Research
How does AI actually become enterprise value?
A research program exploring why organizations with access to similar AI technologies produce radically different economic outcomes.
This page is the home of the research program: its central question, the three streams it is organized in, the frameworks each stream develops, and the publications and articles that document it. It evolves as the work does.
The causal chain at the center of the program
One research agenda, eleven connected frameworks
Each framework addresses one point where the conversion chain breaks: the leadership that cannot yet govern AI, the architecture and governance debt that accumulates, the way pilots are built, the operating model that scales them, the measurement that makes value visible, and the capital allocation that reinvests it.
01Three research streams
AI Value Realization
Why do similar AI capabilities produce such different returns?
Organizational and economic research on how AI investment converts into decisions, actions and measurable value. Includes AVRT and its measurement system, the debt concepts (leadership, architecture, governance) and capital allocation across AI portfolios.
Frameworks in this stream
- AI Value Realization Theory (AVRT)
- AI Leadership Debt
- Architecture Leadership Debt
- Governance Debt
- AVRT Measurement System
- AI Capital Allocation Theory
- AI Portfolio Decision Framework
- Executive AI Literacy
Publications & evidence
- AVRT working paper v3 (in progress; executive brief published on LinkedIn, Aug 2026)
- AVRT Series, Parts 1–3, and the AI Strategic Portfolio series (LinkedIn, 2026)
Related articles · AI Value Realization Theory (AVRT) · AI Leadership & Executive Literacy · AI Portfolio & Strategy · Operating Models & AI Governance
AI Value Realization Theory: executive brief
The executive summary of the framework: the conversion chain, the five metrics (ALDI, DCR, ACR, VCR, AIRR), the weakest-link logic, the J-curve, the 10-question self-diagnosis, the positioning map, the critical metric by industry and the 3, 6, 12 and 18–24 month roadmap.
Read · 9 min readBeyond AI-First: Turning Artificial Intelligence into Business Performance
You do not need to become an AI-first company to create value from artificial intelligence. But you do need to rethink how AI changes the way your business operates, decides and generates economic results: where the value is by industry, a worked case that takes gross margin from 30% to 31%, what Microsoft, McKinsey and BCG converge on, why value goes uncaptured (AI Leadership Debt) and how to know whether it is captured (the AVRT conversion funnel).
Read · 15 min readThe P&L Can Be Right — and the AI Decision Can Still Be Wrong
Gartner now measures the hidden workforce costs of AI: a 3–4× higher cost of specialized talent, up to 30% of displaced roles rehired by 2029, a "mortgage on talent". Every one of those costs was caused by a decision, not by the technology. This working note names the aggregate, AI Leadership Debt, and adds the missing term to the AI value equation.
Read · 11 min read02Three research streams
Industrial AI
How do machine learning and agents change the control of complex physical processes?
Applied research at the frontier of process control and AI: ML-driven adaptive control (MPC/APC) for non-linear mining processes, digital twins and operator training simulators, computer vision at scale, and operational AI agents integrating thousands of real-time signals.
Frameworks in this stream
- Evolution-Ready PoC
- Expert Validation Loop
- Agent Harness Engineering
Publications & evidence
- Barrientos C., Neira N., Olivares J., et al. “Non-linear Adaptive APC: ML-Driven Optimization for Copper Mining.” APC Peru, 2025.
- Barrientos C., et al. “Adaptive APC Strategy using ML for SAG Grinding.” APC Peru, 2024.
- Earlier work: geometallurgical predictive models and AutoML in Chilean mining (CORFO / PIA-CONICYT); Hydrometallurgy (Elsevier).
Related articles · Agentic & AI-native Engineering
We develop 10 times faster. Are we also accelerating the errors?
Spec-Driven Development, Context Engineering and Loop Engineering govern the inputs and the execution of AI agents. None of them validates the output as knowledge. The case for an Expert Validation Loop, and why converting incorrect intelligence faster is accelerated risk.
Read · 8 min readBeyond the LLM: Agent Harness Engineering and AI governance
A year building AI agents for industrial environments left an unexpected conclusion: the more it scaled, the less consistent it became. The problem wasn't the model. It was the design of the whole system around it: memory, context, orchestration, observability and governance.
Read · 4 min read03Three research streams
Healthcare AI
Can machine learning support treatment decisions for older adults with cancer?
Clinical decision support for systemic therapy in older adults with cancer. PROTEGER, the platform developed at Oncoger.ai, analyzes each case in three stages, eligibility, dose intensity and prognosis, from the geriatric assessment: the AI recommends, the oncology and geriatric-oncology team decides. Multivariable models with external validation, reported under TRIPOD-AI, and a prospective validation under way. Part of PRECISION-AI, a five-year CORFO-funded technology program led by Universidad de Chile and Universidad del Desarrollo.
With Hospital Clínico Universidad de Chile, Universidad del Desarrollo, Universidad Mayor, Hospital Dr. Franco Ravera Zunino, Hospital Digital (Ministerio de Salud), University Medical Centre Utrecht; international partners MIT Jameel Clinic and the London AI Centre for Value-Based Healthcare.
Frameworks in this stream
- Expert Validation Loop
Publications & evidence
- Navarrete G., Martínez G., Barrientos C., Venegas R., et al. “Machine learning models as a supporting tool to decide systemic therapy for older adults with cancer.” Journal of Geriatric Oncology, 17 (2026) 103031. DOI 10.1016/j.jgo.2026.103031 · ISRCTN11875539.
- Ensemble model with AUC 0.84 (95% CI 0.78–0.89) in external validation (272 patients, two tertiary centers, 2021–2025), after a 357-patient development cohort.
- Platform validating in 7+ hospitals in Chile, Peru and Brazil, with over 1,000 patients included in the machine learning models (Oncoger.ai).
Research
Publications
Peer-reviewed research
- Navarrete G., Martínez G., Barrientos C., Venegas R., et al. “Machine learning models as a supporting tool to decide systemic therapy for older adults with cancer.” Journal of Geriatric Oncology, 17 (2026) 103031. DOI ↗
- 20+ peer-reviewed publications in Hydrometallurgy (Elsevier), SPIE Astronomical Instrumentation and Space Terahertz Technology, 2005–2016.
Industry research
- Barrientos C., Neira N., Olivares J., et al. “Non-linear Adaptive APC: ML-Driven Optimization for Copper Mining.” APC Peru, 2025.
- Barrientos C., et al. “Adaptive APC Strategy using ML for SAG Grinding.” APC Peru, 2024.
Working papers
- Barrientos C. “AI Value Realization Theory (AVRT): a conversion model of AI value in organizations.” Working paper v3, 2026. Available on request.
Articles and outreach
- Articles, series and carousels published on LinkedIn and reproduced in the Articles & research section, grouped by topic.
Currently exploring
Open questions on the research agenda
Four questions currently shaping the work. Comments, counter-evidence and collaboration are welcome.
- 01
Can AI value realization be measured as a conversion system?
AVRT Measurement System
- 02
Can AI Leadership Debt explain persistent differences in AI ROI between organizations?
AI Leadership Debt
- 03
How should organizations allocate capital across AI initiatives under organizational uncertainty?
AI Capital Allocation
- 04
How should AI PoCs be architected to maximize learning without creating Architecture Leadership Debt?
Evolution-Ready PoC · Architecture Leadership Debt