When 95% Of Code Is AI-Written, What's Left For Humans?
Malte Kosub, is cofounder and CEO of Parloa, the company behind an agentic AI platform reimagining how enterprises connect with customers.
gettyIn 2025, “Microsoft CTO Kevin Scott predicted that 95 percent of programming code [would] be AI-generated by 2030.” Just one year later, many businesses, including my own, have already achieved that metric.
We have quickly moved toward a world where the vast majority of code is machine-generated. Amid the rapid change and urgency to adapt to this new reality, a looming question remains: What happens to human engineers?
I know that engineers aren’t being replaced, but the job has fundamentally changed. The work is no longer writing code. It’s creatively designing the systems that agents operate within and teaching those agents to reason the way humans do.
Organizations are no longer repurposing software. They are rebuilding what humans do. That distinction matters because it fundamentally changes the problems engineers and technology leaders need to solve.
When the industry moved from on-premise software to the cloud, the objective was clear: Make existing systems more scalable, resilient and future-proof. AI presents a categorically different challenge. Instead of simply building software, you’re designing systems that must replicate aspects of human judgment. That means confronting questions the industry is still working to answer:
• How long can a prompt be before accuracy degrades?
• How do you maintain guardrails when customers use a system in unanticipated ways?
• How do you design for reliability when agents can take unpredictable execution paths at runtime?
None of these are code generation problems. They are system design problems.
Within 12 months, I watched our engineering team’s center of gravity shift from “How do I implement this?” to “How do I design and supervise a system of agents that will implement this for me?” The engineer’s role leaped forward, and with the leap came a new set of obstacles.
The first is architectural: Engineers now spend most of their time working on the scaffolding around agents, the harness and the guardrails. The second is contextual: Agents lack the human perspective that makes software trustworthy in practice. Agents need to be trained on the judgment, institutional knowledge and edge-case awareness that humans carry.
We coined a term internally for where engineering is heading: product-minded architect.
• Architect, because the creative work is now focused on the architecture and the design decisions, system coherence and trade-offs that hold up under enterprise complexity.
• Product-minded, because knowing what to build has become more critical than knowing how to build it. If you were building the wrong things 20% of the time before and you’re now 10 times faster, you are producing 10 times more wrong things. AI amplifies output dramatically, but it does not fix unclear thinking. It makes unclear thinking significantly more expensive.
The engineers who are furthest along think of themselves less as architects and more as founders, responsible for what gets built, why it exists and what happens when it fails. They are becoming guardians of quality and choosing the right products for the right use case. Engineering today is about taste and curation.
As AI agents quickly produce code, it is the engineer’s responsibility to infuse human thinking and creative originality into that production. This means providing the organizational and contextual intelligence to ensure a compliant, relevant output across every system.
At the same time, engineers now have significantly more pull requests in flight than a year ago. You can increase throughput with agents, but you can’t increase the human brain’s ability to hold context across multiple parallel streams of AI-generated work.
In the past, our engineers pair-programmed. Now, they pair-conceptualize. Human teams work together to build systems and full products, not just tasks. They treat AI agents as their co-workers and are responsible for mapping agents’ blind spots and calibrating outputs against human judgment.
When an agent writes the code, who owns the bug? Who owns the security gap? Who owns the edge case that only surfaces at scale? The answer is the same as it has always been: the humans who designed the system and chose what to deploy.
In consumer software, you can increasingly build working products without deeply understanding the system underneath. In enterprise environments, when you’re operating across jurisdictions, compliance regimes and decades of legacy infrastructure, that approach breaks almost immediately. You are not just deploying software. You are also designing systems that must operate reliably under strict data governance rules, layered internal trust boundaries and regulatory scrutiny.
Once you introduce agentic systems, this complexity compounds because you need observability, containment and failure isolation for systems whose runtime behavior you cannot fully predict.
Trust in these systems must be earned progressively. An agent starts by recommending, graduates to acting with a window for human objection and eventually operates autonomously, but only after proving itself over hundreds of decisions. That is how we think about deploying AI agents in enterprise contact centers, and it is exactly how we govern our own engineering systems internally.
Programming languages were always an intermediary between human intent and machine execution: Assembly, C, Java and Python were each one step closer to how humans think. AI is the next step in that arc. Eventually, the intermediary disappears. You describe what you want, and the machine builds it. That makes the quality of human intent and skill even more consequential.
The cost of producing software is approaching zero. However, the cost of knowing what to produce and taking responsibility for it is not. That is where experienced engineers and enterprise leaders show their value: designing what should exist, under which constraints and with which safeguards and then standing behind those decisions when systems run in production at scale.
AI is an amplifier. It doesn’t change what good judgment looks like. It just makes the consequences of having it or not having it arrive much faster. If your team’s output increases 10 times in the next year, does your judgment capacity scale with it, or does it stay exactly where it is today?
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
