Everyone I talk to right now is focused on the same things when it comes to AI. Which model to use? How to structure prompts? Do I fine-tune or go with RAG? How many data scientists to hire? These are real questions worth asking. But in my experience, they're not the question that's actually blocking most organizations.
The thing that's blocking most organizations isn't the AI. It's that the AI has nowhere to go.
I've been in software long enough to have watched the same pattern play out with every major architectural shift. When the web came along, companies discovered their core business logic was trapped inside desktop applications with no HTTP interface. When mobile happened, the same logic was trapped behind web UIs that only rendered HTML. Now AI is happening, and the logic is trapped inside monoliths and stored procedures that can't be called from the outside without triggering a cascade of unintended consequences. In some Notebook LM generated research this was referred to as Logic Entombment (1), and that concept really described this core problem perfectly. It's when your critical business rules, like your pricing logic, your eligibility checks, or your workflow validations, are so tightly fused with the infrastructure around them that no outside system can call them cleanly. The logic still runs. It just can't be reached cleanly or easily.
Here's a concrete way to think about it. Imagine you want to build an AI agent that helps sales reps generate quotes. Seems straightforward. The AI needs to check pricing rules, apply discount eligibility, and validate against product configuration constraints. Simple enough in theory. Now imagine all of that logic lives inside a stored procedure in your order management system. And that stored procedure doesn't just calculate prices, it also writes to the audit log, fires the UI refresh event, and sends a notification email to the assigned account manager. You can't call the pricing logic without triggering everything else. There's no read-only mode. There's no clean interface. The AI hits a wall.
This is not a hypothetical. I've seen variations of this at multiple organizations. The AI project stalls not because the model was wrong or the data was bad, but because the business logic was never designed to be called by anything other than the application it was built inside.
The frustrating part is that the logic itself is usually fine. It's correct. It encodes years of hard-won business rules that actually work. The problem isn't the rules, it's that they're buried deeply in a pile of other business rules that are tangentially related. They're doing their job perfectly and are completely inaccessible to anything new at the same time. They're entombed, buried alive.
This matters more now than it did even a couple of years ago because of how AI agents work. A copilot that assists a human can tolerate a lot of friction. A human can navigate the UI, copy a value from one screen and paste it into a prompt, and manually fill in the context. That's not great, actually, let's be honest, it's horrible, but it works.
Agents are different. An agent needs to call discrete functions with specific inputs and receive structured outputs. It needs to invoke your pricing logic the same way your existing application does, without triggering side effects and without going through the UI. If it can't do that, it can't do its job. End of story.
This is the door problem. You can have the smartest model in the world, and it still can't walk through a wall.
The good news is that this is a solvable problem. But it requires being honest about what it actually is. It's not a data problem. It's not a model problem. It's an architecture problem that has been quietly accumulating costs for years, and AI just made it impossible to ignore.
Over the next few posts, I'm going to get into what makes entombed logic specifically hostile to AI agents, how to diagnose where your organization actually sits today, and what to do about it without blowing up your roadmap. But the starting point is just recognizing the real blocker for what it is.
Your AI doesn't need a better brain. It needs a door.
(1). Quick note about Logic Entombement. I was using Notebook LM to do research for this article, and it threw out that label. I really liked it. I searched to see if it could find the source so I could give it the credit it was due, but I couldn't find anything. It may just be a pithy thing Notebook LM came up with (in which case nice job), or it may have lost the original reference somewhere, in which case my apologies to the source.