Power Automate connectors were training wheels.
That is not an insult. Training wheels are useful. They let you go somewhere you could not go without them. The question is whether you still need them.
The short answer is: increasingly, no.
What connectors actually solved
The problem connectors solved was specific. Non-developers cannot write HTTP calls. They cannot deal with authentication headers, pagination, response parsing, or API versioning. So Microsoft built a wrapper. Each Power Automate connector translated a complex API surface into a set of pre-baked triggers and actions, something you could configure with dropdowns instead of code.
That democratised integrations for a class of user who would otherwise have no access to them. Business analysts, operations leads, and admins who had never opened a code editor could now connect Salesforce to Teams, or SharePoint to an approval flow. Four hundred plus connectors later, the platform looked impressive.
But it was always a bridge. The connector was not the destination. The connector existed because non-developers needed to call APIs and could not. Remove that constraint, and the connector stops being necessary.
The connector tax
Every abstraction has a cost. With Power Automate connectors, you pay in a few ways.
Capability gaps. The connector is never the full API. It exposes what Microsoft's connector team decided to expose, in the shape they decided. If the underlying API has a field or endpoint the connector does not cover, you either go without it or drop down to raw HTTP actions, at which point you are writing the wrapper yourself anyway.
Version lag. APIs evolve. Connectors update on their own schedule. New fields, new endpoints, deprecations: the connector catches up eventually. Sometimes.
Licensing tiers. Dozens of connectors are premium: Salesforce, ServiceNow, SAP, Dynamics 365 itself. Connecting those systems requires an additional license. The connector model does not just expose functionality; it prices it by connector.
For many use cases, Power Automate connectors are still the right tool. But they come with a ceiling, and that ceiling is lower than most people realise.
What AI changes
Two recent releases from Anthropic show the direction clearly.
Claude Routines are automated Claude Code tasks that run independently — no device needs to stay active. You configure a task once: a prompt, a repo, and what should trigger it. Three trigger types are available: scheduled (hourly, nightly, weekly), API trigger (an HTTP POST to a dedicated endpoint), and webhook (GitHub events like pull requests). Real teams are using these for nightly bug triage, deploy verification, alert triage, and automated code review.
What matters here is not the scheduling. It is that AI is now running complex, multi-step tasks autonomously, calling external systems without any pre-built connector wrapper standing in between.
Claude Managed Agents takes this further. It is a platform for deploying cloud-hosted AI agents in production: long-running, with persistent state, capable of spawning other agents to handle parallel work. Companies like Notion, Asana, and Sentry are running these now. The agents call tools, interact with APIs, and handle multi-step workflows, not through a library of pre-built connectors, but by understanding what the system can do and constructing the interaction from that understanding.
Model Context Protocol is a clean standard for exposing tools to AI agents, it makes integration easier and more predictable for AI to work with. MCP was specially developed for AI to work with API's, but it is not a prerequisite.
AI can make HTTP requests directly, read API documentation, and handle authentication without any protocol layer in between. This works even without MCP. The connector was needed because AI could not call APIs. Now it can. MCP or not.
The wrapper was a solution to a problem that no longer exists.
What actually survives
Connectors becoming optional does not mean Power Platform becomes irrelevant. The useful parts were never the connectors themselves.
Dataverse survives. Not because of the connector to it, but because of what it holds: your data model, your relationships, your business logic embedded in the platform. The schema of your lead qualification process, the custom tables built over five years, the plugin logic that encodes institutional decisions. None of that is in the connector. It is in the platform.
Governance and compliance survive. Entra ID, audit trails, data loss prevention policies, tenant-level access controls. These do not disappear because an AI agent can call an API. Regulated industries need to know who accessed what and when, and that requirement is not going away.
Domain knowledge survives most of all. An AI can call the Dynamics 365 API. It cannot know why the lead qualification process works the way it does in your specific tenant. It cannot know that the field called "annual revenue" is actually updated by a manual sync from a spreadsheet that runs every Tuesday. It cannot know that the automation which looked broken last quarter was actually correct, because the business rule changed mid-year and nobody updated the documentation.
The connector was a translation layer between business intent and API calls. AI does that translation natively now. What remains valuable is knowing what to translate: the domain knowledge behind the intent.
The real disruption is the inversion
The disruption is not connectors disappearing. It is the inversion of how automation gets built.
Tools used to make you think like the platform. You learned the connector model, the trigger types, the action shapes, the naming conventions. You shaped your requirements to fit what the tool could express.
Now the platform thinks like you. You describe intent. The system constructs the implementation.
"Flow builder" as a skill set becomes less valuable over time. Knowing which Power Automate connectors to chain together and in what order is a procedural translation task. AI commoditises procedural translation tasks first, because that is exactly what it is good at.
"Automation strategist" is the skill that compounds. The ability to describe what a business actually needs, to know whether the output is correct, to catch the edge cases that live in institutional memory rather than documentation — that is not something AI can replicate from the outside. That is where the value concentrates.
Intent over instructions. That is the shift.
The question worth asking
Power Automate connectors are still useful today. This is not an argument to throw away what works.
But if you are spending time building expertise around which connectors exist, how to configure them, and how to work around their limitations, that investment has a different return horizon than it did three years ago.
The question worth asking: are you building skills around the connector library, or around the business problem it was built to solve?
One of those compounds over the next five years. The other does not.
