Brian Greenberg is CIO of RHR International, professor of cybersecurity at DePaul University, board member, trusted advisor and speaker.
getty
In July 2025, Jason Lemkin tried an experiment in vibe coding … building software by describing what you want in plain English and letting AI handle the rest. Lemkin is a well-known software investor, not an engineer, but the early results were promising enough that he documented his progress along the way. Then, Replit’s AI agent deleted his app’s production database, which contained records on over a thousand executives and companies. He said he told the AI, 11 times in all caps, not to change anything without permission. The agent then claimed the data was unrecoverable. That turned out to be wrong. Fortunately, the rollback worked.
Each failure in that story has an established engineering solution. Production data should be kept separate from development. Dangerous commands need safeguards. The AI behaved as any unsupervised tool would. What was missing was the engineering judgment.
That judgment usually comes from senior engineers, who have already made their fair share of mistakes, and learned from them. They ask tough questions, challenge assumptions, spot risks early, simplify complex problems and help the business avoid expensive mistakes.
As AI embeds itself in software development, technology leaders have to ask: Whose engineering judgment is the AI following?
Most discussions about AI coding assistants still focus on speed, which makes sense because speed is easy to see and measure. McKinsey found that developers finished some common tasks up to twice as fast with generative AI. However, results varied a lot depending on the task and the developer’s experience. Junior developers sometimes took longer, and human oversight was still needed to catch mistakes, provide context and handle complex requirements.
AI is useful. But writing code faster is not the same as doing it better.
AI Amplifies The System Around It
AI acts as an amplifier, making an organization’s strengths and weaknesses more apparent. The biggest benefits come from improving the systems around the tools, not just the tools themselves.
If you give an AI assistant vague, unclear requirements, weak security, some testing and a rush to deliver, it will just produce poor work faster. But if you give it clear instructions, strong engineering practices, automated checks and clear verification steps, it can help a team work better.
AI doesn’t take away the need for experienced judgment. Instead, it spreads whatever judgment, or lack of it, is already guiding the work.
AI behavior is shaped by the instructions, codebase, examples, tools and rules around it. Choosing to design that environment carefully is up to leadership.
A senior engineer’s main value isn’t just how fast they can write code. It’s the way they approach their work.
They make sure everyone knows what needs to be built before starting. They know the difference between a quick prototype and a real system that handles customer data. They understand which practices can grow with the project and which ones are essential. They don’t mistake a likely answer for a proven one, and they don’t call it a success just because the code runs or tests pass.
These habits are hard to measure, but they mitigate failures.
Turning Standards Into Everyday Practice
Most organizations already have engineering standards. The problem is these standards often sit in documents that people read once, are used inconsistently or are only remembered by a few experienced staff.
AI gives us a chance to turn that knowledge into everyday actions.
This means going beyond prompts like “write secure code” or “act like a senior engineer.” That's too vague for actual work. You need clear, reusable instructions that tell an AI assistant how to clarify requirements, plan changes, handle uncertainty, apply security, test its work and check results.
That guidance should be tied to clear, automatic checks. If a security rule can be checked with a scanner, use the scanner. If behavior can be tested, require a test. If a claim about the environment can be checked with a tool, check it instead of trusting the model’s confidence.
NIST’s Secure Software Development Framework makes a similar point at the organizational level: Secure practices should be integrated into the software development life cycle, not bolted on afterwards.
The same goes with AI-assisted engineering. Security, testing and checking should be built into the workflow, not added at the end.
An Experiment In Encoding Judgment
This approach led me to create and open-source a project called Senior Engineering Partner.
It’s available as a skill for Claude Code, but its main instructions also work with other AI's that can plan, act and check their own work. The tool isn’t the main focus. The project is an experiment in turning the behaviors I expect from experienced engineers into clear instructions.
It drives a specification-to-plan-to-test-to-verify loop. It applies a phase-aware rigor ladder, so a prototype isn’t burdened with every production control, while holding a constant security floor: no hard-coded secrets, validation at trust boundaries, protection against injection, isolation, least privilege, authentication, vetted dependencies and a credible recovery path.
It also puts verification first. The assistant is told to check before making claims, never make up flags or APIs, automate anything that can be checked and separate evidence from belief. Real failures become test cases, so lessons learned turn into regression tests instead of just being written down.
Ask yourself what your best engineers always do before they start coding. Decide which security and quality practices are absolutely necessary. Find out where important knowledge exists only in people’s heads. And figure out how you’ll know if an AI system actually followed the right process, not just gave a convincing answer.
AI will almost certainly speed up software development. But making your engineering organization better is up to leadership. The real advantage won’t go to whoever uses the most powerful model. It will go to organizations that build systems where AI supports the discipline and standards of their best engineers.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
>> Home