility. IU Indianapolis reduced its retention gap from 19% to 12.7% through data-informed proactive advising with explainable AI and bias mitigation integrated into production systems. May 2026 evidence confirms expanded deployment momentum: Reynolds Community College achieved highest enrollment in 6 years and $1M+ cost savings via SAS Viya analytics; University of Arizona deployed systems achieving 90% early-warning accuracy within first 12 weeks; Community colleges nationally report 11-18pp retention gains from trigger-based at-risk workflows. Long-term audited evidence from Georgia State University's 13-year production deployment (tracking 800+ risk indicators per student) demonstrates sustained ROI: every 1% increase in retention generates $3.18M in tuition revenue, accumulating to $60M+ from 23-point graduation improvement since 2003. International deployments expand the evidence base: India's Ministry of Education UDISE+ system now tracks dropout and retention nationally, with preparatory level dropout falling 2.3%→1.8% and secondary level 8.2%→7.0% (2025–2026), signaling government-wide adoption of learning analytics infrastructure; 3 Nigerian universities with 19,961 student records demonstrate Hist Gradient Boosting effectiveness in distinct sociocultural contexts. Market research shows predictive analytics segment of education learning analytics growing 22% CAGR (2025-2030) from $10B to $27B, representing 57% of the total education analytics market. Emerging sector emphasis on explainability is strengthening: systems like RADAR achieving ~93% accuracy combine early detection with transparent decision-making to enable educator validation rather than blind adherence to algorithmic recommendations. Institutional governance is maturing: NYC Department of Education (serving 1.1M students across 1,700+ schools) now requires bias and equity review for all AI tools including learning analytics before deployment, establishing a 'traffic light' framework (green/yellow/red) categorizing appropriate use cases and safeguards.
These successes, however, sit alongside persistent structural barriers intensifying in 2026, confirmed by August 2026 research: 89% of U.S. schools deploy student activity monitoring for risk assessment yet documentation shows disproportionate flagging of LGBTQ+ and disabled students; joint research across 200+ higher education institutions found 68% retain raw interaction logs beyond academic term and 35 institutions initiated audits after discovery of differential outcomes (7% lower success rates for Black/Hispanic vs. White students). A critical 30-author research perspective demonstrates that prediction accuracy ≠ intervention outcomes: "a dropout score does not tutor a student." Risk scores only affect outcomes through organizational decision changes—advising allocation, resource availability, protocol changes—and high identification accuracy alone cannot overcome weak intervention infrastructure. Practitioner analysis confirms this gap: ~93% of four-year institutions run early alert systems, yet outcome evidence is thin; high-performing programs succeed through narrow measurable objectives, speed-to-contact within 48 hours, and clear ownership, not through model sophistication. Regulatory momentum has reversed sharply: effective July 24, 2026, the US Department of Education eliminated the disparate-impact investigation tool under Title VI, now requiring proof of intentional discrimination only—this removes the federal oversight mechanism for investigating potentially biased learning analytics systems, intensifying governance uncertainty for institutions deploying prediction models. Data-driven systems relying on attendance, behavior, and grades as primary signals also miss earlier emotional and relational precursors—sense of belonging, emotional well-being—that predict disengagement before grades or attendance shift, leaving a detection blindness even when identification accuracy is high. A meta-analysis of 936 learning analytics papers found 70% lacked any learning outcome measures, suggesting field-wide research has drifted from educational improvement. Independent analysis of 1,000+ student success initiatives found 40% showed little or no measurable impact. Deployment infrastructure gaps are substantial: implementing a complete early warning system requires significant infrastructure (LMS integrations, ML models, data governance, GDPR/COPPA compliance, staff training), and entry costs remain prohibitive for most schools and districts; median full-time adviser manages 296 advisees, creating bandwidth constraints on contact delivery. FERPA audit analysis (June 2026) documents that major SIS vendors (PowerSchool, Infinite Campus, Ellucian, Anthology) are shipping AI-powered predictive analytics for student retention and behavioral risk scoring in production systems, but 4 critical compliance gaps persist: subprocessor opacity, model training ambiguity, de-identification failures, and audit trail gaps. Fairness remains acute and increasingly visible: large-scale real-world studies across 600k+ students in 80 education systems demonstrate bias concerns in ML-based risk prediction; deployed systems document false negative rates of 19-21% for Black and Hispanic students compared to 6-12% for White and Asian students, and the Wisconsin Dropout Early Warning System disproportionately flagged African American and Hispanic students despite low actual risk. Evidence further shows that student risk prediction tools have wrongly labeled 19% of Black and 21% of Latinx students as likely dropouts despite their later earning degrees. Yet only 23% of administrators actively assess for algorithmic bias. Privacy-preserving technical approaches are advancing: research demonstrates synthetic educational data (SynEdu-HEDL framework using rule-based and probabilistic generation) enables institutions to develop early warning systems without exposing real student records; LSTM-GNN models trained on synthetic data achieve 78.1% AUC after fine-tuning with just 5% real data, addressing data scarcity barriers while maintaining FERPA compliance. Fundamental statistical research (Likelihood Ratio Wall, ACM FAccT 2026) proves that rare-event prediction systems (student dropout 3-8% base rates) face irresolvable fairness constraints at the mathematical level—high precision on positive predictions requires tools far more discriminative than current instruments provide, and demographic groups subject to historic under-service face structurally lower maximum achievable fairness metrics independent of algorithm choice. A related fairness failure—algorithmic exclusion—occurs when AI systems lack sufficient data on certain populations to make meaningful predictions at all, systematically failing the most marginalized students. Regulatory constraints are now sharply constraining K-12 deployment: COPPA 2026 (effective April 22) requires parental consent for any AI-powered learning analytics features and mandates data minimization. FERPA governance remains inadequate—the 1974 framework was designed for file cabinets, not cloud-based AI systems; the average U.S. school district uses 1,449 EdTech tools as potential "school officials," affecting 55M K-12 students. Generative AI integration is advancing with Panorama's Solara platform in production, but the harder problems of equitable intervention design, regulatory compliance, institutional capacity, and algorithmic fairness remain unresolved.
— Peer-reviewed study on Open University Learning Analytics Dataset (27,522 enrolments) achieving day-28 gradient boosting ROC-AUC 0.790, demonstrating early risk prediction feasibility with explicit caution on calibration and human oversight.
— Achieving the Dream launched DHSS Initiative (January 2026) with five named institutions receiving $500K grants each (Clovis CC, Durham Tech CC, North Central SC, Prince George's CC, Fayetteville State); 2-year rollout of unified data systems and predictive analytics dashboards.
— 89% of U.S. schools deploy student activity monitoring systems for risk assessment; research documents disproportionate flagging of LGBTQ+ and disabled students; Department of Education eliminated Title VI disparate-impact enforcement (July 2026), removing federal oversight mechanism for deployed biased analytics systems.
— State-level deployment of Gujarat AI early warning system (co-developed with Wadhwani AI and UNICEF India): Year 1 flagged 167,000 at-risk students (55% girls); Year 2 identified 118,000 additional; documented intervention success with targeted scholarships and teacher outreach.
— Preprint framework proposing Student Digital Twins for early intervention, grounded in Course Signals (Purdue) and GPS Advising (Georgia State) deployments; addresses the prediction-to-intervention gap via causal, actionable intervention.
— De Montfort University's Insight and Understanding Framework embedded across 2,300+ students and 90+ faculty in School of Computer Science; national teaching fellowship (55 awarded 2026) recognizing sustained improvements in student engagement and continuation with human-centered oversight.
— Peer-reviewed empirical study using 15,901 records (2017–2025) from Durban University of Technology achieving ROC-AUC 0.888 (full model) and 0.835 (early-warning year-1 model) with 78.9% detection of future non-graduates before Year 2.
— Joint research by University of Michigan, Institute for Ethical AI in Education, and EDPB on 200+ HE institutions: 68% retain raw interaction logs beyond term; LearnAI platform shows 7% lower success rate for Black/Hispanic vs. White students; 35 universities initiated bias audits post-publication.