In this article
In this article
Browse the research library
Research
The research library
No forms, no gates, and every number carries the source it came from. Where a figure is somebody's forecast rather than a measurement, the page says so, which is exactly what makes the rest of the numbers worth your time.
Published reports
- 01Why enterprise AI adoption is failingSix research houses, six methods, one answer: what stops these projects is the data and the way the work is run, not the model.What the measured evidence says about why pilots stall before production, and why the cause is usually the data and how the work is run.Free to read, no form. Every citation visible, every forecast labelled as one.
- 02The enterprise data foundation layer for AISearching your documents should be one thing the system can do, not the only thing it knows how to do.Where company data really sits across core systems and a decade of shared drives, and why document search should be one option the system picks rather than the only thing it knows.The routing decision tree and the connector list, published in full.
- 03The scaling cliffA pilot is a different system from production, and the sums change with it.What happens to an AI workload between pilot and production, why the sums that worked at ten users fail at ten thousand, and where running it yourself stops making sense.It finds that a hosted open-weight endpoint is the cheaper answer for many workloads. The arithmetic is published so you can run it on your own numbers.
In preparation
An original benchmark on enterprise documents
Three open-weight models tuned on invoices, filings, purchase orders, safety data sheets and mixed-script records, measured against a frontier model on accuracy and speed.
Ask us to send it when it landsHow we publish
The rules we hold ourselves to when we publish, and the reasoning behind each one.
Nothing sits behind a form.
Every citation stays visible.
A forecast is labelled as a forecast.
Published as pages, not files.
We publish what the evidence says.
Citation library
Every source the reports lean on, linked back to whoever published it. 39 of 41 keep a stable public copy. Where one does not, we say so rather than pointing you at somebody's mirror.
Links open on the publisher's own site
- The Root Causes of Failure for Artificial Intelligence Projects (RR-A2680-1) (opens in a new tab)
RAND Corporation, Ryseff, De Bruhl & Newberry · Independent research · 2024
65 practitioner interviews. More than 80% of AI projects fail, twice the rate of IT projects without AI. Four of the five root causes are organisational, not technical.
The GenAI Divide: State of AI in Business 2025
MIT NANDA · Independent research · 2025 · no stable public copy
300 public deployments reviewed, 52 structured interviews, 153 survey responses. 95% of enterprise GenAI pilots produced no measurable P&L impact against $30–40bn of investment. No stable public URL; cited from the published report.
- Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (opens in a new tab)
Gartner · Industry analysis · 2025
A forecast, not a measurement, we label it as such wherever it appears. Cites escalating cost, unclear value and inadequate risk controls. Also estimates only ~130 of thousands of self-described agentic vendors are real.
- Cost of a Data Breach Report 2025 (opens in a new tab)
IBM Security / Ponemon Institute · Industry analysis · 2025
600 breached organisations across 17 industries. 97% of organisations with an AI-related breach lacked proper AI access controls; shadow AI added roughly $670,000 to average breach cost.
- Prompt injection is not SQL injection (it may be worse) (opens in a new tab)
UK National Cyber Security Centre · Regulator or statute · 2025
The national technical authority's position: because a language model cannot separate instruction from data, prompt injection may never be fully mitigated. Design for blast radius, not prevention.
- Top 10 for Large Language Model Applications (opens in a new tab)
OWASP · Industry analysis · 2025
Prompt injection is LLM01, the first-ranked risk for language-model applications.
- Seven Failure Points When Engineering a Retrieval Augmented System (opens in a new tab)
Barnett, Kurniawan, Thudumu, Brannelly & Abdelrazek · Peer-reviewed · 2024
CAIN '24. The canonical taxonomy of where retrieval-augmented systems break. Establishes that the failures are architectural rather than tunable.
- From Local to Global: A Graph RAG Approach to Query-Focused Summarization (opens in a new tab)
Edge et al. Microsoft Research · Peer-reviewed · 2024
Graph-based retrieval beat vector retrieval on comprehensiveness win-rates of 72–83%, using over 97% fewer tokens for root-level community summaries. Win-rates are LLM-judged, not ground truth.
- FRAMES: Fact, Retrieval, and Reasoning Measurement Set (opens in a new tab)
Krishna et al. Google & Harvard · Peer-reviewed · 2024
Single-step naive retrieval scores ~0.40 accuracy against 0.66 for multi-step and 0.73 with oracle documents. Multi-hop questions are where naive retrieval collapses.
- Don't Do RAG: Cache-Augmented Generation (opens in a new tab)
Chan et al. National Chengchi University · Peer-reviewed · 2024
For bounded, stable corpora, preloading the whole corpus into cached context removes the retrieval hop entirely, large latency reductions at comparable or better accuracy.
- DistilBERT, a distilled version of BERT (opens in a new tab)
Sanh, Debut, Chaumond & Wolf, Hugging Face · Peer-reviewed · 2019
40% smaller, 60% faster, ~97% of language understanding retained. Still the clearest published statement of what distillation actually costs you.
- A Benchmark to Understand the Role of Knowledge Graphs on LLM's Accuracy for Q&A on Enterprise SQL Databases (opens in a new tab)
Sequeda, Allemang & Jacob, data.world · Peer-reviewed · 2023
Accuracy on enterprise question answering rose from 16.7% on a raw schema to 54.2% when the same questions were grounded in a knowledge graph.
- Deep Entity Matching with Pre-Trained Language Models (Ditto) (opens in a new tab)
Li, Li, Suhara, Doan & Tan · Peer-reviewed · 2021
VLDB 2021. The entity-resolution method behind a workable authority index: F1 of 96.5% matching two company datasets of 789K and 412K records.
- Regulation (EU) 2024/1689, the AI Act (opens in a new tab)
European Union · Regulator or statute · 2024
The consolidated text as amended in July 2026. The Commission's own explainer page is the easier starting point.
- The State of AI (opens in a new tab)
McKinsey & Company (QuantumBlack) · Industry analysis · 2025
88% of organisations use AI somewhere; only ~39% report enterprise-level EBIT impact. Workflow redesign is the single attribute most correlated with that impact.
Voice of the Enterprise: AI & Machine Learning
S&P Global Market Intelligence · Industry analysis · 2025 · no stable public copy
The share of companies abandoning most AI initiatives before production rose from 17% to 42% year over year. Cost, data privacy and security lead the cited obstacles. Subscription research; no public URL.
- OpenLineage, an open standard for lineage metadata collection (opens in a new tab)
OpenLineage (LF AI & Data) · Industry analysis · 2026
The spec we emit lineage against, so your lineage is portable rather than locked to our tooling.
- Apache Iceberg (opens in a new tab)
Apache Software Foundation · Industry analysis · 2026
Our default open table format: specification-first and engine-independent, which is what makes a vendor-neutral lakehouse possible.
- Model Context Protocol specification (opens in a new tab)
Model Context Protocol · Industry analysis · 2026
How we expose governed queries as tools an agent can call, with the server acting as the policy gatekeeper rather than the model.
- Presidio, data protection and PII anonymisation (opens in a new tab)
Microsoft · Industry analysis · 2026
The pattern our classification stage follows: regex and checksum recognisers for structural identifiers, NER for contextual ones, confidence scored per column.
- Regulatory framework for artificial intelligence (opens in a new tab)
European Commission · Regulator or statute · 2026
The Commission's plain summary of the AI Act.
- Regulation (EU) 2026/1744, the Digital Omnibus on AI (opens in a new tab)
European Union · Regulator or statute · 2026
The 2026 amendment that moved the high-risk deadlines.
- ISO/IEC 42001:2023, Artificial intelligence management system (opens in a new tab)
ISO/IEC · Regulator or statute · 2023
The certifiable standard for how an organisation governs AI.
- ISO/IEC 23894:2023, Guidance on AI risk management (opens in a new tab)
ISO/IEC · Regulator or statute · 2023
Guidance, not certifiable; feeds into 42001.
- ISO/IEC 27001:2022, Information security management systems (opens in a new tab)
ISO/IEC · Regulator or statute · 2022
Often paired with 42001 where AI handles sensitive data.
- Artificial Intelligence Risk Management Framework (AI RMF 1.0) (opens in a new tab)
NIST · Regulator or statute · 2023
Voluntary US framework: govern, map, measure, manage.
- AI RMF Generative AI Profile (NIST AI 600-1) (opens in a new tab)
NIST · Regulator or statute · 2024
The generative-AI companion to the framework.
- Regulation (EU) 2016/679, the General Data Protection Regulation (opens in a new tab)
European Union · Regulator or statute · 2016
Applies fully to AI systems that process personal data.
- Regulation (EU) 2022/2554, Digital Operational Resilience Act (opens in a new tab)
European Union · Regulator or statute · 2022
ICT resilience rules for EU financial entities, applicable since January 2025.
- Regulation (EU) 2024/2847, Cyber Resilience Act (opens in a new tab)
European Union · Regulator or statute · 2024
Security requirements for products with digital elements.
- Directive (EU) 2024/2853, Product Liability Directive (opens in a new tab)
European Union · Regulator or statute · 2024
Extends product liability to software and AI.
- SR 11-7, Supervisory Guidance on Model Risk Management (opens in a new tab)
Federal Reserve and OCC · Regulator or statute · 2011
The reference text for model inventories and independent validation.
- SB 26-189, Automated Decision-Making Technology (opens in a new tab)
Colorado General Assembly · Regulator or statute · 2026
Replaced the 2024 Colorado AI Act; effective 1 January 2027.
- HB 149, Texas Responsible Artificial Intelligence Governance Act (opens in a new tab)
Texas Legislature · Regulator or statute · 2025
Intent-based rules, in force since 1 January 2026.
- Guidance on AI and data protection (opens in a new tab)
UK Information Commissioner's Office · Regulator or statute · 2025
How UK data protection law applies to AI.
- Model AI Governance Framework (opens in a new tab)
IMDA and PDPC, Singapore · Regulator or statute · 2024
Voluntary framework, with a 2024 generative-AI edition.
- AI Verify testing framework (opens in a new tab)
AI Verify Foundation · Regulator or statute · 2026
A voluntary testing toolkit that complements the Singapore framework.
- Act on Promotion of Research, Development and Utilisation of AI-Related Technologies (opens in a new tab)
Government of Japan · Regulator or statute · 2025
A framework law in force since September 2025, with no penalty regime.
- Digital Personal Data Protection Act 2023 and Rules 2025 (opens in a new tab)
Ministry of Electronics and Information Technology, India · Regulator or statute · 2025
Consent-based data protection law, operational since the 2025 Rules.
- OECD AI Principles (opens in a new tab)
OECD · Regulator or statute · 2024
The first intergovernmental AI standard, updated May 2024.
- State of FinOps 2026 (opens in a new tab)
FinOps Foundation · Industry analysis · 2026
Practitioner survey on managing AI spend.