AI Coding Tools Are a Mirage -- Real Value Lies in Workflow Reinvention

AI’s first real product is coding -- and that changes everything. Most of the value from AI won’t be captured by the models themselves, but by the applications and workflows they enable, just as past platform shifts saw infrastructure layers commoditized while value moved up the stack. The hidden consequence? Companies betting everything on foundational models may be building expensive rails for someone else’s train. This post is for leaders, builders, and investors trying to separate transient hype from structural advantage -- and it reveals where the real moats will form: not in models, but in reimagined workflows, restructured teams, and newly automated processes that only emerge after years of messy iteration. The immediate signal is coding; the long-term implication is a complete reshaping of knowledge work, with AI as the invisible engine, not the visible product.

Why Coding Was the First Domino -- And Why It Hides the Real Shift

The most important signal from the past year isn’t that AI got better. It’s that agentic coding went from useful to transformative -- and did so in a way no one could have precisely predicted. As Benedict Evans notes, “clearly agentic coding went from being kind of useful to really changing everything.” This wasn’t a gradual improvement. It was a phase shift.

And here’s the kicker: the fact that coding worked first wasn’t random. It was inevitable -- not because AI is inherently good at code, but because software developers were the first users, building tools for their own work. “The first thing that people are doing with LLMs,” Evans says, “is to make more compute.” It’s recursive creation: the people with access to AI are using it to build more AI and more software, accelerating the very infrastructure they depend on.

But this creates a dangerous illusion. Because coding has product-market fit -- customers are “pulling it out of your hands” -- the industry has narrowed its focus almost entirely on developer tools, code generation, and AI-augmented engineering. That focus blinds us to a deeper reality: this is still early days, and we’re optimizing for the wrong layer.

Most companies aren’t software shops. Most knowledge work isn’t coding. So the real question isn’t “How do we build better AI coding tools?” It’s “What happens when another domain -- law, finance, logistics, healthcare -- gets its own coding moment?”

We don’t know when that will happen. We don’t know which field it will be. But when it does, it will unfold the same way: a narrow, high-leverage use case will emerge, demand will outstrip supply, pricing will go haywire, and then -- slowly -- the chaos will settle into new workflows, new roles, and new business models.

And just like with mobile data in 2009, the infrastructure will get built, the usage will explode, and the profits will go to someone else.

"They built this amazing piece of incredibly sophisticated, very expensive global infrastructure... and they didn't make any money from it because all the value moved up stack."

-- Benedict Evans

This quote isn’t about AI. It’s about mobile networks. But it’s also a warning for AI. The telcos built the pipes. Apple and Google built the apps. Netflix and Uber captured the value. The infrastructure layer scaled, commoditized, and margin-crushed -- while the applications layer thrived.

AI models are on the same path. The trillion-dollar capex race isn’t a moat -- it’s a trap. Because if all models become roughly equivalent, differentiated only by cost and speed, then pricing power evaporates. And when developers can switch between OpenAI, Anthropic, and open-source models as easily as they switch cloud providers, the foundation model becomes invisible infrastructure -- like TCP/IP, like AWS, like the browser.

The system responds by routing around the bottleneck. Developers don’t care which model runs under the hood. They care whether the tool works. And that means the real innovation isn’t in the model -- it’s in the user interface, the workflow integration, the guardrails, the data pipelines, the tooling.

The UI Lie: Chatbots Are Not the Product

Most AI companies today are selling chatbots. That’s a problem.

Because as Evans puts it, “the chatbot itself is like a kind of weird limited v1 UI.” It’s the default interface, not the final one. And it’s terrible for most real-world tasks.

Why? Because real work isn’t conversational. It’s structured, iterative, and context-heavy. You don’t “chat” your way through a financial model, a legal brief, or a supply chain audit. You use tools -- spreadsheets, databases, dashboards, workflows.

So the current wave of AI adoption is stuck in a liminal zone: powerful enough to automate tasks, but too unstable to trust with jobs.

This creates a fundamental tension: where do you put the probabilistic software (the LLM) and where do you put the deterministic system (the database, the ERP, the CRM)? Do you embed AI into Salesforce? Or do you let AI pull data from Salesforce, synthesize it, and suggest actions? The answer is both -- but the balance determines who captures value.

And right now, most enterprises are groping in the dark. They don’t know what to automate. They don’t know how to retrain teams. They don’t know whether to hire fewer juniors -- or retrain them for higher-value work.

"Do you hire junior people and if so, what are they doing and why were you hiring junior people in the past?"

-- Benedict Evans

That question isn’t about headcount. It’s about the purpose of work. If AI automates the grunt work -- code reviews, document drafting, data entry, report generation -- then what’s left? Judgment. Strategy. Exception handling. The stuff that can’t be templated.

But here’s the catch: you can’t train for that in a chatbot. You need systems that surface edge cases, not average outcomes. You need interfaces that support decision-making, not just generation.

And that’s where the real opportunity lies: in rebuilding software not as a set of predefined workflows, but as adaptive systems that learn from how people actually work -- not how consultants think they should.

The Real Platform Shift: More Software, Not Less

Everyone assumes AI will reduce the need for software. The opposite is true.

AI will create way more software -- not because we need more features, but because every company will become a software shop.

Think of the current enterprise software stack: SAP, Workday, Salesforce, Oracle. These are monolithic systems built for scale, not flexibility. In the middle: a messy layer of spreadsheets, email, shared drives -- the “shadow IT” that keeps things running. On the edges: hundreds of SaaS apps, each solving a narrow problem.

Now add AI. Suddenly, you can:

• Automate workflows between systems
• Generate custom tools for niche tasks
• Let non-engineers build lightweight automations
• Let AI agents query systems directly, without a UI

The result? Fragmentation accelerates. Instead of a few big apps, you get a thousand micro-tools -- some built by vendors, some by consultants, some by internal teams, some by AI itself.

And that means SaaS consolidation reverses. The Microsoft- or Salesforce-style “platform” becomes less relevant. Why? Because AI can bridge systems without requiring full integration. Why buy a $50K/year seat when an AI agent can do the job for $500?

This isn’t hypothetical. It’s already happening in consulting, law, and finance -- where firms are using AI to automate discovery, due diligence, and reporting. But they’re not using off-the-shelf tools. They’re building custom pipelines, because the generic chatbot doesn’t cut it.

And that’s the real shift: AI isn’t replacing software -- it’s unlocking a new layer of software that was too expensive or too hard to build before.

The Capex Mirage: You Can’t Spend Your Way to a Moat

There’s a dangerous narrative in tech right now: if you spend enough on AI, you’ll win.

Meta, Google, Microsoft -- they’re spending 80% of revenue on capex. $700 billion this year alone. That sounds like dominance. But it also sounds like desperation.

Because capex at that scale isn’t a moat -- it’s a financial gravity problem. You can’t sustain it. And when the efficiency of models improves 10x, 100x -- as they have -- yesterday’s frontier becomes tomorrow’s commodity.

Evans’ point is brutal: “Just because demand for tokens is infinite doesn’t mean you can’t get to a different price equilibrium.” Look at mobile data. Demand grew 2,000x. Prices collapsed. Telcos spent trillions. Profits? Flat for 20 years.

AI will follow the same path. The companies that win won’t be the ones with the biggest models -- they’ll be the ones who figured out how to build valuable products before the infrastructure commoditized.

And that means the real advantage isn’t in compute -- it’s in time. The teams that spend 2024 iterating on workflows, retraining talent, and building AI-native tools will have a 12- to 18-month lead when the next domain hits its “coding moment.”

Because when that happens, it won’t be about who has the best model -- it’ll be about who has the best workflow.

Key Action Items

• Over the next quarter: Audit your internal workflows to identify tasks that are repetitive, high-volume, and context-rich -- these are AI’s sweet spot. Start with coding, support, or reporting, but don’t stop there.
• Within 6 months: Shift AI experimentation from chatbots to embedded tools -- integrate AI into existing software (CRM, ERP, email) rather than building standalone agents.
• Over 12--18 months: Re-evaluate junior hiring and training. If AI automates entry-level tasks, what higher-value skills should new hires develop? Invest in judgment, not execution.
• Now: Assume foundational models will become commoditized. Build your strategy around workflows, data, and user experience -- not model choice.
• Immediately: Monitor token usage and cost. Many companies are “token maxing” -- using expensive models for trivial tasks. Implement cost controls and ROI tracking.
• Over the next year: Partner with domain experts (lawyers, doctors, engineers) to co-design AI tools. The best applications won’t come from AI teams -- they’ll come from people who understand the work.
• Long-term: Prepare for fragmentation. The era of monolithic SaaS platforms may end. Invest in interoperability, APIs, and modular tooling to stay agile.