AI Coding: The Reality Beyond the Hype

## Disposable Code: The End of Permanent Software

[A] "So someone just spent two hours building a piece of custom enterprise software, right? They used it for a single month and then they permanently deleted it."

[B] "Yeah. And I mean, that might actually be the smartest business decision they made all year."

[A] "Which is wild because, you know, it completely breaks every rule we've been taught about how to invest in technology."

[B] "Absolutely. I mean, we are trained to hoard code, right, to protect it. We treat it as this foundational asset that needs to generate an ROI over like a decade."

[A] "Exactly. So throwing it away just feels almost irresponsible."

[B] "Right. It feels like setting money on fire."

[A] "But if you are a manager or a business leader, especially those of you tech curious managers tuning in from across Scandinavia, you are probably wrestling with this exact tension right now."

[B] "Yeah, everybody is."

[A] "You see these incredible AI coding demos online. You hear the endless hype, but you're trying to figure out, like, what does this actually mean for my daily operations?"

[B] "Right. Risk management, budget allocation, all of that."

[A] "Exactly. And that tension is really the core of what we are unpacking today in this deep dive."

[B] "Yeah. Our mission today is to go through a stag of insights from our source material that reveal a massive shift in how we think about software."

[A] "Right. Moving from building these permanent architectural monuments to creating what they call disposable code."

[B] "Disposable code. Yeah. Turning spreadsheets into intelligent financial engines and fundamentally changing the job of senior software engineers."

[A] "The way you think about your software budgets and your engineering talent is, honestly, it's about to be turned upside down."

[B] "Definitely."

## When Building Custom Software Takes Two Hours

[A] "So let's ground this. Let's look at the specific example in the research. There was a user who needed a project tracking tool, but the requirements were incredibly narrow. It was only meant to be used by three people, and the entire project was only going to last for six weeks."

[B] "Right. And in the traditional model, if you're a manager looking at that requirement, you basically face a lose-lose scenario."

[A] "You really do."

[B] "I mean, you either purchase a heavy software as a service, which means you're paying thousands of dollars for a massive platform with, you know, a hundred features you don't even need. And you're locked into an annual contract for a six-week project."

[A] "Exactly. Or the other option, you force your team to manage everything in a rigid manual spreadsheet that eventually just becomes a total nightmare to update."

[B] "Right. The classic version 2 final .xlsx problem."

[A] "Yeah, exactly. But instead of doing either of those things, this user just had an AI build the exact tool they needed."

[B] "Just for them."

[A] "Yeah. It took two hours to create. They ran the project for six weeks, and when they were done, they literally threw the code in the trash."

[B] "Just deleted it."

[A] "Gone."

[B] "Gone. And the reason this is suddenly viable comes down to the underlying economics of why the SaaS industry even exists in the first place."

[A] "Okay. Unpack that a bit."

[B] "Well, companies like Salesforce or Workday, they thrive because human programming is staggeringly expensive. You have to pay teams of highly skilled developers to write, test, and maintain code over years. So to make that math work, the software company builds the platform once and spreads that massive development cost across millions of users."

[A] "Right. It's cheaper for you to just rent a tiny fraction of their system than to build your own from scratch."

[B] "Exactly. But, and here's the paradigm shift, the moment an AI can write functional code on demand, the cost of generating that code drops to near zero."

[A] "Near zero. Wow."

[B] "Yeah. You don't need to spread the cost across thousands of users anymore. You can build a hyper-specific tool for, like, three people on the finance team for a two-week sprint, and the economics still work perfectly."

[A] "Okay. So the bottleneck shifts completely."

[B] "Exactly. We are no longer constrained by the cost of how we build software. The new constraint is what we choose to build. The value moves away from typing technical syntax and moves toward conceptual clarity."

[A] "Meaning you have to actually know what you want."

[B] "Precisely. It requires the product owner to deeply understand the business context and communicate that intent clearly to the AI model."

## The Shadow IT Nightmare

[A] "Okay. Well, I'm going to push back heavily here."

[B] "Go for it."

[A] "Because as a business leader, if I tell my entire organization, hey, everyone, you can suddenly build your own custom temporary apps whenever you feel like it, that sounds like a catastrophic shadow IT nightmare. I mean, marketing is building unsupported tools. HR has their own hidden databases. None of it is vetted for security. If traditional SaaS is the concrete foundation of a house and this disposable code is like a pop-up tent, I mean, I cannot build the core infrastructure of my company out of pop-up tents. The whole thing will just collapse."

[B] "And that is a perfectly valid fear. In fact, the sources actually validate that exact concern."

[A] "Okay. So I'm not crazy."

[B] "No, not at all. This is not about tearing down your core infrastructure. The permanent, stable elements of your business, your central databases, security protocols, customer records, that is your concrete foundation. You still rely on robust SaaS and traditional engineering for those heavy, critical layers."

[A] "So the foundation stays concrete."

[B] "Always. The disposable AI code is meant for the thin, concept-driven layers on top. It's for localized, fast-moving problems that, frankly, do not need to exist for 10 years. The new skill for leaders is ruthlessly distinguishing between the two, you know, having the courage to treat the thin layers as disposable while fiercely protecting the permanent foundation."

## Spreadsheets Become Autonomous Analytical Engines

[A] "Okay. So if the business side is now empowered to handle those thin layers, we have to look at the tool where they already spend most of their time."

[B] "Right. The spreadsheet."

[A] "It's a spreadsheet."

[B] "Yeah. The sources detail this massive shift where AI isn't just generating standalone apps, it's completely transforming the spreadsheet into an autonomous analytical engine. It takes the ultimate business canvas and changes it from a static organizer into, like, an active participant in the analysis."

[A] "Yeah. There is a really detailed case study in the material about building a car leasing model."

[B] "Oh, right. Vehicle leasing is notoriously complex. You're balancing contract lengths, residual vehicle values, strict mileage limits, varying tax deductions."

[A] "That's a lot of variables."

[B] "Yeah. And building a robust comparison usually takes a season to analyst days. And even then, you're terrified you hard-coded a wrong number somewhere."

[A] "Exactly. But in this case study, the user opened Excel integrated with Anthropic's Claude. And in exactly 15 minutes, the AI constructed an eight-tab model comparing three different cars across three different contract lengths."

[B] "Fifteen minutes."

[A] "Yep. And it built a sensitivity analysis, meaning it automatically set up scenarios to show the absolute worst-case and best-case costs if the user drove, you know, way over or way under the expected mileage."

[B] "The speed is impressive, obviously. But the part that changes the game for me, the real aha moment, is the AI acting on unprompted context."

[A] "Yes. That was huge."

[B] "Because the user just casually mentioned they were a sole trader. They didn't explicitly ask for tax advice or for a whole tax model to be built."

[A] "Right. But the model instantly recognized the financial implications of that specific employment status. It independently created a new tab dedicated solely to calculating tax effects. Completely unprompted. It compared the costs for a sole trader versus a limited company. It even paused the process to ask the user for their exact marginal tax rate."

[B] "Wow. It was anticipating the next logical business requirement without needing to be told."

## Cell-Level Intelligence and the Data Advantage

[A] "Okay. Let's get into the mechanics of how that actually works, though, without completely destroying the workbook."

[B] "Good point."

[A] "Because, historically, when you ask an AI to interact with data, it just gives you a wall of text or, like, an isolated table. If it tries to modify a complex Excel file, it usually breaks a formula in cell B4, and suddenly the entire sheet is just showing hashtag REF error codes everywhere."

[B] "Yeah, it's a mess. So how is it actually integrating into the math safely now? Well, this is a massive technical leap. The AI is no longer just reading the raw text on the screen. It's actually mapping the structural relationships within the entire workbook."

[A] "Oh, I see."

[B] "It tracks dependencies down to the individual cell level. So when the AI decides to add that tax calculation, it understands that the total cost cell relies on the monthly payment cell, which relies on the interest rate cell. It sees the whole chain. So if you ask it to change an underlying assumption, it can trace that logic backward and modify the specific input without severing the dependency chain."

[A] "Which gives you full traceability. You can actually click through and verify the math rather than just blindly trusting a black box text answer."

[B] "Exactly. Traceability is what turns this from a neat parlor trick into a tool you can actually trust for serious business decisions."

[A] "Right. But there is another mechanism at play here that is quietly rewiring the financial industry, and that's data integration."

[B] "Right. For a long time, the tech world was obsessed with which language model was fundamentally smarter. But that's no longer the primary battlefield. The new competitive advantage is who has the data. Because a smart AI without current data is just, you know, confidently guessing."

[A] "Exactly."

[B] "And Anthropic has been securing these massive licensing partnerships with legacy financial data providers. We're talking about institutions like LSEG for real-time market feeds."

[A] "Right. The London Stock Exchange Group."

[B] "Moody's for credit risk. S&P Capital IQ. Morningstar. Pitchbook. So the AI can pull the live morning pricing straight from the market, cross-reference it with a credit rating from Moody's, and dynamically update the sensitivity analysis in your local spreadsheet all in real time."

[A] "Yes. And the efficiency gains from this are just difficult to overstate. There's a striking metric from Norway's sovereign wealth fund, Norges Bank Investment Management."

[B] "Oh, yeah. I saw this."

[A] "They estimate they have already saved 213,000 work hours just by deploying Claude inside their spreadsheet workflows."

[B] "213,000 hours. Which, I mean, that forces a rather uncomfortable realization, doesn't it?"

[A] "How so?"

[B] "If a robust financial model drops from three days of work to 15 minutes, and the AI is fetching the live market data and running all the calculations, what are we actually paying the human analyst to do?"

## Calculation Becomes Cheap, Judgment Becomes Expensive

[A] "Ah. Well, the sources frame this perfectly. They say calculation becomes cheap, judgment becomes expensive. The fundamental value chain of knowledge work is completely flipping."

[B] "So math is just a commodity now."

[A] "Exactly. The machine handles the math flawlessly. But the machine cannot look at that car leasing model and tell you if your baseline assumption of driving 10,000 miles a year is actually realistic based on your lifestyle."

[B] "Right. It doesn't know if I'm planning to move to the suburbs next year."

[A] "Exactly. Only the human can weigh those qualitative factors. The strategic value of having a flexible contract versus locking in a lower monthly cost for three years."

[B] "Right. The role of the professional shifts from being a manual calculator to being an evaluator of reality."

## What Happens to the 20-Year Veteran Engineer?

[A] "Okay. But if the business professionals are suddenly generating their own disposable applications and running their own complex models, that creates a massive friction point."

[B] "What do you mean?"

[A] "Well, what on earth is left for the 20-year veteran software engineer to do? The ones who maintain the core business systems. It feels like they would be the very first ones automated out of a job."

[B] "You'd think so, but the research argues the exact opposite."

[A] "Wait, really?"

[B] "Yeah. The popular narrative focuses entirely on the amateur, you know, marketing manager, hacking together a quick tool over the weekend. The sources refer to them as vibe coders."

[A] "Vibe coders. I love that term."

[B] "People who are just feeling their way through the process until the app works. And if a vibe coder makes a mistake, their temporary pop-up tent collapses. The app crashes, they restart it, and the damage is completely contained."

[A] "Low stakes."

[B] "Extremely low stakes. But when a senior engineer integrates AI into a legacy core system, the stakes are exponentially higher."

[A] "Oh, because they're working with the concrete foundation."

[B] "Exactly. They are working in environments that handle millions of live transactions, where a single inefficient database query can bottleneck an entire production server. Their blast radius is massive."

[A] "The blast radius. The amount of damage they can do with one mistake."

[B] "A catastrophic amount. The sources outline a nightmare scenario called a silent rewrite."

[A] "Okay. What is that?"

[B] "Imagine you unleash an AI agent to clean up a large code base. It works quickly, submits the changes, the human reviewer looks at it, everything seems to compile correctly, and they push it live."

[A] "Sounds fine so far."

[B] "Right. It sits in production for three months. But under one very specific, rarely used logical branch, the AI accidentally flipped a positive financial variable to a negative one."

[A] "Oh, man. So it's just quietly corrupting data for months before anyone even notices."

[B] "Exactly. Because the blast radius is so large, senior engineers have to adopt incredibly rigorous, deeply constrained management techniques. They can't just let the AI run wild."

## Commit Discipline and Context Window Management

[A] "So what do they do?"

[B] "Well, the first major shift in workflow is enforcing strict commit discipline."

[A] "Commit discipline. Let me see if I can puzzle this one out. A commit is when a developer officially saves a chunk of new code to the main system, right?"

[B] "Yep, that's right."

[A] "So if an AI is writing code 10 times faster than a human, it could potentially touch a dozen different files in an hour. And if it bundles all of that into one giant save, the human reviewer is just looking at a massive wall of changed text. You have the final answer, but you have absolutely no idea how the AI got there."

[B] "You hit the nail on the head. When a human reviews code, they need visibility into the logic between the steps. If the AI compacts hours of work into a single massive commit, that visibility vanishes. So the required discipline is forcing the AI to make one distinct save for every single logical change."

[A] "Okay, so you might end up with 20 micro commits instead of one big one."

[B] "Exactly. You need that granular trail to actually audit the machine's reasoning. It's like a math teacher forcing a student to show their work on a complex equation. The final answer doesn't matter if the underlying formula is flawed."

[A] "That's a perfect analogy. The sources also talk about managing an AI's context window. And honestly, this feels very relatable."

[B] "Yeah. AI's context window is the total amount of text and instruction it can hold in its active memory at one time."

[A] "To me, that sounds exactly like a human's working memory at like 4:00 p.m. on a Friday afternoon."

[B] "That is a remarkably accurate way to conceptualize it. I mean, early in the session, the model's memory is empty. It's sharp. It tracks all your variables perfectly. But by Friday afternoon, after you've been working on a complex problem all day, your mental budget is spent. You start dropping important details. You make careless mistakes."

[A] "Right. And when an AI's context window gets stuffed full of code, error messages, and chat history, it does the exact same thing. It degrades and starts hallucinating."

[B] "And the natural human reflex when the AI makes a mistake is to just keep typing in the same chat box."

[A] "Oh, yeah. We try to argue with it."

[B] "Right. We try to correct the error because we dread the idea of starting over and re-explaining the entire premise of the project."

[A] "But arguing with a tired AI just makes it worse, doesn't it?"

[B] "It actively poisons the context. The model relies heavily on pattern matching. When you argue with it in a degraded state, it begins to weigh its own previous errors and your frustration just as heavily as your original instructions. So the output just spirals."

[A] "Completely spirals."

[B] "The necessary discipline is to immediately stop. You close the session entirely."

[A] "Just walk away."

[B] "Walk away. You write a clean, concise, executive summary of what has been accomplished and what the current error is. Then you open a brand new session with a completely empty memory. Paste in the summary and start fresh."

[A] "Wow. You have to treat its memory as a strict budget. When it's spent, you just put it to bed and start a new shift."

[B] "Exactly."

## System Instructions as a Rigid API

[B] "There is one more critical workflow change for senior engineers, and it revolves around how they give instructions in the first place. The sources discuss a system file, usually called claude.md, and they insist it has to be treated like an API."

[A] "Right. An application programming interface. This is a massive mindset shift for developers."

[B] "Oh, so. Usually people treat system instructions like a casual notepad. They type things like, make sure the code is secure, or try to write clean functions. It is entirely subjective."

[A] "What does clean even mean to a machine?"

[B] "Exactly. That ambiguity is the root of the problem. For a senior engineer, those instructions cannot be casual suggestions. They must be treated with the exact same deterministic rigor as a public API."

[A] "Meaning they have to be precise."

[B] "Completely precise. The AI parses that specific markdown file before the chat even begins to establish its foundational behavior. Every rule in that document must be exact, unambiguous, and mechanically testable."

[A] "So if you're in a chat session and you find yourself having to remind the AI, hey, don't forget to use this specific database format, that isn't a problem with your prompt."

[B] "No, it is a failure of your system instructions. If you have to repeat yourself, there is a bug in your core interface. You don't just keep chatting. You go back and rewrite the markdown file to eliminate the ambiguity at the source."

## From Bricklayers to Building Inspectors

[A] "Huh. Looking at all these new disciplines, granular commits, strict memory budgeting, treating instructions as rigid APIs, it completely shatters the narrative that AI coding replaces senior engineers."

[B] "It does. It seems like it actually requires more engineering, not less."

[A] "The nature of the work evolves, definitely. The value of a veteran engineer shifts away from remembering syntax or manually typing out implementation details. Their value concentrates entirely on architectural clarity. It's about understanding failure models, knowing exactly how a system will break before the code is even executed."

[B] "They move from being bricklayers to being building inspectors."

[A] "Yes. Their job is to look at a complex AI-generated system, audit the structural integrity, and have the authority to say, I stake my professional reputation on this being safe to run in production."

## Judgment as the Final Moat

[B] "And we arrive right back at the core theme of the research, judgment."

[A] "Yep. The manual labor of typing and calculating is disappearing for everyone across the board."

[B] "Everywhere. Whether you're a product manager generating a temporary project tracker in two hours, or a business leader evaluating a massively complex financial model that took 15 minutes, or a 20-year veteran engineer auditing an AI's code submission. The physical act of creating the output is no longer the bottleneck."

[A] "What remains, and what is exponentially increasing in value, is human judgment. It is the ultimate asset a leader has left to cultivate."

[B] "Let's bring this all together. If you're listening to this, you need a clear, actionable way to apply this to your own operations. The next time you sit down to plan a project, you need to ruthlessly audit your workflows. Stop paying for expensive, permanent SaaS monuments when a localized, disposable AI tool will solve the problem for a fraction of the cost. Let the business teams build the thin layers."

[A] "Absolutely. But simultaneously, you must double down on protecting your core infrastructure. Empower your teams to use intelligent spreadsheets to accelerate their analysis, but explicitly train them that their primary responsibility is aggressively verifying the machine's assumptions against reality."

[B] "Right. And for your engineering teams, transition their focus from writing lines of code to architecting secure, verifiable systems. It is a complete paradigm shift in how we value work."

[A] "It really is. But before we wrap up, I want to leave you with a final thought to mull over. Throughout this deep dive, we've established that AI has commoditized the calculation, the coding, and the syntax. Human judgment is our final moat."

[B] "Right. You have to evaluate the car lease. The senior engineer has to verify the code submission. That is the boundary of our current reality."

[A] "But look at the trajectory we just mapped out. We are seeing models integrate deep, real-time data from global financial institutions. We are seeing tools that can trace their own mathematical logic down to the individual cell level in a spreadsheet and spontaneously suggest tax strategies based on unseen context."

[B] "Yeah, it's moving fast."

[A] "If these systems are rapidly becoming this adept at pulling real-world data and tracing their own reasoning, how long until they become sophisticated enough to start reliably auditing and evaluating their own judgment?"

[B] "That is the horizon line. The moment the AI stops just running the math and begins successfully challenging the premise of the equation itself. When it stops just building the temporary tent and starts telling you whether the ground is stable enough to camp on in the first place, think about that the next time you open a blank spreadsheet or approve a massive software budget. Are you truly paying for a monument or are you just paying for the illusion of permanence?"

[A] "Thanks for joining us on this deep dive into the changing landscape of code. We will catch you next time."