Organisations that have begun leveraging artificial intelligence (AI) in their software development lifecycle see marked improvements in productivity, but remain cautious about letting the technology handle some tasks.
In particular, AI still isn’t ready to manage more complex workflows and activities that require stringent accountability and auditability.
At digital payments operator Nets (Network for Electronic Transfers Singapore), AI has enabled its software development team to spend progressively less time on routine and “less-engaging” tasks, and devote more time on higher value-added activities, such as architecture decision-making and solutions design.
“The biggest job shift over the last one to two years is with the tasks that occupy most of our developers’ and testers’ time,” said Bernard Chew, group CIO at Nets.
Its software developers are tapping AI increasingly for tasks, such as code development, test case generation, code review, and documentations, Chew told FutureCIO via an email interview.
They use AI to refine user specifications and code scaffolding, as well as to perform code reviews and accelerate troubleshooting.
Test teams also leverage AI to generate test scenarios and improve test preparation, he said.
In addition, engineers use AI-assisted knowledge retrieval to more quickly understand legacy applications and dependencies.
“Software engineers’ focus has shifted left, with more attention on outcome-based design,” he noted.
His team adopts “a blend of specs-driven and test-driven development”, establishing boundaries and necessary guardrails within which AI agents operate to support code and test generation.
Nets’ security teams, too, are beginning to tap AI-powered capabilities for vulnerability management, he said.
Still limited in complex reasoning
But while AI has led to productivity gains, there still are gaps in more complex enterprise processes, according to Chew.
“These often require multi-stage coordination, deep business context, and human judgment,” he said. “Ultimately, AI is becoming a powerful co-pilot, but it is not yet at a stage where it can independently manage complex, high-risk workflows that require reasoning, explainability, and accountability.”
In more complex solution designs, for instance, AI can support implementation, but cannot be expected to handle large-scale architectural decisions that require balancing business objectives, regulatory requirements, security, resilience, and operational risks, he noted.
Enterprise initiatives also often require collaboration across business, technology, risk, compliance, operations, partners, and regulators.
“While AI can support individual activities, it should not be expected to effectively orchestrate complex end-to-end decision-making,” he said.
More care is further required for companies that operate in highly regulated sectors, where AI explainability and auditability are critical.
Established in 1985 by a consortium comprising multiple Singapore banks, Nets currently is owned by DBS, UOB, and OCBC.
“In payments and financial services, critical decisions must be transparent and auditable,” Chew said. “Whilst there are AI models that provide auditable steps and logs, these are still maturing.”
“The use case for AI agents executing payments or fully automated anti-money laundering and KYC (Know Your Customer) decisions still face challenges in accountability, governance, customer consent, and regulatory compliance,” he noted.
In fact, while software developers are open to giving AI agents more autonomy for lower risk, reversible work, such as documentation, desire for human oversight increases as tasks move closer to production.
Some 79% requires human approval for production deployment and just 38% say their codebase is ready for a fully autonomous agent, according to a study from Agoda, which polled 800 software developers across seven Asian markets, including Singapore, India, Indonesia, and Vietnam.
Some 79% highlight hallucinations or incorrect outputs as a top concern.
Asked about barriers to adoption, 28% cite cost while 24% point to integration complexity and 19% note the lack of governance.
AI cost also increasing concern
Tokenmaxxing, for one, was once a popular workplace practice to motivate employees to “max out” their token consumption as a way to encourage them to experiment with AI. The practice begun losing its shine when reports emerged of IT departments busting their annual budgets within months.
With AI expected to be integrated into more workflows, there are growing concerns that token costs could escalate out of control.
It underscores the need for organisations to be thoughtful about training their AI agents with the right context, said Purna Doddapaneni, a partner at Bain & Company, who is part of the consultancy’s performance improvement practice, advising clients on platform architecture and technology transformation.
Failing to do so would result in agents consuming more tokens than needed and still end up hallucinating, Doddapaneni told FutureCIO via a video call.
Adopting good harness engineering will ensure organisations focus on the context as well as business outcomes they are trying to drive, so they are not needlessly wasting tokens, he noted.
More importantly, harness engineering builds the scaffolding around agentic workflows — from start to end — so AI agents deliver the expected outcomes, he said.
Software engineers have to lay down the rules, stating what agents can and cannot do, he added.
Such efforts should include putting in deterministic rules and concepts, where they can validate what the agents are creating before the latter are put into production, he said.
Doddapaneni also urged enterprises to measure the business value they want out of their AI adoption.
“If I’m only measuring the volume of codelines generated, then the AI agents will just burn through tokens to generate codes,” he said.
Instead, organisations should be measuring what is needed to take a software feature from idealisation to a customer’s hands, he noted.
“Start measuring the right business value so you can optimise for the right things,” he added.
He also highlighted the importance of engineering hygiene, where businesses should decide what are the right tools and AI models they need.
“Do they need to have the biggest model [to apply to a code base] if the problem needs only a basic model?” Doddapaneni said.
Governance still the key
The biggest challenges in integrating AI into development workflows also often are not technical.
Instead, they are governance, risk, and organisational challenges, said Chew, who noted that Nets has worked to resolve such issues through standardised tooling, common DevSecOps platforms, and shared governance models.
The Singapore organisation set up governance committees, policies, and oversight mechanisms, which are regularly reviewed to ensure Nets keeps up with changes in the AI market, he said.
Because it operates in a highly regulated environment, Nets needs clear accountability for AI-generated outputs, he noted.
These means consistent standards must be established for approved tools, acceptable use cases, model risk management, and human oversight, he said.
Without standardisation, for instance, there are risks of fragmentation where different teams may adopt different AI tools, creating inconsistency, security concerns, and governance challenges, Chew said.
Agoda CTO Idan Zalzberg said in the report: “Across industries, developers are seeing stronger productivity gains and are beginning to use agents across more of the development lifecycle. But greater capability brings new responsibilities.
“The real cost is not simply the model call. It is the work required to make AI trustworthy in production. The organisations that will excel in this next phase will be the ones that manage cost carefully, build the right safeguards, and keep human judgment where it matters most,” Zalzberg said.
If done right, AI-powered software development can yield significant gains.
Asked where AI had generated the biggest returns for his developers, Chew pointed to the early and middle phases in the development lifecycle, such as software testing, knowledge transfer and management, and security.
The technology has helped to significantly reduce time spent on code generation, refactoring, troubleshooting, documentation, and code reviews, he said.
This is particularly valuable when the team has to deal with multiple code bases across different technologies, he added.
Because Nets operates a mix of modern and legacy platforms, he explained that AI has helped new engineers become better acquainted with the intricacies of these systems.
It allows them to spend less time figuring out how to manage the system and spend more time on solutioning, he added.
Chew noted that the fastest measurable gains have come from testing tasks, spanning test data generation, regression test preparation, test coverage improvement, and test documentation improvement.
“Performing comprehensive testing has traditionally been a resource-intensive activity and AI is helping to reduce this substantially,” he said.
Software developers in Asia also report productivity gains from tapping AI, with 55% claiming to save at least seven hours each week, the Agoda study found. This figure reflects an increase from 18% who say likewise last year.
Specifically, 84% leverage the technology for code review as well as to write code, while 81% each tap AI for planning and for documentation.
In addition, 71% are testing AI agents for document generation and 70% are doing likewise for code refactoring.
Another 66% are piloting AI agents to autonomously generate code, while 63% do so for research and analysis.
Furthermore, 53% already use AI agents in production workflows or broad organisational use, and 22% are tapping AI agents for selected production workflows.
However, 86% still review or validate AI-generated outputs most of the time or always.
A Bain & Company report further predicts that agent-assisted software engineering will climb to more than 50% by year-end, up from the current 25%.
The consultancy expects the figure to increase further to 90% by mid-2027.
It adds that this growth momentum will drive a change in developers’ role, which will transition from code executors to agent architects and orchestrators, working alongside hybrid human-agent teams.
“This shift is redefining the role of engineering as the foundation for broader enterprise transformation. It’s not enough for engineering teams to deliver code five times faster; business teams must generate demand at the same pace, and operations must match that speed to deploy, scale, and support solutions in production,” Bain consultants commented on the report.
“Unlocking the full value requires an end-to-end transformation across the entire delivery chain,” they wrote. “What once looked like ambitious progress now represents the baseline for a fundamentally different era of software development.”
According to Bain, efforts to bring AI into development workflows may experience roadblocks if organisations focus only on optimising a single activity, such a code generation or test creation. Doing so can simply push bottlenecks elsewhere.
“Unlocking real value requires broader changes in behaviour and organisations,” Bain notes. “Without new workflows, measurement, and guardrails, companies are likely to see adoption plateau and only minimal new value.”
Human developers remain essential, with evolved skills
Humans, too, will remain a necessary part of the software development lifecycle, even as the role of AI expands.
At SAP, about 30% to 40% of the entire software development cycle is AI-assisted, with its developers tapping Anthropic’s agentic coding tool Claude Code, amongst other coding assistance tools, said Manik Saha, managing director of SAP Labs East Asia.
“The myth that you don’t need developers is false because you will need better developers in agentic AI coding, where they need to review the code and ensure it’s optimised and ready for enterprise systems,” Saha said, in a video call with FutureCIO.
Furthermore, focus on specs-driven development has increased, he said.
This requires expertise on how to construct specifications to allow agentic AI tools to understand the intent and create codes based on that intent, he explained.
“We’ve moved away from small features to writing complete specs, which include boundary exceptions and what the [agentic] behaviour should be,” he said. “And you need engineers to ensure this code is optimised.”
Engineers also are critical to guide software tests, where AI-driven test parameters are needed to understand what agents are supposed to do, how they are supposed to behave, and the guardrails needed to ensure agents do not run out of control.
When enterprises create a new agent, they need to ensure all the baselines, governance, behaviour of agents, data boundaries, and authentication remain secured and in place, Saha said.
Chew said: “Our intention is not to replace engineers. In a highly regulated payments environment, success will come from balancing innovation with governance, security, and operational resilience.”
Production-grade systems in financial services require more than code generation, he said, noting that technology teams must provide guardrails, monitoring, integration, and governance frameworks that support innovation while maintaining operational risk and sustainability.
“The challenge will no longer be writing code. It will be sustainability, security, and integration of the codes and tools that are built,” he added.
To better support business teams, Chew noted that DevOps will need to evolve to become platform engineers, solution architects, integration specialists, security and governance experts, and reliability engineers.
Saha said: “Developers still need to know basic coding because they still need to read the code and understand it.”
Software engineering has changed and developers’ role has evolved, where AI today is just another step in development and helping developers do more, he said.
In fact, SAP now looks for developers who already are able to work with AI-assisted tools, he added, noting that this was not a required skillset five years ago.
New software engineers also have to go through a two-month training program to understand the AI ethics work environment and governance model, he said.
Engineers will need to look beyond simply generating code and better understand what customers need and build the necessary products, Doddapaneni said.
Platform and DevOps teams are gaining in importance, he said, adding that engineers will be needed to solve new-world problems.
They also have to look at quality requirements, where testing can no longer be an afterthought, he noted.
In addition, context engineering will be essential, to ensure AI agents generate the right output and consume the right tokens, he said.
