AI adoption is now widespread across the region. On the developer front, Gartner predicts that by 2028, AI coding costs will surpass the average developer’s salary as token consumption surges.
Gartner further warns that if we continue without a governed engineering operating model, the cost of AI tools will continue rise faster than the productivity gains they are intended to deliver.
What will the widening gap mean for the industry and developers?
While we acknowledge that AI has become a regular part of software development workflows, many engineers remain cautious, seeing AI as an accelerator rather than a dependable substitute for human judgment.
In an interview with iTNews Asia, Amit Choudhary, Global Chief Operating Officer at Thoughtworks, agrees, making the argument that while AI may make code faster and easier to produce, organisations still need engineers who hold a myriad of roles. They have to decide when AI should be used, choose the right model, manage costs and autonomy, design scalable systems, and take responsibility for security, quality and business outcomes.
iTNews Asia: A recent McKinsey study shows that software developers can complete coding tasks up to twice as fast with generative AI.
Has AI-assisted development matured from an experimental novelty to an accepted practice today?
I also wonder if AI-coding tools and assistants can really deliver the productivity gains with minimal or no human oversight?
AI-assisted development has decisively won the adoption argument, including in APAC and the productivity improvements at the task level are real. To illustrate this, an Agoda study found that 95 percent of developers across Southeast Asia now use AI weekly.
That case is settled. The trust argument is not.
McKinsey's claim of 2x coding speed measures generation velocity, which gives a misleading signal. Enterprise value comes from delivery velocity not generation velocity. At Thoughtworks, our experience as Client Zero, where we made build-versus-buy decisions across internal tools and enterprise platforms, taught us that the unit of change is the end-to-end process, not the point tool.
Speeding up code generation by 50 percent is irrelevant if security reviews, architectural integration and deployment pipelines remain choked. That is the operational failure. The strategic failure is deeper: a technically sound solution to the wrong economic problem is not productivity. It is precision waste.
For instance, in Singapore, Agoda found that 77 percent of developers cite inconsistent AI outputs as their biggest barrier and two-thirds always review AI code before merging. Accelerating point-tool code generation without continuous architectural oversight does not create business value; it simply manufactures software defects at scale.
The result is Brownian motion across the organisation: localised kinetic energy with zero net forward progress for the business.
iTNews Asia: Does faster code necessarily mean cheaper or better software?
Faster code is necessary to stay competitive, but not sufficient on its own. AI reshapes the speed-cost-quality trade-off by accelerating delivery and lowering initial build costs. However, without engineering governance, it creates a dangerous illusion of progress.
Evaluating true financial impact requires Total Cost of Ownership thinking across three phases: pilot, scale and long-term maintenance. Unchecked AI code generation is a high-interest financial loan. It gives you cheap build speed today in exchange for unpayable comprehension debt tomorrow.
Architecture review, security verification and deployment demand non-negotiable human engineering judgment. Independent audits show raw AI-generated code contains 2.7 times more security vulnerabilities than human-written code. Singapore's regulators (MAS) have reinforced this through mandated end-to-end lifecycle controls. Enterprise software demands deterministic safety.
When comprehension debt compounds, maintenance debt accumulates. Systems do not degrade gradually. They break suddenly. Think of it this way. When bread becomes cheap and abundant, the premium shifts to butter. AI is doing exactly that: making code generation nearly free.
The engineers who command a premium are not the ones who type the most code. They bring judgment, architecture and governance that AI cannot replicate. Faster code without that judgment does not make software cheaper or better. It only makes failure faster.
-Amit Choudhary, Global Chief Operating Officer at Thoughtworks
iTNews Asia: With companies continuing to spend on AI, how do you see this impacting how organisations hire and use tech talent? Do you see this as a boon or bane for engineers and developers?
AI is an extraordinary boon for true engineers but an existential threat for pure syntax typists, because the talent challenge is not a tech supply problem, it is a business absorption problem.
In my client conversations, I emphasise that the constraint has shifted. The CIO's engineering team is no longer the bottleneck; they are ready to move fast. Today's constraint is the CXO and business leaders who lack the readiness to absorb tech velocity. Yesterday, enterprises stalled because IT was too slow. Today they stall because business leadership cannot iterate and scale fast enough.
This elevates 'transmissibility': the human capability to articulate a clear mental model, frame business context and specify system constraints so an AI agent can execute. It is the difference between an engineer who says the API is ready and one who explains why the downstream process must change before it delivers value.
The talent struggle in Asia is not about finding people who write prompts. It is about finding engineering leaders who bridge that gap, translating technical delivery into actionable business terms.
As the Dean of NTU's College of Computing and Data Science noted, 'Artificial intelligence is a multiplier. But if the multiplicand is zero, the outcome is zero.'
The real question for Asia is not how many people can use AI. It is how many can give AI something worth multiplying.
iTNews Asia: If we look at how we can give AI something worth multiplying, how then should the developer’s role change or evolve?
The developer role is not evolving gradually. It is being structurally redefined. Singapore's Minister for Digital Development, Josephine Teo, puts it well: developers are no longer just typing code, they are becoming 'orchestra conductors' who frame problems, validate outputs and direct overall system performance.
At Thoughtworks, we frame this through a clear operational taxonomy: citizens build, agents execute and experts govern. Business analysts use AI to draft user stories, reducing turnaround significantly. QA engineers move quality controls upstream by pre-generating automated test scenarios, while Tech Leads govern architecture, eliminate redundant AI outputs and manage token cost-efficiency at scale.
I have found that successful role evolution requires operating at the intersection of three forces: engineering, economics and psychology. The engineering and economics shifts are visible and measurable. The psychology shift is harder and more consequential. Of all the psychological challenges, stakeholder selection and management is where most AI transformations succeed or fail.
The hardest part of role evolution is not teaching developers AI tools. It is training engineering leads to identify business stakeholders who are genuinely ready to lean in and sustain that alignment when resistance surfaces and the original business case gets tested. If you do not get this right, even the best AI engineering team will fail to create value.
iTNews Asia: How can businesses determine whether AI coding tools are delivering genuine value?
Value is determined by business challenges being solved, with or without AI. That is the economics lens, the only one that matters for this question. To capture genuine value, leaders must follow three operational rules.
- First, project selection must start with economics, not engineering. Engineering is rarely the binding constraint in enterprise AI projects today, but most organizations start with the engineering, not the problem. Extending Antoni Gaudí's philosophy that design without functionality is an extravagance, I would say that engineering without economics is a hobby. Leaders must evaluate AI projects by asking: does this fundamentally alter our cost structure, revenue model, or time-to-market?
- Second, measure end-to-end process outcomes rather than isolated task speed. GovTech Singapore found that a 24 percent gain in writing speed yielded only a 12 percent overall productivity improvement. Measuring speed at the point-tool level systematically halves enterprise ROI.
- Third, the build vs. buy decision is a leading indicator of whether value is being created or destroyed. At Thoughtworks, we built proprietary agents where our model required distinct IP and deployed externally where time to value mattered more. Getting this decision wrong destroys value before you write a line of code, especially because now the threshold for justifying a build is lower than most organizations realize.
If executive dashboards are celebrating token consumption or lines of code written, they are measuring noise. Genuine value is realised only when AI changes what your business can do, not just how fast your engineers can type
iTNews Asia: Let’s look at Singapore as an example, is the AI talent challenge in the country more about developing engineers, and not simply producing more coders?
Singapore’s challenge centers on developing engineers, not producing coders. As a high-cost innovation hub, the country cannot compete on labor volume; it must lead on engineering governance, system architecture and strategic value. That structural reality makes Singapore’s talent challenge unique when compared to the rest of the region.
Three forces make this distinction visible in market data. First, Singapore sits at the premium end of the regional market. Second, local compensation data reflects this split: salaries for AI agent engineers jumped 18 percent and AI compliance leads jumped 22 percent, while pay for generalist mid-level coders fell 4 percent. Third, Singapore’s strict regulatory environment makes unmanaged AI code generation an unacceptable liability, making governance a non-negotiable baseline.
iTNews Asia: Will this be similar to other countries in Asia? If not, why?
The broader region adapts with greater caution, with 67 percent of Asia-Pacific developers reporting sanctioned AI tool use, compared to 88 percent in the United States. Neighboring markets still rely on developer volume as a competitive lever, an option Singapore lacks.
Singapore is the critical early-warning indicator for the region. National initiatives back this reality: IMDA’s AIxTech programme is upskilling 40,000 technology professionals, alongside a national strategy targeting 15,000 high-tier AI practitioners. Singapore is not training syntax typists; it is building leaders who translate architectural decisions into business outcomes.
iTNews Asia: Looking at developing talent, what skills employers and universities should prioritise as AI coding agents become more autonomous?
The skills that matter are no longer below-the-architecture line; the routine execution layer that AI is rapidly automating. They are above it: the design, governance and judgment layer that AI cannot replicate.
AI is rapidly automating syntax writing, basic implementation, unit testing and routine maintenance. What remains indispensable is domain modeling, system decomposition, security engineering and trade-off evaluation under ambiguity.
But above-the-architecture is not just a technical space. In my experience, the skills required span three forces that no single institution currently teaches well: engineering, economics and psychology. Engineering schools teach the first. Business schools teach fragments of the second and third. Neither teaches the intersection.
The collaboration I would most like to see is between MBA programs and engineering schools, producing what I would call tech-first managers: leaders who can simultaneously govern AI systems, evaluate economic trade-offs and manage the human dynamics of transformation.
The agentic era also demands a return to age-old management disciplines. Ruthless prioritisation. Sustained domain focus. And faster Think-Do-Learn-Iterate cycles. Good program and project managers who can hold these disciplines under pressure are in critically short supply. Most organisations discover this too late, when their AI programmes stall not for technical reasons but for management ones.
If training remains focused on below-the-architecture execution, we risk cognitive surrender: developers passively accepting AI outputs without understanding system mechanics, building organisations drowning in technical debt they can no longer diagnose.
The prescription is straightforward. Stop training people to write code. Train them to architect the boundary lines that keep autonomous agents secure, compliant and effective.
.jpg&h=271&w=480&c=1&s=1)
