In partnership with Elastic
AI is progressing so fast that each month or two, models bring change and deliver better performance than the last. Frontier AI models are on the horizon with increased potential to draw insights from large amounts of data, and identify vulnerabilities in software.
And indeed, organisations are getting into the thick of AI. Asia-Pacific leads the way in AI adoption and experimentation, with 78 per cent of respondents in a recent BCG survey saying they use AI at work at least weekly, compared to 72 per cent worldwide.
However, the research firm also found that only 57 per cent of respondents in Asia-Pacific say their organisations are transforming workflows to accommodate the shift with AI. That begets the question: with such rapid development in AI, how can an organisation ensure their innovation is driving value?
Waiting for a flawless data foundation – where every silo is cleaned and synced – is a trap, as such a perfect state rarely materialises in a working business, Ajay Nair, general manager of Elasticsearch at Elastic, told Techgoondu in a recent interview.
For companies and IT teams, sitting on the sidelines is no longer an option. AI adoption is becoming a multi-cycle journey similar to the shift toward cloud-native architectures, where organisations must carry their legacy data forward while trying to innovate.
Where does agentic AI fit into an organisation?
Meaningful progress happens when the focus shifts from individual tools to the system, as a whole, pointed out Nair.
He suggests identifying the quantifiable gains that an organisation would gain from the adoption of agentic AI. Depending on where the technology is applied, different benefits can be found, ranging from increased speed to complete tasks, faster innovation and go-to-market with new products, or increased security and stability of applications.
Once the desired benefit is identified, look at “surfaces” where the rules of engagement are already well-defined, such as DevOps or Site Reliability Engineering (SRE) workflows, and use AI to accelerate these predictable processes. For example, an SRE can use agentic AI alongside observability processes that gain insights from logs, and metrics to identify and address bottlenecks in IT infrastructure.
How can an organisation build a solid foundation for agentic AI?
Nair suggests that foundation building should happen alongside AI efforts rather than being treated as a separate, prior task.
The most effective way to gain momentum is to start small and move quickly. By identifying a singular, bounded dataset or a specific workflow, such as simplifying the onboarding process for new engineers, teams can secure the early wins needed to build organizational confidence.
Agentic AI isn’t a standalone application, but a pipeline involving multiple components, ranging from data ingestion, data retrieval for relevance and scale, context engineering, controls, and large language models (LLMs).
As part of this pipeline, ingesting relevant data at scale is critical to ensure quality output from an AI agent. Context engineering is a collection of practices that can be combined, bringing data from disparate formats and sources across an organisation’s IT infrastructure to the LLM, helping them get a full picture behind and ask to arrive at a decision.
Another part of the pipeline that’s important is ensuring trust. For a developer, that would mean having controls in place to ensure that an AI agent is only privileged to access the minimum amount of data that it requires to perform a task.
At the same time, the AI agent should only be able to have limited capability for action; or require a human in the loop in its initial days, so its decisions are checked before any action is taken.
Quality data that pertains to a certain task will be required for AI to understand the background behind a task and act on it – for example, receipts and customer response playbooks in the case of resolving customer queries.
Instead of a massive, daunting overhaul, organisations can use agent builders – applications that connect prepackaged components that constitute the “AI stack” like data ingestion, analysis, and guardrails, refining their infrastructure as they go.
How do we determine the value that agentic AI drives?
This approach requires a change in how return on investment (ROI) is measured. It is easy to mistake high activity for progress, such as using AI to churn out thousands of knowledge-base articles.
Real impact is found in outcomes, like whether those articles actually enable support engineers to close customer tickets more efficiently.
Experimentation is important, said Nair, but there has to be a systematic way to adopt AI. “Because your gains on velocity come when your system becomes better. Not if your individual pieces get better.”
There are several other ways to measure agentic success, including agent cost per completed task (ACCT), which normalises costs by calculating the total expense required for a successfully completed task, no matter how complex; and effective context utilisation (ECU), a composite metric combining task success rate and accuracy relative to cost, ensuring agents operate efficiently and reliably.
What else helps make agentic AI implementation a success?
The strategy for success is one that plans for flexibility. LLMs are rapidly developing and frontier models are just over the horizon, said Nair, which means it’s important to avoid vendor lock-in and choose open platforms that can adapt as the landscape shifts.
To get up to speed, anyone can use a three-step approach.
First, start small and start quickly by identifying a singular, bounded dataset or a focused workflow where AI can prove its value in a contained environment. This reduces the risk of failure and allows the team to gain the hands-on experience necessary for larger deployments.
Second, plan for flexibility by avoiding a singular, rigid end-to-end stack. Since the industry is changing so rapidly, it is vital to pick a platform that offers features to mix and match agentic AI capabilities that is open and flexible enough to swap models or integrate new data sources as they emerge.
Finally, be curious. Successful workers will be those who stay engaged with the technology’s evolution and don’t let the fear of technical debt prevent them from experimenting in real-time.
“Pick a singular data set or a singular workflow where you think AI can help and start experimenting,” said Nair. “That is the best way to go faster.”
