Agoda CTO on why AI still needs human oversight

• Agoda CTO Idan Zalzberg says engineers now focus more on reviewing AI output and guiding its use.
• AI coding tools can speed up development, but humans still make the decisions.

Artificial intelligence tools now sit inside many developer workflows. Code assistants, AI agents, and generative models can help write code, suggest fixes, and handle repetitive tasks. Yet even as adoption grows, many engineers remain cautious about how much they should rely on these systems.

Recent industry surveys show that while most developers use AI in some form, many still question whether its outputs are reliable enough for production systems. The tension between speed and trust has become one of the main issues facing engineering teams that want to integrate AI into their daily work.

Trust still develops through experience

Trust in AI tools does not appear overnight. It builds over time as engineers learn where the tools perform well and where they fall short. “Trust is something that needs to be earned,” Zalzberg said. “As engineers use AI daily, they learn where it can be trusted and where it cannot based on their own experience.”

Two factors have helped increase confidence in recent months. One is that the underlying models are improving quickly. The other is that engineers are learning how to work with these systems more effectively.

“First, the models and agents are getting significantly better,” Zalzberg said. “Since we ran the survey, numerous new models and agent updates have been released that deliver better adherence to instructions and greater ability to self-correct.”

At the same time, developers are improving the way they interact with the tools. “Second, our own use of the tools is improving. As we learn how to work with them and how to instruct them in ways that leave less room for ambiguity, we are able to generate more predictable and trustworthy results,” he said.

In practice, this means engineers are becoming more skilled at writing prompts, setting clear instructions, and verifying outputs.

Agoda engineers remain responsible for AI-generated code

Even when AI helps produce code, the responsibility for that code still rests with the engineer who used the tool.

At Agoda, teams treat AI outputs in the same way they treat code written directly by a developer. “First, it’s critical for us to maintain that engineers are accountable and responsible for their work, whether they used AI or not,” Zalzberg said. “AI is just a tool, and engineers within different teams should have the oversight and measures needed to ensure the code it generates is correct and maintainable.”

Any code generated by AI is still tied to the engineer who requested it. “As such, code generated by AI is still attributed to the specific engineer who used AI to produce it,” he said.

Quality control processes also remain the same. AI-generated code goes through the same checks as human-written code before it reaches production. “In terms of practices to ensure AI-generated code is high quality, we rely on the same mechanisms we use for human-written code: linters and other static analysers, strong reliance on automated testing to validate behaviour, and gradual rollouts to catch concerning changes,” Zalzberg said.

Some of those safeguards now also involve AI tools. “Some of these processes are themselves assisted by AI, which helps us apply them at larger scale regardless of how the code was written,” he added.

Unpredictability remains a core challenge

One of the biggest technical hurdles with generative AI systems is that they behave differently from traditional software.

Standard programs produce the same output when given the same input. Generative AI systems do not always follow that pattern. “The biggest challenge when using AI is its unpredictability,” Zalzberg said. “Unlike ‘standard’ software, we can’t always tell exactly what the code will do by simply reading it.”

The challenges go beyond just code generation. “We can’t validate outputs with simple tests when the AI generates free text, and we can’t assume running the same code multiple times will produce identical results,” he said.

This shift has forced engineering teams to rethink how they evaluate software behaviour. “The way to address this is by becoming experts in evaluation, which is a critical skill for AI engineers,” Zalzberg said.

Evaluation methods allow teams to assess generative systems using statistical measurements rather than simple pass-or-fail tests.

“Evaluation is the method by which we assess generative AI systems in a statistical wayallowing us to determine whether changes to our code or prompts are producing better or worse results on average,” he explained.

This area continues to develop quickly. “This is an active field of research worldwide, and we continuously try to track emerging best practices as they evolve,” Zalzberg said.

Engineers focus more on judgement and oversight

The presence of AI in development workflows has not removed the need for human engineers. In many ways, it has increased the need for careful oversight.

Developers may now ask AI assistants to generate pieces of code, but the core role of the engineer has not changed. “It’s true that engineers no longer write all the code themselves and instead often instruct an AI assistant to generate parts of it,” Zalzberg said. “However, that was never the core role of an engineer.”

He pointed out that the central task has always been solving technical problems. “The main responsibility has always been making technical decisions at different levels to address business problems,” he said.

That responsibility remains with human engineers. “While AI can make small decisions on its own, the majority of decision-making still sits with the human operator,” Zalzberg said. “Whether it’s choosing the tech stack, defining the architecture, or designing the data model, the engineer must remain on top of it.”

Even when AI generates code, developers still need to review and challenge its output.

“When AI produces code, engineers still need to challenge it and ensure the output aligns with the technical direction of the project,” he said.

As a result, certain skills are becoming more important. “This requires engineers to develop even stronger critical thinking skills and, to some extent, stronger communication skills,” Zalzberg said.

Engineers must also design the systems that guide AI tools. “Additionally, engineers are responsible for creating environments where AI is more likely to produce the desired results — for example by building the harness or framework within which AI agents operate,” he said.

AI productivity gains take time to appear at Agoda

Many developers report that AI tools save time, but measuring the real impact across large engineering teams can be difficult. Survey results show that developers often save between one and six hours per week when using AI assistants. Yet time saved in code generation does not always lead to faster product releases.

“The report shows that most developers save between one and six hours per week, but productivity gains in code generation do not automatically translate into better outcomes at the end of the pipeline,” Zalzberg said.

Early tests at Agoda showed clear gains.

“When we first rolled out AI tools, we were able to run controlled experiments that showed roughly a 27% productivity uplift,” he said.

Those experiments became harder once AI tools spread across the entire organisation. “Today, however, AI usage is so widespread that we no longer have a clean benchmark for comparison,” Zalzberg said. Instead, teams look at broader signals such as development velocity. “Instead, we track our overall velocity improvements over time, such as pull request throughput,” he said.

The improvements did not appear immediately. “In many areas it took a considerable amount of time before improvements became visible, largely because engineers needed time to properly integrate AI into their workflows and codebases,” Zalzberg said.

Recently, some of those gains have started to show up more clearly. “Recently, however, we have begun to see some material improvements overall,” he said.

Cultural differences influence AI adoption

Agoda operates across multiple Asian markets, and the company has noticed differences in how developer communities approach AI tools.

“It’s definitely interesting to see how cultural differences influence AI adoption,” Zalzberg said.

Some teams prefer to move carefully before deploying new tools. “Some cultures are more naturally risk-averse and prefer to spend more time ensuring guardrails are set up ‘perfectly,’ while others prefer to experiment first and adjust or backtrack if needed,” he said.

A diverse workforce can help balance those approaches. “We’ve always taken pride in our multicultural employee base, and in this case it’s beneficial as well,” Zalzberg said.

Different viewpoints can help teams learn faster. “By trying different approaches and hearing different viewpoints, we learn faster as a team about what works and what doesn’t,” he said.

Listening to both cautious and experimental voices remains important. “The important thing is to listen to all perspectives — both the voices saying ‘slow down’ and the ones saying ‘speed up,’” he said.

Clear frameworks may define AI maturity

As AI tools spread across engineering teams, many organisations are now trying to define what responsible use looks like at scale. Zalzberg believes that an “AI-mature” organisation focuses on two areas: using AI within engineering workflows and embedding AI inside the products the company builds.

“There are two aspects to being an AI-mature engineering organisation,” he said. “One is effectively using AI in the engineering process itself, and the other is embedding AI into products built by the engineering organisation.”

Both areas present similar challenges. “In both cases there is significant value — for both customers and the company — but the complexity of working with non-deterministic systems is substantial,” he said.

Three principles tend to guide that process.

The first is measurement. “Evaluate and measure. Ensure AI-driven processes continue to improve by measuring them objectively and assessing whether changes are producing positive or negative outcomes,” Zalzberg said.

The second is governance, which includes tracking how AI tools are used across the organisation. “Governance. Track AI usage and costs, and understand who is using it and how. At Agoda, for example, we developed a single AI gateway for all models so we can monitor usage centrally,” he said.

The final principle is careful control over where AI makes decisions. “AI-driven decisions should either be reviewed by humans in some form or limited to areas where mistakes have no material consequences,” Zalzberg said.

Some tasks remain too sensitive to automate fully. “You might not let AI decide the price charged to a customer for a room, but you could let it help decide which hotels to recommend first based on user preferences and hotel characteristics,” he said.

That distinction may shape how engineering teams adopt AI tools in the years ahead.

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology eventsclick here for more information.

AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.