As Artificial Intelligence reshapes how engineering and operations teams work, the real challenge for most organisations becomes ensuring they have the visibility and controls needed to operate it with confidence.
Yadi Narayana, field CTO, Asia-Pacific & Japan, at Datadog, discusses how organisations can identify blind spots in AI adoption, build trust, and make wise investments to prepare for the future.
Common blindspots
According to Narayana, one of the most common blind spots is assuming that whatever happens in the pilot will also translate to production.
“When we trial something, it’s usually in a controlled environment; it has a limited number of users. We carefully select the type of data we want to try with and we also define some of the tasks we want to try,ā he said.
The second blind spot is putting too much focus on the model itself.
Although the model is the most visible part in an AI application, it is only a small part of a much larger ecosystem. The information it receives, the way prompts are constructed, API failures, and problems in the wider application can all influence the outcome.
“The third, which is also something common, is ownership,” said the Datadog executive.
When an AI system produces a wrong and risky outcome, there should be clear guidelines on who will respond.
Building trust through observability
Addressing these blind spots requires more than building better AI models.
For Narayana, to help organisations feel more confident about using AI in production, observability is non-negotiable.
“We should be able to see what the AI is doing. We need to understand why it behaves the way it behaved, and how it intervenes when it moves outside the boundaries we have set,” he said.
He explained that observability provides the evidence to investigate what happened rather than relying on assumptions.
However, visibility alone is not enough.
Narayana underscored the importance of pairing visibility with proper guardrails to define what an AI system can access, what actions it can take, and when human intervention is needed.
“While observability does not create trust by itself, trust will depend on governance, testing security, and accountability,” he added.
While observability does not create trust by itself, trust will depend on governance, testing security, and accountability. Yadi Narayana
A peek into the future
As AI systems continue to evolve, Narayana believes observability itself must also evolve.
“Today, in the majority of cases, a user asks a question and the model produces an answer. In this environment, observability is mainly about understanding the input, the output, the response time, the cost, and the quality of the answer,ā he said. āNow we know AI is already moving beyond that. We are seeing systems that can plan work, select tools, access company data, call APIs, and take some of their own actions.”
He believes the future will involve multiple AI agents working together to handle different parts of a task.
As AI systems become more autonomous and involve multiple interacting agents, observability must evolve beyond monitoring model outputs to tracking how AI systems make decisions and execute actions.
Organisations need end-to-end visibility into the entire decision chaināfrom the user’s request and the AI’s interpretation to the tools and data accessed, the decisions made, and the resulting business outcomes.
This level of transparency enables organisations to investigate unexpected behaviour, enforce guardrails, pause or contain actions when necessary, and ensure AI operates within defined permissions.
Because AI systems continuously change as models, data, prompts, and applications evolve, evaluation must become an ongoing process rather than a one-time pre-deployment exercise.
AI observability may evolve from being more of a technology-specific lens or a technology monitoring discipline into part of an organisation- how they manage and govern their whole digital workforce. Yadi Narayana
“AI observability may evolve from being more of a technology-specific lens or a technology monitoring discipline into part of an organisation- how they manage and govern their whole digital workforce,ā he said.
He added: āIt could even evolve there. From understanding what it is doing, from seeing where the behaviour changes, and it’s a continuous posture; it needs to be done even today,” he said.
Preparing for what’s ahead
With AI becoming increasingly autonomous, Narayana believes organisations should start preparing now.
The first step, he said, is defining what success should look like for every AI use case.
“Many organisations are still relying on broad measures such as accuracy or availability. But a customer service assistant, for example, should be measured on whether it has resolved the issue, used the correct information, protected sensitive data and avoided unnecessary escalation, as an example,” he said.
Second, organisations must build traceability across their AI workflow.
“When something goes wrong, do we have an ability to understand what model, data source, tool, prompt or application component influenced that result?” he said.
He also underscored the importance of establishing more practical guardrails around what AI systems are allowed and not allowed to do.
This includes defining the data they can access, the tools they can use, the actions they can perform independently, and the situations where human approval is required.
The fourth priority is ownership, which involves ensuring that all teams agree on who owns the use case, who defines the boundaries, and who responds when a system behaves unexpectedly.
“We would also avoid treating AI operations as a completely separate discipline, because wherever possible, the capability should connect with our existing engineering, security, risk and governance practices,” he said.
Wise investments
Given limited budgets, Narayana believes that if CIOs could invest in only one capability to minimise AI blind spots, it should be production evaluation for one of their most important AI use cases.
“Most organisations can tell whether an AI application is available, but very few can confidently say whether it is producing a useful, accurate, safe and cost-effective outcome,” he explained.
Rather than launching an organisation-wide programme, he recommends starting small by choosing one meaningful AI application and defining the small number of measures that matter most.
“I recommend this investment because it forces an organisation to answer a very fundamental question: what does working well actually mean in an AI system? Because once that is clear, decisions about everything else- monitoring, controls, model selection, future investment become far more grounded,” he shared.
Bright but not blinding
For Narayana, the future is not about eliminating every possible risk but about reducing the blind spots that prevent organisations from deploying AI with confidence.
The future of AI is as bright as ever, but with proper investments, knowledge and preparation, it will never be blinding. There are ways that organisations can better understand how their AI systems behave, respond when things go wrong, and continuously improve over time.
