When The Platform Retires Your Workflow
4 Min ReadA mid-sized insurance firm built its entire claims review process on AWS Kendra in 2024. They trained their team of 14 analysts to structure queries through it, replaced two senior roles with contractors who managed the system, and reported 40% faster document retrieval in their first year. Last week, AWS moved Kendra to maintenance mode. No new features. No long-term roadmap. The system still works, but the clock is now ticking. The analysts who learned to depend on it are now scrambling to evaluate replacements while their leadership asks how they missed the warning signs.
They didn't miss them. There were none. That's the actual lesson.
- AWS moved three first-generation AI services to maintenance mode: Q Business, Kendra, and Bedrock Agents. These tools launched just two years ago. Companies that adopted early now face migration costs and workflow disruption with no advance notice that these products had expiration dates.
- Platform dependency is no longer a slow burn risk. It's a quarterly re-evaluation requirement. The velocity of AI product retirement has compressed from years to months, and enterprise buyers are learning that "cloud-native" also means "cloud-disposable."
Here's the system that prevents you from getting caught in the next platform shutdown.
First, map your AI dependency load. Open a spreadsheet. Column one: every AI tool your team or business uses daily. Column two: the provider. Column three: whether you could replace it in 30 days without breaking a critical process. If the answer in column three is no, you have a single point of failure.
Second, build parallel literacy. For every AI tool you depend on, one person on your team should be able to perform the same task manually or through a different system at 70% speed. Not as a backup plan. As a forcing function. If no one on your team can do the work without the tool, you don't control the work. The tool does.
Third, adopt with exit in mind. Before you integrate a new AI service into a core workflow, draft the replacement plan. Not if it shuts down. When its capabilities become table stakes and the provider moves upmarket or retires it. Write down what you would migrate to, how long it would take, and what data you need to export. If you can't answer those questions cleanly, don't integrate it into anything load-bearing.
Fourth, favor composability over convenience. Single-vendor AI stacks feel efficient. They are also traps. Choose tools that expose APIs, allow data export, and interoperate with competitors. Pay the integration tax now, or pay the migration cost later. The second bill is always higher.
Fifth, treat first-generation AI tools as rentals, not purchases. If a product launched in the last 24 months and doesn't have a clear enterprise moat, assume it has a three-year lifespan maximum. Use it to learn. Use it to accelerate. Do not use it to replace institutional knowledge or build dependency you can't unwind.
This doesn't mean avoid new tools. It means control the terms of the relationship. The insurance firm's mistake wasn't adopting Kendra. It was architecting their entire process around a tool they didn't control and couldn't replace.
The judgment AI can't replicate here is recognizing when efficiency becomes fragility.
Every platform will tell you integration is seamless. Most will deliver on that promise. What they won't tell you is that seamlessness runs in one direction. Getting in is easy. Getting out is expensive. The human skill that matters now is knowing the difference between a capability you're renting and a capability you're surrendering.
Speed is valuable. Autonomy is more valuable. The professionals who will stay in control over the next five years are the ones who use AI to move faster without forgetting how to move at all.
This week clarified the new rules. AI capabilities are abundant and temporary. Platforms will compete by speed, not stability. Providers will retire products faster than enterprises can adopt them.
Your strategic position doesn't come from picking the right tools. It comes from building systems where no single tool is irreplaceable. The winners in the next phase won't be the teams that adopted first. They'll be the teams that adopted with their eyes open, built redundancy into every dependency, and kept human capability in reserve.
Next week, we'll look at the opposite move: when to go all-in on a platform and how to know the difference. Until then, audit your stack. If losing one tool would break your workflow, you don't have a workflow. You have a liability.