From generative AI to agentic AI: Why business value depends on context and control 

AFTER a surge of interest in personal AI agents earlier this year, the focus has shifted from what they can do for individuals to a more serious enterprise question: Can agents be trusted to act on behalf of a business? Agent-building tools such as OpenClaw have shown how users can build AI agents to take on tasks ranging from planning a day’s schedule to writing complex software code. But for organisations, the stakes are higher. 

Beyond individual use, organisations are now exploring how agents can be deployed across business workflows to drive business results, whether it is to improve productivity, strengthen operations and deliver measurable results.

At its core, an AI agent is a system designed to act. It can assess information, reason through a task, plan the next steps and execute them towards a defined goal.

For organisations, this marks the shift from generative AI to agentic AI – from systems that generate or retrieve answers, to systems that can take action in ways that affect customer experience, operations and business outcomes, says Ajay Nair, general manager for Elasticsearch at Elastic.

Focus on impact, not activity 

The appeal of AI agents is clear, but their usefulness depends on where and how they are deployed. Instead of rushing to deploy the technology, Nair says there are three key factors that businesses need to consider before they adopt agentic AI.

First, organisations need to be clear about the value they want to create. Agentic AI can deliver different outcomes depending on where it is applied. Used in product innovation, for instance, it may help teams move faster from idea to iteration.

Applied to observability or security workloads, it may instead support stability by helping organisations manage risks, detect issues and build resilience. 

Second, businesses need to identify the workflows where agentic AI can be applied effectively. The technology is best suited to processes with clear rules, repeatable steps and well-documented playbooks. This is why many early implementations have started in engineering workflows, such as software programming. Other use cases may require more groundwork before they are ready for agentic AI. 

Third, organisations need to define how success will be measured. It is not enough for an AI agent to generate more output or complete more tasks. In customer support, for example, an agent may produce hundreds of support articles, but if customers’ issues remain unresolved, the business impact is limited. 

Businesses, Nair said, should not confuse activity with impact. Moving quickly on AI may create the appearance of progress, but speed alone does not translate into business value.

For many businesses, excessive experimentation without clear direction has become a costly and frustrating exercise. This often stems from insufficient strategic thinking around AI adoption, says Nair.

The starting point, then, should be a concrete use case and clear measures of success, so that AI initiatives are tied to business value rather than innovation for its own sake. 

Can AI agents be trusted?

Once a viable use case has been identified, the next hurdle is trust: Can an AI agent arrive at sound conclusions and take responsible action?

For organisations, trust depends on the quality and relevance of the results an agent produces. That requires accuracy, transparency and, crucially, the right context, explains Nair.

The risks are far higher in a business setting than in casual use. An error in approving a loan or processing an employee reimbursement, for instance, can have serious consequences. “This is unlike using ChatGPT casually for planning a holiday,” he added.

The concern is not merely theoretical. As AI agents become capable of taking actions rather than simply producing answers, questions of accountability, control and oversight become more pressing.

Countries around the world including Singapore have also been paying closer attention to trust in AI.

The Ministry of Digital Development and Information launched a model AI governance framework for agentic AI earlier this year, aimed at guiding organisations in deploying agents responsibly. The framework recommends technical and non-technical measures to mitigate risks, while emphasising that humans remain ultimately accountable.

For businesses, trust starts with relevance: whether an AI application has the right context to perform a task well. That means designing an AI stack that can draw on the appropriate data, retrieve the right information and generate insights that are fit for purpose. As Nair puts it, it is hard to have a smart agent if the data behind it is poor.

Guardrails are equally important. These include controls over the types of data an agent can access, limits on its read-and-write privileges, and human oversight to ensure that proposed actions are safe, legitimate and aligned with business rules and regulations.

“There are certain tasks, like coding, that we’d trust AI more with, but in general, people need to know that the LLMs they are working with are pulling the right insights from the data they are being provided, and whether the data is relevant to begin with,” Nair adds.

Organisations therefore need to integrate AI solutions securely with their proprietary data, while ensuring that these systems can retrieve relevant insights to solve specific business problems, he notes. “That’s the essence of context engineering.”

Context engineering is the practice of giving AI systems the right information at the right time. For an AI agent, that means having enough relevant, timely and trusted context to make better decisions and take appropriate action.

From context to accountability

The challenge, then, is not simply to have more data, but to make the right data usable at the moment an AI agent needs it.

Context engineering requires a strong data foundation, though organisations do not necessarily have to build every component from scratch. Even those without mature data architecture can get agentic AI applications running within months if they choose tools that fit their needs, says Nair.

The key is modularity. Modern AI tools allow organisations to assemble the components needed for agentic workflows, including data ingestion, retrieval, security, governance and evaluation.

Elastic’s search capabilities, for instance, are designed to understand the intent behind queries, retrieve relevant insights from stored data and connect those insights with large language models.

These capabilities are brought together in Elastic’s Agent Builder, which is aimed at helping organisations implement agentic workflows more quickly while retaining flexibility over the components and models used for specific tasks.

But deployment is not the finish line. AI agents need to be monitored and evaluated continuously after they go live. Each action or change made by an agent should be logged, checked for validity and traced back to the data it accessed, adds Nair.

That makes security and observability essential to agentic workflows. With the right capabilities in place, organisations can monitor what an agent did, verify whether it had permission to act, and ensure that its actions were appropriate, traceable and aligned with business rules.  Find out more about how Elastic drives agentic AI.