Enterprise AI Trend Forecast: 8 Transformations in the Next 6 Months
In the podcast 'Human in the Loop,' which covers the front lines of enterprise AI, Sam Denton, Ben Shariffstein, and Felix Sue predicted the trends in AI utilization that may occur over the next six months. This article organizes their discussion and provides an easy-to-understand explanation, incorporating professional insights, specific examples, and quotes.
1. Shift in Evaluation Metrics: From Text to Action
1-1. Conventional Evaluation Methods
Until now, 'text quality evaluation' has been the mainstream for measuring AI effectiveness in enterprises. Sam points out, 'Currently, 90% of evaluation is focused on text, and 10% on action.'
1-2. Implementation Examples of Action Evaluation
Sam predicts, 'In six months, it will be completely reversed, with 90% focused on action evaluation.' For example, a workflow where an agent asks, 'Should tool call A or B be executed?', the chosen action is automatically performed, and the effectiveness is measured by subsequent production metrics (such as user behavior results on a BI dashboard). Establishing such an evaluation infrastructure will be the key to improving AI performance.
2. Redefining Work-Life Balance
2-1. Return to 9 to 5
Sam expresses hope that 'by delegating overnight jobs to agents, employees will be freed from the situation of being glued to their PCs, leading to a revival of the 9-to-5 workday.'
2-2. The Dilemma of Usage
On the other hand, Ben cites a real-world example, noting that 'because code can be generated in a short time, people end up taking work home,' mentioning the possibility that AI utilization may not necessarily lead to improved work-life balance.
3. Sandbox Environments and Reinforcement Learning Data Collection
3-1. Building Enterprise RL Environments
Sam's final prediction is that 'leading companies will build sandbox environments to collect continuous learning (RL) data for agents.' Mechanisms that test real operational data on reproducible clones to acquire optimal action sequences will become widespread.
3-2. Pre-requisites for Data Integration
However, Felix points out, 'The first priority is to integrate scattered data and establish interfaces that AI can access.' He cited the Bonobos case to emphasize the importance of unifying diverse data sources.
4. Stagnation of Investment in Mid-Market Companies
4-1. The Temporary 'AI Give-up' Segment
Felix sounds an alarm, noting that while big tech and startups continue to invest, 'there will be cases where mid-market companies, finding it difficult to see ROI and seeing no growth in usage after introducing chat interfaces, will decide to quit AI.'
4-2. Opportunities for Re-challenge
However, they also predicted that this is merely a 'temporary pause,' and that efforts will resume once interfaces are refreshed and data is prepared.
5. The Future of Asynchronous Agents
5-1. Discovering the 'Three-Quarter Mile'
Felix compared chat-based AI to a 'quarter-mile' sprint, stating that 'asynchronous long-duration task execution by agents is the three-quarter mile, which will bring about the next leap forward.'
5-2. Large-Scale Horizontal Scaling
Ben also explained the advantages of dynamic scaling, noting that 'parallel processing of 10,000 agents from a single instance allows for deep research and complex tasks to be completed in a short time.'
6. Dramatic Reduction in Model Costs
6-1. Distillation and GPU Efficiency
Ben pointed out that 'prices are falling rapidly due to model distillation and improved GPU utilization, making tasks that were previously too expensive now economically viable,' emphasizing that 'we should focus on the value AI brings rather than worrying about costs.'
6-2. Shifting Bottlenecks
It was also noted that 'as LM costs decrease, other infrastructure expenses, such as search APIs, may become the new bottlenecks.'
7. Full-Scale Agent Adoption: From Efficiency to Capability
7-1. From Information Retrieval to Business Automation
While traditional chat has been an 'information retrieval tool,' the future demands the ability to automate and augment entire jobs as 'business automation tools.'
7-2. Rethinking UX Design
With this shift, 'new interface designs that intuitively demonstrate agent capabilities, rather than just chat UIs, are essential.'
8. The Rise of Vertically Specialized AI Companies and Corporate Choices
8-1. Buy vs. Build
Ben pointed out that 'AI startups specialized in specific industries, such as legal and customer support, are increasing, forcing companies to decide whether to buy or build.'
8-2. In-house Development as a Differentiation Strategy
Market leaders are considering the development of proprietary agents that leverage their own user data to differentiate themselves from competitors, concluding that 'strategic choices aimed at industry leadership, rather than just being above average, are essential.'
The next six months will be a period of significant movement in enterprise AI implementation strategies, ranging from a paradigm shift in evaluation methods to cost reduction and the full-scale adoption of agents. Based on the forecasts in this article, let's prepare the optimal data infrastructure and UX design to push AI utilization to the next stage.