The AI Architectural Trap: Avoiding One-Way Doors
Stay updated with us
Sign up for our newsletter
When I was at Amazon, I learned the “one-way door vs. two-way door” framing, a simple guide for hard decisions: one-way doors are irreversible choices that deserve deep scrutiny, two-way doors are reversible ones where speed matters more than perfection. This somehow became one of those management classics you can now bump into in different places such as the famous silent reads of six-pager narratives, two-pizza teams, and working-backwards press release I’ve seen engineers rely on it for architecture decisions, and sat in leadership meetings where the CEO invoked it to arbitrate strategic calls. Now, in the AI-era, the “one-way door vs two-way door” framing has become more relevant than ever.
The trap of moving (too) fast
Most companies have successful AI pilots by now, and many have early wins. But those wins were relatively easy to get. AI transformed software development first, and building something quickly is essentially a solved problem. The real challenge starts after that: scaling, meeting enterprise standards, iterating fast enough to match user expectations as models and capabilities keep shifting.
If the pilots and early wins were built on disposable foundations, and they usually are, teams will hit a wall. You can’t integrate new models or ingest new data if you’ve hard-coded your assumptions from day one. Without explicit guidance, AI development tools won’t build around two-way doors. I’ve seen teams ship AI-generated code that nobody could fully read or untangle. That’s a new kind of one-way door, one you didn’t even know you were walking through. Building principles and intentional architectures will matter as much as “vibe” in the AI era. Teams won’t reap the benefits of AI in a sustainable way if they build once to ship, rebuild for the real world, and then do it again when the ecosystem shifts.
The data decision you’ll regret first
The earliest and most irreversible decisions are made in the data layer. This is the very first one-way door developers will encounter. Predicting data volumes over an application’s lifetime has always been hard, and well-architected systems are built to handle elasticity and evolving data formats. Today, data is created and consumed by machines and is growing at an unprecedented pace. An application may start dealing with text and become multi-modal after a while. Schemas change constantly as industry standards are still forming. I’ve seen production teams plan a full application downtime just to change a data structure. That’s the one-way door you didn’t see coming. If integrating a new data provider means rewriting your pipeline, or if adding a metadata field requires a migration, you’re in trouble, and you may not realize it yet when celebrating the launch of your first AI project.
Data isn’t just growing and evolving. It’s also becoming constrained. Where it lives, how long it’s retained, and who can access it are increasingly dictated by regulatory and sovereignty requirements. I’ve seen cloud-first customers quietly move workloads back on-premises. Not for cost reasons. Because they didn’t trust AI with their most sensitive data. If your data model assumes a single region or provider, adapting later can mean a full rewrite. That’s a costly one-way door you didn’t plan for.
You don’t have a dog in this AI model fight
We can’t assume the best model today will be the best one tomorrow, or the right one for what your application will eventually need. Model lock-in is an unnecessary bet. If you build your entire app logic around a specific version of Codex or Claude, what happens when a smaller model is faster for your latency-sensitive customers? Or when a domain-specific model would boost accuracy for a particular use case?
We’re also in the infancy of AI economics. Token pricing, GPU supply constraints, and the economics of large general-purpose models versus small, specialized ones will all keep shifting. Building with model flexibility in mind just requires thinking about it from the start. It will pay off.
Model choice isn’t only about quality or cost. In some environments, it’s dictated by where the model runs and how data is handled. If switching models means moving data across regions or providers, what looked like a technical decision quickly becomes a compliance problem. That’s another form of lock-in, and another one-way door.
“And you say there are agents now?”
There’s so much to be excited about with agents: the autonomy, the leverage, the potential for continuous improvement. But there’s still so much that’s undefined. How tight should the harness be? Who decides? Who’s responsible when an agent goes off script? Will you even know it did? How do you measure whether an agentic system is improving or quietly degrading?
The battle for the agentic platform layer is just getting started, with hyperscalers, data platforms, and AI framework players all staking their claim. You can’t avoid some bets here, and 99% of companies shouldn’t build their own. Everyone needs to move fast, and there will be one-way doors. Just go in with eyes wide open, understanding the cost and feasibility of changing your mind. Be especially intentional in areas that are being completely reshaped by the agentic revolution and where winners aren’t certain yet, such as observability, identity, access controls, and orchestration.
Adaptation is the only moat
The AI landscape twelve months ago looks nothing like today. Anything tightly coupled then is already outdated. Anything tightly coupled today will be outdated a year from now.
The bottleneck used to be “can we build it?” AI solved that. The new question is “will we have to unbuild it?” That shift is subtle, but it changes everything. Moving fast used to mean shipping faster than your competitors. Now it means staying reversible while they lock themselves in.
Not every one-way door can be avoided. This marathon is starting at sprint speed, and everyone is moving fast, including your competitors. But there’s a difference between the one-way doors you walk through deliberately, eyes open, and the ones you don’t notice until you’re already through them. The “one-way door vs. two-way door” framing doesn’t ask you to slow down. It asks you to know which one you’re choosing.
