1. Elixir is still a pretty leaky abstraction over Erlang. In my experience it's not enough to learn just Elixir, I regularly had to dive into Erlang library source code to debug or answer questions. This somewhat negates the benefit of a small, stripped down syntax when you often have to learn another one in conjunction.
2. Maturity of ecosystem. This should improve over time but I've had challenges finding high quality libraries, especially for things like database connection drivers or making network requests. It's often hard to tell how well-supported or complete a library is and regressions were a regular occurrence.
3. Documentation. In practice I rarely found official documentation complete or even helpful (outside of big projects like Phoenix / Ecto). Even core Erlang libraries had surprising chunks missing. It's been awhile but I remember it being very hard to figure out what options were supported in Erlang's TLS wrapper. I ended up stitching together pieces from the OpenSSL documentation, Erlang source code, and lots of trial and error.
4. OTP overlap with other scheduling systems. This isn't a design flaw as much as a potential footgun depending on how you deploy Elixir code, but there is a lot of overlap between the cluster management support in Erlang/OTP and, for example, container orchestration in Kubernetes. Both support scheduling some concept of tasks, configuring redundant copies of tasks, detecting problems, and restarting tasks according to policy. Deploying an OTP application on top of Kubernetes (on top of Linux) results in 3 layers of OS-like task scheduling to learn, teach, maintain, and debug through, all with distinct tooling, terminology, and limitations.
All in all, I found Erlang/OTP to be a pretty interesting and compelling platform, especially for certain special purpose applications. If I ever use it again I'll probably skip the extra layer of Elixir syntax and write straight Erlang.