Tech Giants Cannot Deploy AI Because Their Main Problem Is Not the Model—It Is Their Own Company
Why are large companies held back in deploying AI not by weak models, but by their own processes and organizations?
Test whether Google and OpenAI become useful everyday interfaces or require constant correction; the next step is to check how many steps the interface actually removes and what dependency it creates in return.
What to watch for
Key takeaways
The “Sam Altman can talk about hundreds of billions for AGI, and Google’s CEO can talk about” scene leads to a working conclusion: inside an ordinary company, however, nothing begins with the model. First you have to understand where the data lives, who owns the process, and what AI is actually supposed to improve. Without that, even DeepSeek, OpenAI, or an in-house model becomes one more system that employees work around.
The working conclusion from “The latest news from the world of technology — part 2/2” is that the week's roundup covers Instagram's content tools, the SoftBank–OpenAI joint venture in Japan, Amazon's promised Alexa announcement, Meta's Ray-Ban glasses ads, and a wave of AI-device startups.
The decision in “The problems of deploying AI in B2B” depends on one criterion: B2B adoption breaks on integration and accountability — a company installs a tool but cannot give it access to the right databases, cannot measure quality, and fears a leak; the pilot yields a pretty presentation while daily work stays the same.
The working conclusion from “Instagram launches content software, Google an assistant” is that Instagram is shipping content tools while Google turns search into an assistant — both companies embed AI where the user already is instead of waiting for the user to come to a separate app.
In the context of “SoftBank and OpenAI: establishment of a joint venture,” this criterion applies: in Tokyo OpenAI announced a joint company with SoftBank to build AI sales assistants — but for now this looks more like reasoning than movement toward real use cases.
The “Current technology limitations” scene leads to a working conclusion: the lack of successful enterprise cases may be a limit of the technology itself — Yann LeCun expects today's LLMs to be forgotten within a year or two because of hallucinations and the statistical nature of generation; and adoption still demands classic digitalization: people, data, and process setup.
The boundary of the “Amazon promises a big announcement for Alexa” case is defined by this point: the promise to upgrade a voice assistant changes nothing by itself — a habit will form only if the new Alexa proves faster and more reliable than the old way of doing the same tasks.
The “The “cringe ads” of AI companies” topic becomes clearer once this point is included: a strong model does not guarantee that a user will start opening a camera and asking about every object. A habit forms only where the new function is faster and more reliable than the old way.
In the context of “Meta's new Ray-Ban glasses ad: a promising use case?,” this criterion applies: everyone is looking for distribution—the place where AI meets an existing user. Without an interface and a process, even the best model remains an expensive API.
The boundary of the “Startups building AI devices” case is defined by this point: do not start by asking which model to buy — take one process, count the time it consumes, identify the data and the cost of an error, then test whether AI reduces manual work; the giants are not exempt from this logic.
What this episode is about
Amazon, Google, and Meta are spending billions on AI, yet their internal data is fragmented, their processes are unprepared, and users rarely turn on new features. OpenAI faces the same problem: powerful technology does not become a product automatically. Enterprise adoption requires tedious work that advertising cannot replace.
Sam Altman can talk about hundreds of billions for AGI, and Google’s CEO can talk about turning search into an assistant. Inside an ordinary company, however, nothing begins with the model.
First you have to understand where the data lives, who owns the process, and what AI is actually supposed to improve. Without that, even DeepSeek, OpenAI, or an in-house model becomes one more system that employees work around.
B2B adoption breaks on integration and accountability. A company installs a tool but cannot give it access to the right databases, does not know how to measure quality, and fears a leak. The pilot produces an attractive presentation, while daily work remains exactly the same.
OpenAI also releases products that do not always find a use case: separate search features, agents, generators, and new interfaces. A strong model does not guarantee that a user will start opening a camera and asking about every object. A habit forms only where the new function is faster and more reliable than the old way.
SoftBank and OpenAI are creating a joint venture in Japan, Meta is developing Ray-Ban glasses, and Instagram is launching content tools. Everyone is looking for distribution—the place where AI meets an existing user. Without an interface and a process, even the best model remains an expensive API.
The central lesson for business is very practical. Do not start by asking which model to buy. Take one process, calculate the time it consumes, identify the data and the cost of an error, and then test whether AI reduces manual work. The giants are not exempt from this logic. Their scale does not make a bad process less bad; it merely makes it more expensive.
The giants are not exempt from this logic. As a result, their scale does not make a bad process less bad; it merely makes it more expensive.
Episode transcript
The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 80 segments: 40 identified, 3 mixed, 31 probable, and 6 unresolved.
Read transcript on a separate page
Loading…