Zoho’s Sridhar Vembu Warns Engineers Not to Let AI Replace Their Understanding

Zoho founder Sridhar Vembu has cautioned software engineers against becoming so dependent on AI coding tools that they lose their understanding of the systems they are building. His comments come amid a growing debate over whether AI is improving software development—or encouraging teams to prioritise output over engineering judgment.

The rapid adoption of artificial intelligence in software development is changing how engineers write, test and maintain code. AI coding assistants can now generate programs, prepare tests, explain errors and help developers move through parts of the development process much faster.

But that convenience has raised another question: what happens when engineers stop understanding the work because an AI system is doing too much of it?

Zoho founder Sridhar Vembu has entered that debate, warning engineers to use AI without surrendering their technical understanding to it.

Responding on X to a post describing extensive AI use inside a software-development team, Vembu wrote: “This is sad. I tell our engineers to use AI but never cede our understanding to AI.”

He compared the industry’s rapid adoption of AI development tools to driving a sophisticated new vehicle without fully understanding how it works.

“As an industry, we are driving too fast in a fancy new car we barely understand how to drive, inviting disaster,” Vembu wrote.

Why Sridhar Vembu is concerned about AI-generated code

Vembu’s comments focus on a distinction that is becoming increasingly important in AI-assisted engineering: using AI to support developers is different from allowing AI to become responsible for the development process itself.

An AI coding assistant can generate a working solution, but engineers still need to understand what the software does, why it was designed in a particular way and what could go wrong.

Without that knowledge, developers may find it harder to identify subtle bugs, security weaknesses, architectural problems or unintended behaviour.

The issue becomes even more significant when AI-generated code is introduced into large and complicated production systems that will need to be maintained long after the original code was written.

Engineer says AI is being used across almost the entire workflow

Vembu’s response followed an X post from a software engineer describing an environment where Claude Code was reportedly being used throughout much of the team’s workflow.

According to the engineer’s account, AI was involved in creating specifications, code, tests, product requirement documents, tickets, ticket resolutions and reports.

The engineer claimed that the team was under pressure to maximise software output while some employees were uncomfortable with the extent to which AI had become involved in the process.

“Nobody knows anything here,” the engineer wrote.

The post continued:

“The specs, code, tests, PRDs, tickets, resolution of those tickets, reports, etc., everything is made by Claude Code. Nobody on my team likes this. They are being forced to ship as much as they can.”

The engineer also described spending long working days—reportedly 12 to 13 hours—directing AI systems rather than deeply working through the software themselves.

Another claim in the post was particularly blunt:

“Nobody is resolving bugs. In reality, nobody is thinking anymore. Everything is done by LLMs.”

These are the engineer’s personal accounts on X and should not be treated as independently verified evidence of practices across the wider technology industry.

The real issue may be over-reliance, not AI itself

The debate that followed Vembu’s post was more nuanced than simply arguing for or against AI.

Some engineers said they had experienced situations in which AI-generated code made them less familiar with a codebase. Others argued that the technology can reduce repetitive work and allow developers to spend more time on higher-level problems.

That distinction is important.

AI can be useful for tasks such as generating boilerplate code, writing first drafts of tests, documenting functions, exploring implementation options or helping locate bugs. The problem arises when speed becomes the only measure of productivity.

A team can produce more code without necessarily producing better software.

When shipping more code does not mean better engineering

Software development has traditionally involved several stages of human judgment.

Engineers have to understand requirements, design systems, select appropriate technologies, anticipate failure cases, review code and troubleshoot problems after deployment.

AI can assist with many of these activities, but removing human understanding from the process can create a different kind of technical risk.

If developers approve code they cannot explain, organisations can accumulate what might be described as AI-assisted technical debt—systems that appear productive in the short term but become difficult to maintain because the humans responsible for them do not fully understand their structure.

That concern is particularly relevant for security-sensitive software and systems handling financial, personal or operational data.

AI coding tools are changing the definition of productivity

One reason the debate is becoming intense is that AI can dramatically increase the amount of software an individual engineer appears capable of producing.

This creates an obvious incentive for companies to increase development targets.

But more output can also create more code to review, test, secure and maintain.

If the organisation measures engineers primarily by how much software they ship, developers may naturally favour whatever allows them to maximise output. AI tools can become central to that process.

The result can create a difficult feedback loop: AI increases productivity expectations, higher expectations encourage heavier AI usage, and heavier usage can reduce the time available for engineers to understand the underlying systems.

That is one interpretation behind the concerns raised in the X discussion.

Some engineers argue AI can reduce workload

Not everyone in the discussion viewed AI dependence as a warning sign.

Supporters of AI-assisted development argue that software teams have always adopted tools that automate repetitive work. From compilers and integrated development environments to automated testing and version-control systems, developers have repeatedly used technology to increase productivity.

AI can similarly help engineers handle routine tasks and focus their attention elsewhere.

One commenter responding to the discussion argued that engineers might be more willing to spend time understanding their systems when deadlines are less restrictive. The commenter suggested that when organisations focus heavily on results and impose tight timelines, employees may naturally rely more heavily on automation.

That raises an important point: AI adoption is also an organisational issue.

The way companies measure productivity may influence how responsibly employees use AI tools.

Why human understanding still matters

Even when AI writes the first version of a program, engineers remain responsible for deciding whether that program should be used.

Human developers need to be able to answer basic questions:

What problem does this code solve?

Why was this particular implementation chosen?

What assumptions does it make?

What happens when the input is unexpected?

How will the system behave under heavy load?

What are the security implications?

What happens when the AI-generated solution fails?

If a developer cannot answer those questions, the existence of a functioning program does not necessarily mean the engineering process has succeeded.

Vembu’s message is not to avoid AI

The most important part of Vembu’s response is that his advice was not to stop using AI.

He explicitly said that he tells Zoho engineers to use AI while warning them not to give up their understanding.

That places his position somewhere between rejecting AI and embracing unlimited automation.

The message is essentially about retaining technical ownership.

An engineer can ask an AI assistant to generate code and still review every important decision. AI can accelerate implementation without becoming the sole source of architectural or technical judgment.

What responsible AI-assisted engineering could look like

A balanced software-development workflow could use AI for speed while keeping critical decisions with people.

AI could generate an initial implementation. Engineers could then review it, test it, examine security implications, compare alternative approaches and modify the result.

The same principle can apply to documentation, requirements and debugging.

The goal is not necessarily to minimise the amount of AI used. It is to ensure that the humans responsible for the system remain capable of understanding and challenging what the AI produces.

That distinction may become increasingly important as AI coding systems improve.

The bigger lesson for software engineers

The debate sparked by Vembu’s comments reflects a larger transition in the technology industry.

AI is no longer merely a tool for answering questions or generating small pieces of code. It is increasingly being incorporated into workflows that previously required sustained human participation.

That creates huge opportunities for productivity, but it also changes the skills engineers need to maintain.

Knowing how to prompt an AI system may be useful. Knowing when not to trust its output may be just as important.

Engineers who understand software architecture, debugging, security, system design and the underlying business problem can use AI as leverage without becoming completely dependent on it.

Vembu’s warning ultimately comes down to that distinction.

ALSO READ:

• Geoffrey Hinton Warns Rogue AI Agents Could Develop Dangerous Subgoals
• Elon Musk’s Advice to 20-Year-Olds in the AI Era: Don’t Limit Yourself to One Field
• Zoho’s Sridhar Vembu Warns Engineers Not to Let AI Replace Their Understanding
• Apple Limits New Siri AI Features to Newer Devices With iOS 27
• AI Is Becoming a Science Problem, Sam Altman Says as Nvidia Security Debate Grows
• New Siri AI Supported Devices: Check Which iPhone, iPad and Mac Are Compatible
• lon Musk’s Advice to 20-Year-Olds in the AI Era: Don’t Limit Yourself to One Field
• Gas Cylinder ₹600 Cheaper? Here’s Why Some Commercial LPG Cylinders Are Being Sold at a Discount
• IBPS PO Prelims Result 2026 Released: Check Result, Scorecard Download Steps and Next Stage
• Sensex, Nifty Slide 1.7% as Crude Oil Surge and Insurance Reform Concerns Hit Markets
• NSE Shares Slip 1.39% a Day After Historic Market Debut; Valuation Stands at ₹4.43 Lakh Crore
• GPT-6 Astra System Card Raises Questions Over AI Reasoning Transparency