Chatbots are great for drafting a function and wishing you luck, but not helpful when it comes to more nuanced requests. But imagine you had a software developer on permanent standby. Someone you could turn to and say, “Hey, I have a question about this,” or “Can you build this for me?” and they’d just do it. Now imagine that friend works ten times faster than a human, never sleeps, and can read any codebase you point them to.
That friend now exists with tools like OpenAI’s Codex and Anthropic’s Claude. While most of the industry is talking about the impact of these tools on productivity, the real story is the value chain, and it has some significant implications for anyone making a three-year platform bet.
By the numbers
The numbers coming out of the labs are striking, even allowing for the fact that they are self-reported by companies whose valuations depend on them being true:
- At Sequoia’s AI Ascent conference, OpenAI president Greg Brockman said that over the course of a single month, agentic coding tools went from writing 20% of your code to 80%, shifting from a sideshow to the main event.
- Google’s Sundar Pichai has said that about 75% of new code created inside the company is now generated by AI and reviewed by humans.
- Anthropic’s Dario Amodei has gone further, predicting that AI would write 90% of code within months and noting that because AI already writes much of Anthropic’s own code, it is accelerating the pace at which the company builds its next models.
It’s important to note that when you look closer at the lab figures, they do not all measure the same thing. Google’s is new code under human review; others are future expectations or estimates, and none of the studies are independently audited.
But while these findings vary, they all signal a clear direction, and this is important information for anyone making platform investments.
See also: AI Chatbot Outperforms Human Clinicians in Probabilistic Diagnosis
From autocomplete to colleague
The first wave of AI coding tools was like a glorified autocomplete tool. They were helpful and occasionally clever, but what’s emerging now is categorically different. These new tools function as on-demand developers who are capable of handling every aspect of building a feature from description to working code.
I’ve shipped five massive feature builds almost entirely this way, and here’s the part that surprised me: the bottleneck wasn’t building the code; it was reviewing it. When a tireless collaborator can produce a complete, working feature in an afternoon, the time suck is no longer the engineering hours; it’s the human judgment required to check it.
The inversion of time and resources changes the whole game. For decades, the cost of software has been the time and resources needed to produce it. We hired teams and guarded our roadmaps because writing code was slow and expensive. Once writing code gets cheap, every assumption we built on top of that cost starts to change.
This extends beyond features as well. The same assistants now write ad hoc database queries on demand. As an engineer would, they connect to live applications and can trace a production defect back to the specific commit that caused it. They can also suggest a fix, which is more important than it sounds. Connecting these models to real-world systems is exactly what’s making them so much more capable. The more you can show the model, the better it can navigate, call additional tools, and reason about what’s occurring. At the end of the day, the buzzword “agentic” has come to mean an engineer who has read everything and never gets tired.
This is not the same thing as “vibe coding,” the casual, consumer-friendly mode where someone describes what they want and ships whatever the AI produces without really reading it. That’s the low-stakes end of the spectrum, fine for a weekend prototype. What I’m describing is the opposite posture: production features that still demand rigorous review.
Remember: AI is still a tool, not a replacement for humans. It surfaces options at superhuman speed, but interpreting them, weighing trade-offs, and owning the final call still needs humans.
Follow the money
I predict that a lot of SaaS businesses will be disrupted because they can now be easily replaced by something built with an AI assistant’s help. When good software becomes fast and cheap to build, a company’s biggest advantage falls apart. A lot of SaaS companies thrived because they had already done the hard work of building their product. A competitor couldn’t easily copy them without a big engineering team and a lot of money and time. Once AI makes building easy, that’s no longer true.
Spending will shift away from buying pre-packaged software toward building tailored capability, and from vendors who competed mainly by having a mature product to those who compete on something harder to replicate. We’ll see successful companies building these new approaches into their products, generating new features on demand, and answering questions about the product with an assistant that understands the systems because it can look under the hood. This then improves the support experience because it can answer questions a human support technician never could.
The caveat, though, is that more code tends to mean more bugs. But the overall trajectory is optimistic: We’ll see support quality improve and broken software become unusual.
Making the call
We started with the idea of a developer on permanent standby. The strange thing is that the more capable that collaborator becomes, the more it surfaces what only humans can do: deciding what’s worth building, judging whether it’s good, and standing behind that call.
So, here’s the question: If software creation is now dramatically faster and cheaper, and the constraint has moved from building to reviewing, what exactly is your product defending? If the answer is that it’s hard to replicate, know that will no longer be the case. If it’s a deep understanding of your customer’s problem, proprietary data, trusted judgment, or an experience that compounds, you’re building on solid ground.
