AI is not 'stagnating in performance.' The battle is in B2B 'integration' and 'operations'
Discussions about generative AI tend to lean toward the idea that 'model performance has plateaued.' However, Winston Weinberg, CEO of Harvey (Legal AI), emphasizes the explosion happening in B2B (especially enterprise) rather than the sense of stagnation in the 'consumer' space. Using his statements as a foundation, I will organize the 'practical points' for current AI application companies to win.
1. 'Model performance stagnation' is a misunderstanding—what is needed is 'context' and 'connectivity'
Weinberg states that while 'performance may seem to have hit a ceiling in consumer use cases,' the important thing there is not further reasoning power in the model, but connectivity to calendars, business applications, and workflows—in other words, 'context.' To put it another way, what users are looking for is not a 'smart chat,' but a workflow that gets the job done.
This is where it counts in B2B. Companies do not run on 'a single AI.' Average business operations cross multiple systems and involve permissions, auditing, and sharing. Therefore, the path to victory is shifting from model differences to product design that delivers value through 'integration and operations.'
2. Enterprise AI will 'explode with a 3-5 year delay'
Weinberg estimates the time lag until significant productivity improvements become apparent in companies to be '3 to 5 years.' The reason is simple: corporate work is fragmented across 'not just 17, but 50 systems,' making integration difficult for agents to complete tasks 'from start to finish.'
This time lag is not a pessimistic factor, but rather a tailwind. The capability is 'already there,' but the organization has not caught up—meaning there is a large capability overhang (underutilization). The moment adoption begins to progress, expectations will jump all at once.
3. 'Multiplayer' is the next battlefield—what Shared Spaces symbolizes
The 'Shared Spaces' launched by Harvey is a 'collaboration layer' where law firms and clients (companies) can work together in the same space. It appears to be a move aimed not just at adding features, but at shifting from 'individual productivity tools' to an 'industry OS.'
What is interesting is that, beyond the initial assumption (company ↔ law firm), it becomes a 'hub' where legal departments within a company can collaborate with HR, compliance, procurement, and others. Because legal documents are at the core of management, horizontal expansion is easy—this is the reason why the 'ceiling' for SaaS value is rising.
4. The essence of scale is moving from 'hero' to 'machine'—the pain of deciding P0
In the early stages of a startup, all-encompassing intervention by the CEO, such as 'clearing Slack every 15 minutes,' is effective. However, Weinberg admits that this becomes poison when scaling, and says that to focus on P0 (top priority), one needs to say 'painful no's.' Prioritization is about discarding good ideas—it is correct only when it feels 'as painful as a breakup.'
This also applies to products. While AI companies are prone to 'winning demos,' if they cannot keep their promises due to infrastructure shortages as customers increase, their GRR (Gross Retention Rate) will collapse. The true value of a company emerges in the phase where operations become more important than appearance.
5. The secrets of negotiation and hiring—'movement' is not 'progress'
His view on deals is thrilling. First, 'movement is not action.' He concludes that just because someone is loud does not mean they are leading, and that it is ultimately a game of 'reading people.' Second, there are situations where you must 'decide not to negotiate.' If you understand the value better than anyone else, go for the 'one point you want' rather than playing tug-of-war over terms.
The same applies to hiring: 'If they are top-tier talent, hire them on the terms they want.' Cutting corners by haggling over a few thousand dollars is a loss in the long run.