The success or failure of enterprise AI depends not on the model but on the 'context'... 7 conditions for the age of intelligent agents

Enterprise-level artificial intelligence (AI) is moving beyond the “experimental” stage and transitioning into practical business deployment, but on the ground, results still often fall short of expectations. Industry experts point out that the reason is not a lack of better models, but a lack of “context.” No matter how outstanding an AI agent’s performance is, if it cannot properly acquire internal corporate knowledge and business context, it will stall at the decision-making stage.

Appen Ltd. Chairman Vanessa Liu recently stated at a joint event with the New York Stock Exchange (NYSE) on theCUBE: “Data is crucial for enterprises to leverage AI. Just as even the best employees need organizational training to adapt after onboarding, AI agents must have access to business context to operate normally.” Senior executives from data infrastructure, finance, enterprise modernization, and open-source AI fields, including Thomson Reuters Corp. CEO Steve Hasker, attended the event to discuss “how to truly implement agents into actual business operations.”

1. Corporate competitiveness ultimately depends on “exclusive data”

Speakers unanimously agree that relying solely on cutting-edge AI models cannot achieve differentiation. The core lies in the internally accumulated long-term data and business knowledge of the enterprise. Chairman Liu pointed out that the company’s proprietary expertise is often not systematically organized. CEO Hasker believes that future competitive agents will no longer depend solely on “how good they are,” but on whether they “possess a defensible data moat in the market.”

2. Users won’t wait, and neither will agents

Speed is now regarded as a basic requirement rather than an option. Ariel Schulman, Chief Product Officer of Bright Data Ltd., explained that when users see “searching the web” on a chatbot screen, their patience timer starts. Bright Data currently provides web scraping data as a response starting point for chatbots, keeping page transfer times under 1 second, with a median of 500 milliseconds. Due to slow data retrieval, users may abandon the agent before it has finished preparing a reply.

3. AI agents handling money need “accounts” and identity systems

Some suggest that if AI agents are to perform financial actions like payments or transfers, they require a certification system similar to human ID documents. Sean Neville, co-founder and CEO of Catena Labs Inc., said banks must be able to verify who the agent represents, what it can do, and why it takes certain actions. This concept aims to ensure accountability and traceability in financial automation through a so-called “Know Your Agent” system.

4. Beyond “vendor lock-in,” now “token lock-in” is the real issue

A warning was also issued: building all systems entirely around a specific AI model could lead to losing cost control in the future. Woodson Martin, CEO of OutSystems Inc., pointed out that companies relying on a single cutting-edge model will face profitability pressures as inference costs accumulate. He emphasized the need for a platform layer that allows replacing models in production without rewriting underlying systems, which is a practical solution for profit and loss management in agent strategies.

5. “Providing” AI tools is not the same as truly “making people use” them

There is a significant gap between actual application on-site and management’s perception. Tye Kim, CIO of WalkMe Ltd., said that 80% of management believe they provide excellent AI tools for employees, but only a small fraction of employees agree. The issue is not the quantity of tools but whether they can naturally present functions at the right moment. Without “context-based guidance” integrated into business processes and at the moment of need, the effectiveness of AI investments will be greatly diminished.

6. Start with the strongest model, then look for cheaper alternatives

Some argue that prioritizing cost reduction is a strategic mistake. Wu Qingyun, representing AG2ai, said that one should first confirm the achievable level with the most powerful model, then compare cheaper open-source models or alternatives to see if they can deliver the same performance. This means that if a company sets a low bar from the start with cheaper models, it may miss out on the capabilities it truly needs. Only afterward can a balance between cost efficiency and performance be achieved.

7. Pilot projects are easy, but operational phases are most prone to failure

The greatest risk is not in pilot projects but in real-world deployment. Bar Moses, co-founder and CEO of Monte Carlo Data Inc., explained that many agents that perform well in initial proof-of-concept (POC) phases tend to encounter issues after deployment, such as referencing outdated data, skipping reasoning steps, over-consuming tokens, or generating hallucinations not caught during testing. Especially since courts have ruled that ultimate responsibility for agent actions lies with the enterprise that created the service, establishing control and monitoring systems becomes even more critical.

Final assessment indicates that the next round of enterprise AI competition will depend less on model performance itself and more on “how accurately they can provide context and how reliably they can operate.” As AI agents increasingly replace human work, proprietary data, internal knowledge, speed, cost control, and accountability structures are likely to become the key factors determining success or failure.

TP AI Notes: This article is summarized based on TokenPost.ai’s language model. The main content may be omitted or differ from the actual facts.