According to The Kobeissi Letter, the combined capital spending on AI by Alphabet, Amazon, Meta, Microsoft, and Oracle is projected to reach around $1.1 trillion by 2027, which would be about 3.2% of the US GDP. At the same time, most companies are struggling to scale these technologies.
McKinsey research shows that while roughly 65% of organizations are actively integrating generative AI into their workflows, only a small number of companies have seen tangible economic impact.
The problem lies in the gap between the public perception of AI as an all-knowing mind and the reality, where basic language models often struggle with specialized business tasks. Waiting for a smarter next version of the model — the next GPT, the next Claude — has become a trap that distracts resources from building an operational layer that tailors AI to specific business needs.
The invisible army of AI architects #
The AI industry is like buying the most powerful computer in the world — it doesn't guarantee high-quality code if there's no programmer who knows how to use it. The general public expects an AI that can tackle any task on its own, but behind the scenes of successful deployments, there's an army of specialists at work.
These engineers focus on creating complex architectural "wrappers." This includes data retrieval systems (RAG), verification cycles, special memory formats, and libraries of reusable skills. In recent years, progress in the real sector has come thanks to these invisible layers, not just by increasing the number of parameters in neural networks.
When companies have to integrate external tools to get a model to work reliably, management often sees it as a failure or a 'crutch.' But modern research shows that this strategy isn't an admission of AI incompetence — it's the only way to build reliable systems.
"Local patch"
Research by Mikhail Arbuzov plays an important role in understanding the paradigm shift — he's a senior applied AI manager with over ten years of experience in machine learning and an inventor on multiple U.S. patent filings in artificial intelligence and telecommunications machine learning, spanning granted patents, published applications, and pending provisional applications. Together with Sisong Bei and other collaborators, Arbuzov has produced works such as* 'Beyond Exponential Decay' *(arXiv:2505.24187) and 'The Architecture of Errors' (arXiv:2605.30628), which argue that reliability should be engineered within bounded task domains — where recurring failure modes can be identified and addressed systematically — rather than assumed to emerge uniformly from ever-larger general-purpose models. Arbuzov's formalization of "patch-local" reliability is among the first rigorous theoretical arguments that dependable AI must be engineered locally, task by task, rather than waiting for ever-larger models — an early formal account of how the industry can engineer reliable AI systems.
Arbuzov says we shouldn't expect a model that works flawlessly across all domains without exception. Reliability is always local. It has to be 'earned' within specific conditions, task by task.
"If a model makes mistakes in arithmetic, you don't need to wait for a more advanced version — you just need to give it a calculator. If it doesn't navigate a specific corporate context, it should be provided with the workflows, data, and constraints that the team already uses. This isn't cheating; it's a proper engineering approach to building business solutions," notes Mikhail Arbuzov.
The economy of change #
Practical evidence for this approach comes from the researchers' own experiments — and from an open-source tool built to run them. Skillfed, an engine that gives AI agents native skill search — a way for a model to find the right reusable, task-specific tools at the moment it needs them — was co-created by Arbuzov and Bei: Arbuzov contributed the underlying retrieval architecture, and Bei built the backend and led the experimental evaluation. It is a search engine designed for AI agents rather than people, and it was downloaded more than 1,000 times in its first week (per public npm and PyPI registry statistics) — work that places Bei among the pioneers of agent-native retrieval.
Using the tasks from SkillsBench, a public benchmark for skill use, the researchers measured what happens when a model can draw on the right skills through Skillfed. In one internal test, when Claude Opus 4.5 was given access to relevant skills, task accuracy rose by more than 20 percentage points. In some of these internal test conditions, smaller, cost-effective models equipped with relevant skills outperformed larger models operating without them. In a separate, tougher internal test, the researchers stripped out the 'oracle' hints that normally tell a model which skill to use, and asked Claude Opus 4.6 to find the right tools on its own from a library of about 26,000 real-world skills (a 26,629-skill snapshot). Even with no guidance, its performance improved by roughly 30% relative to its no-skill baseline. These results are in addition to the group's originally reported SkillsBench figures — a lift from 17.5% without retrieval to 22.8% with retrieval, against a 36.8% oracle ceiling, on the same snapshot. Together, the numbers suggest that a model's raw "intelligence" matters less than an operational layer that lets the system effectively find and use the knowledge already available.
There is a reason the right skill helps so much. The SkillsBench research found that skills a model generates for itself add little, because they mostly restate knowledge the model already holds; a skill earns its value only when it carries information specific to the task — a procedure, a format, an organization's own terrain — that is not in the model's training or cannot easily be recovered from it. That is why locating the right skill among tens of thousands, rather than enlarging the model, is where the gains come from — the same logic behind Arbuzov and Bei's argument that reliable AI is engineered in the layer around the model, not inside it.
Industry leaders' consensus #
This view on technology development is shared by recognized industry leaders. American computer scientist Andrew Ng has repeatedly emphasized that progress in AI-agent workflows will become a major driver of development, possibly even more powerful than creating next-generation foundational models. Shopify CEO Tobi Lütke coined the term 'context engineering,' describing it as the art of presenting a task in a way that allows the model to work flawlessly.
One of the founders of OpenAI and former AI director at Tesla, Andrej Karpathy, calls this process the science of filling the context window, while Satya Nadella from Microsoft sees AI implementation as fundamental changes to workflows. All the experts agree that the battleground for efficiency has shifted from training the model's 'brain' to designing the architecture around it.
Operational stage #
The future described in the forthcoming article 'Frontier and Localhost,' co-authored by Arbuzov and his collaborators, suggests that learning is shifting from the model training stage to its operational stage. The industry has created this working layer (tools, memory, instructions) everywhere, but has not yet developed strict engineering standards for managing it.
Sisong Bei compares AI deployment to onboarding a new employee.
"When a person joins a company, they are introduced to the environment, familiarized with regulations, and trained to use internal systems. AI requires a similar approach — providing context, norms, and examples. The advantage is that while a person needs lengthy training, the 'adaptation layer' for AI is created once and then scaled," says Sisong Bei.
Conclusion #
Instead of waiting for the arrival of an expensive 'superintelligence,' companies can get working systems today by directing resources toward creating an adaptation layer. The information everyone is waiting for is already here, spread across tools, skills, and coordination mechanisms.
The breakthrough is in taking responsibility for deploying the model, for creating the infrastructure that turns raw algorithmic capabilities into a reliable business tool.
Today's success in AI is a matter of the quality of engineering work in integrating it into a real business context. Those who build the operational layer first will gain a competitive advantage in the coming years.
© Copyright IBTimes 2026. All rights reserved.