People and AI: What the Big Labs Got Wrong

The evidence according to O'Reilly

Tim O'Reilly has measured value for decades against one constant criterion: create more value than you capture. He now applies this yardstick to artificial intelligence.

In the interview given to Wired, the publisher and internet pioneer makes a sharp claim: the big labs have misread the future and misunderstood what people really want. His observation shifts attention from the technology to the conditions of adoption.

O'Reilly argues that frontier models are optimized for specific use cases far removed from what people desire. He calls the current setup an architecture of control, designed to track users and bind them to the product. The comparison recalls Microsoft in the 1990s, when lock-in defined the dominant strategy.

For those who lead people inside organizations, this reading carries immediate operational weight.

The gap between frontier AI and real use

O'Reilly identifies a precise gap: advances in so-called frontier AI move the technology away from the needs of ordinary people. The largest models turn out to be better for some tasks and worse for others.

He cites a concrete example. Some users judge advanced writing models, such as Fable and Sol, inferior to lower-tier models for producing text.

Anthropic and OpenAI disagree with this assessment, as the interview itself records. The structural point remains valid nonetheless for those who design work: the race for maximum capability assumes that model power coincides with user value.

That assumption deserves empirical verification inside every organization. The useful question becomes operational: which of people's tasks draw real benefit from frontier models, and which work better with lighter, controllable tools?

Architecture of control versus participation

The strongest distinction O'Reilly proposes separates three elements: the model, the orchestration infrastructure (harness) and the application. He calls for a clean separation between these layers.

His argument dates back to the 1990s. While others debated open-source licenses, he insisted on the architecture of the system and its capacity to enable participation.

An architecture of participation leaves people in control of their own special sauce, the specific logic that makes a tool useful in their context. An architecture of control keeps that logic with the vendor.

For organizations the difference matters. When people can adapt tools to their own processes, adoption grows organically. When tools remain opaque and constraining, adoption stays superficial and fragile.

The real bottleneck: leadership

The evidence accumulated on this desk points to a board-level figure: just 7% of leaders are ready to govern AI deployments. This figure concerns those who commission, evaluate and regulate adoption, distinct from the employees who use the tools every day.

O'Reilly's reasoning reinforces this reading. When the labs misunderstand people's needs, the leaders' task becomes bridging that distance inside the organization.

An organization equipped with excellent tools and unprepared leadership achieves worse results than one with mediocre tools and literate leadership. The bottleneck lives in the capacity for judgment, far from the technology budget.

Adoption and readiness remain distinct concepts. People adopt rapidly; leaders, by contrast, struggle to build the conditions, the evaluation criteria and the governance rules that make adoption productive.

Converting internal talent beats the external market

A second data point steers talent strategy: converting internal talent systematically beats external recruiting for AI capability. People already inside the organization bring context, relationships and process knowledge.

The documented adoption figures confirm it. Organizations that record high rates of voluntary participation among their own professionals achieve superior performance compared with the isolated hiring of external specialists.

The link to O'Reilly's thesis is direct. A capability rooted in context requires tools that people can shape, and an architecture of participation favors precisely this internal learning.

For the CHRO the priority becomes the design of L&D pathways, internal mobility and skills evaluation criteria. Building capability now reduces the skills debt that would otherwise become an acquisition cost within a few quarters.

Botsitting as a design failure

A third signal deserves attention: the passive supervision of AI output absorbs hours of weekly work in many functions. This phenomenon, called botsitting, measures a wrong workflow design.

When people passively monitor output, the problem lies in the process, far from the model's limits. AI has been inserted into a flow designed for humans, without a redefinition of roles.

This is a failure of change management, distinct from a technological failure. The remedy requires redesigning the steps, clarifying where human judgment adds value and where the machine operates autonomously.

O'Reilly offers the right frame. A clean separation between model, harness and application makes it possible to rebuild the workflow around people. Process design, therefore, precedes the choice of tool and determines its real performance.

The design question for decision-makers

The evidence converges on an operational conclusion for every top role. The technology exists; the decisive variable remains the organizational condition designed by leaders.

Here is what changes for those who decide:

• CEO: bring to the board a conversation about organizational readiness, distinct from adoption rates.
• CHRO: set the priorities of L&D, talent and organizational design around internal conversion.
• CFO: assess the investment in developing people as a line item with documentable ROI.
• Talent & Compensation Committee: monitor human capital metrics, from leader readiness to skills debt.

The distance between what the labs build and what people use defines the governance challenge described by this evidence. To explore further analyses on human capital, the blog section remains useful.

The final design question stays direct: what conditions are leaders creating today so that people can shape AI around real work?