It is the system, not the model: the key to a competitive AI strategy
In 2006, Netflix set the world of artificial intelligence a challenge: to improve its recommendation algorithm by 10% and win a million dollars. Hundreds of research teams from around the world worked for three years to solve what appeared to be a purely technical problem. The winner managed to improve the algorithm by 10.06% and received their prize. But Netflix never implemented that winning solution.
Why didn’t the most sophisticated algorithm ever go into production? Because Netflix, back in 2009, realised something fundamental: the model was just one part of the system. This is a key lesson.
When Netflix evaluated its implementation in a real production environment (millions of concurrent users, dynamic behaviour patterns, and new films being added constantly, etc.), they discovered that the winning model had been optimised for a static snapshot of the past. The computational complexity of the winning algorithm meant that the model’s update times were prohibitively long. The 10.06% improvement did not justify the added complexity, the exponential computational cost, the slower update times, or the interpretability issues it entailed.
The success of AI lies not in the model, but in the system that powers it
The quality of the data feeding into the model had also improved over the three years of the competition.
User behaviour patterns changed, the film database expanded and consumer devices diversified.
The winning model had been optimised for a photograph from the past, not for the dynamics of the actual system in production three years later.
Aware of this reality, Netflix took a strategic decision that marked a turning point for the industry: rather than implementing the winning solution as it stood, they used their algorithmic insights as a reference and built an integrated system that combined simpler yet robust models, rigorous data governance, scalable infrastructure, continuous evaluation, and radical explainability of the system’s operation.
The result was a reliable system that worked. We can apply this metaphor to any organisation embarking on an AI strategy that aims to be successful.
Trust in AI solutions is built through integrated, reliable, and scalable systems
Nowadays, when we look at the adoption of AI in corporate environments, we see that this widespread trend has become firmly established: the focus is on the algorithm or the model.
Eighty per cent of the budget is spent on developing a sophisticated model, whilst the remaining 20% is allocated to everything else: governance, data quality, infrastructure, security, compliance, people, etc.
It is a misplaced investment that leads to projects which shine in test or pilot environments but fail in production environments.
A successful strategy must also measure: production reliability, explainability of decisions, regulatory compliance, user adoption and business impact (ROI). If your solution is 95% accurate, but users do not trust its decisions, it has failed as a system.
In essence, a successful AI strategy is the convergence of multiple dimensions operating in an integrated manner. We must realise that a sophisticated algorithm without the support of these dimensions is like building a skyscraper without foundations.
At Izertis, having worked with hundreds of organisations on digital transformation and AI solutions, we operate on the premise that trust in AI solutions is not built on sophisticated models, but on integrated, reliable and scalable systems. This is the essence of Trust AI, our strategic approach.
To this end, Izertis has structured Trust AI around four key dimensions:
1. People: trust and organisational adoption
Without teams that understand and trust algorithmic decisions and the impact of AI solutions on their business, the strategy will fail. It includes training, internal communication, change management and building trust in the adoption of AI.
2. Governance: framework for decision-making and accountability
Well-defined policies, clear roles, traceability of decisions and ongoing auditing are essential. Rolling out solutions without a sound corporate governance framework is a black box that is unacceptable from a business perspective.
3. Technology: infrastructure and scalability
An AI solution that works in test environments but not in production with millions of records is merely a proof of concept; that is why the infrastructure, automation, monitoring, observability and security architecture are just as critical as the AI solution itself.
4. Processes: data quality and continuity
Biased, incomplete or out-of-date data leads to business-wise flawed AI solutions. For this reason, you must give careful consideration to data governance, quality assurance and continuous improvement.
Without these four dimensions working in unison, the sophistication of the solution is irrelevant.
At Izertis, our commitment to Trust AI reflects this maturity: trust in AI solutions stems from robust systems that integrate four dimensions (people, governance, technology, processes), not from sophisticated algorithms in isolation.
Trust in AI solutions stems from robust systems
If your company is considering an AI project, we invite you to reflect on the following: Are you focusing on a model or a system?
It is the difference between a transformation initiative that fails quietly in production and one that generates a measurable and sustainable impact.
If you have already started your AI roadmap and wish to assess the maturity of your AI system across the four dimensions of our framework, our team of Trust AI experts will help you identify your strengths, vulnerabilities and specific areas for improvement.