How Grab is rebuilding its engineering culture around AI speed

Suthen Thomas Paradatheth, chief technology officer at Grab, argues that artificial intelligence makes writing code cheap while making checking code rare.

This change forces leaders to rebuild management, hiring and physical operations in South-east Asia before speed causes problems.

The hidden costs of building software fast

As Grab to releases new software tools at a record pace, leaders must now figure out which numbers show a better business instead of just a busy staff.

Old ways of measuring work can push companies toward the wrong goals, meaning executives can no longer rely on legacy metrics to track actual output.

“90 per cent of our engineers are using some form of AI coding assistance daily,” Paradatheth notes. “We did not mandate anything. We made the tools available, taught people the skills on how to use them effectively, and then cut them loose.”

When output grows but work gets harder to measure

“Engineering productivity always has a bunch of caveats and asterisks,” he reports.

Still, the numbers tell a story. Using merge requests as a proxy, Grab has seen around a 40 per cent increase in output per person, with turnaround time for similar-sized tasks dropping by 20 to 30 per cent.

He adds that engineering output represents only a small piece of the puzzle, noting that true AI effectiveness requires company-wide institutional change alongside individual improvements.

Changing team roles and the risk of easy coding

Giving more people the ability to build software speeds up work, but it raises new questions about hiring and oversight. “Software engineering fundamentals still matter,” Paradatheth says.

SEE ALSO

Grab seeks AI edge in product development using agentic engineering, vibe coding

On hiring, he draws a line. “You cannot come in and say, ‘I could ask any agent to do this, but I do not know what it has done.’ We also look for AI fluency and a sense of ownership, acting like an owner rather than waiting for instructions.”

Lines between departments fade when non-technical staff can automate their own work

The legal team built an automated tool to slash first-pass NDA reviews from hours to minutes, while the design team created a similar tool called Mosaic to generate brand-specific illustrations in a fraction of the usual time.

“If a tool is going to go out in production, if our end customers are going to be exposed, then it needs a review by production engineers,” he explains. “For internal use, we do not want engineering to be gatekeepers.”

Keeping humans accountable for independent systems

Removing internal gatekeepers speeds up development, but it shifts the pressure elsewhere. Weak checking mechanisms risk turning high engineering output into fragile systems.

To combat this, Grab structures its workflow around four operational steps:

• Change engineering work from typing direct prompts to handing off tasks, allowing systems to operate on their own
• Organise business and technical data cleanly so independent systems can pull context without crashing
• Expand checking processes with automated testing to prevent software bugs and verify unexpected behavior
• Use secondary models to check outputs, monitor the automated judges, and assign final responsibility to a specific human.

Software creation changed when systems stopped waiting for human typing

This workflow shifts how software is generated. An engineer now assigns a task to an agent and steps away, returning later to review the generated code, which amplifies speed through asynchronous workflows.

“Code can be generated in reams. The new bottleneck is people cannot review all the code that is created,” Paradatheth warns. “Ultimately, we say you are accountable for what gets to production. So, how do you review it?”

To manage this, the company invests heavily in “harness engineering,” keeping the codebase legible for AI while scaling automated tests to catch errors before they reach production.

“The accountability chain does not end with an AI; it ends with a human being,” he says. “Every leader in my organization is expected to deliver a change to production using agentic engineering.”

Building rules directly into the software

Keeping humans accountable for checking code relies on controlled platforms, presenting leaders with the challenge of giving computing power to staff while maintaining oversight.

Grab’s answer is GrabGPT, which grew out of a failed internal chatbot. When AI excitement surged in 2022 and 2023, external tools posed a clear security risk, so the infrastructure team built a secure in-house interface instead.

“Despite the name, GrabGPT is not just using one vendor,” Paradatheth says. “You can use closed models like Gemini, Claude, and GPT, and open-weight models like Qwen.”

GrabGPT also acts as a router and abstraction layer, with an audit log, controlled onboarding, usage metering, and cost control.

Finding hidden software problems faster

Using this centralised router moves the security focus away from individual tool choice and toward overall system behavior, preparing leaders for an environment where structural weaknesses can grow much faster.

“A useful mental model is that AI is an amplifier of everything,” Paradatheth says. “If you have great software engineering practices, it will amplify that. If there was latent risk in your system, that risk is now amplified.”

Unpredictable system behaviors can expose weaknesses that used to be hidden

“You need to start thinking in the context of latent risk,” Paradatheth warns. “Vulnerabilities may have always been there, but now you have an agent that acts in a non-deterministic way. You cannot ship code into production and hope for the best.”

That same power, however, can work in reverse, automated reviews can now spot and fix weaknesses far faster than human reviewers alone.

Fixing tiny delays in physical operations

This bandwidth for risk assessment extends into the physical logistics network connecting businesses to consumers, where robotics is seen as a tool to eliminate wasted operational time, not replace human workers.

“A robot could be loaded at the counter, meet the driver at the curb, and on the other end, deliver straight to the door, handling the first and last meters of the journey,” Paradatheth says.

Since these walking stages consume roughly ten percent of a driver’s time, removing them lets couriers complete more orders and increase their earnings.

Adapting autonomous systems to varied city streets

While robotic extensions can fix walking delays, achieving full self-driving capabilities requires navigating economic and regulatory conditions that vary drastically between cities.

“For autonomous vehicles, it is going to be a journey before it expands more widely in Southeast Asia,” Paradatheth explains.

The road ahead is complex. “Challenges include unit economics and adapting to the diversity of conditions. Cities are packed with motorbikes and bicycles, and there often are not dedicated bicycle lanes.”

For autonomous delivery robots, the company builds everything in-house. For passenger vehicles, it takes a multi-partner approach to integrate them smoothly into the existing marketplace. TECH IN ASIA