At Echomind, we use self-hosted AI models for customer support instead of relying entirely on external LLM APIs. This gives us more control over model behavior, costs, updates and how customer data is handled, while allowing us to choose and combine different models depending on the task.
Ask any team that has been using AI in a production environment for long time, and they’ll all tell you the same thing: the model itself never keeps them up at night. But what happens when that model changes under their heels—that’s what keeps them up.
Where We Started
Our transition to self-hosting didn’t happen overnight. In the early stages, we tested a number of options: OpenAI’s GPT models, Claude, models via AWS Bedrock, and a few others. Initially, using an external API seemed like a sensible solution: we didn’t have to maintain any infrastructure, and we gained access to cutting-edge models without having to develop them ourselves.
The shift to self-hosting wasn't a single decision. It came out of a few separate problems that kept showing up as support volume grew.
The Dependency Problem
If you build your solution on someone else’s API, you automatically adopt their development strategy. The model you tested six months ago may not necessarily meet your customers’ needs today. Providers are constantly updating, phasing out, and retraining models, and the quality of support changes along with it—often without warning.
These are not mere assumptions. Users of various AI products regularly note that the model behaves differently than before: more cautiously, less consistently, or simply not the same way it did a few months ago. It is difficult to determine from the outside whether this is a real deterioration or simply a shift in expectations, and we are not claiming that any specific provider’s model has objectively become worse. But there is a technical reason why this happens: prompts and settings that were optimized for one version of the model do not always work the same way after an update. The model interprets instructions slightly differently, edge cases are handled differently, and a setting that worked reliably for several months may start to behave erratically, sometimes without any changes on your part.
And one thing remains true in any case: when a model is controlled by a third party, you have no control over whether it will change or when that will happen. For a system that interacts with real customers on a daily basis, this is a very inconvenient situation.
Self-hosting doesn’t eliminate the need for updates—we continue to improve and adapt over time. The difference is that updates are rolled out on our own schedule, are first tested in real-world support scenarios, and don’t appear in your customers’ conversations without prior notice.
The Economics Didn't Line Up Either
The prices for the most powerful external models were designed for infrequent, high-value queries, not for the volume generated by a real customer support service. When there are thousands of interactions per day, processing each conversation through a high-end API quickly becomes expensive. At the same time, cheaper third-party options weren’t actually cheaper in the sense that mattered: in terms of quality, they turned out to be roughly on par with the models we could deploy and configure ourselves.
This gap is where self-hosting became a worthwhile investment: support quality comparable to a premium third-party API, with a cost structure designed for high-volume daily conversations and pricing that we could pass on to our customers, rather than endlessly absorbing the vendor’s markup.
What Self-Hosting Actually Buys Us
Self-hosting does not mean that Echomind is permanently tied to a single internal model. It means that we choose the model best suited for a specific task and change that choice according to our own schedule, rather than waiting for a vendor to release a new version and being forced to accept whatever changes they decide to make.
In practice, this means that we control when and whether to change the model, and we set the cost structure ourselves without relying on the vendor to determine it. We have full visibility into what works and why, and we’re free to change or combine models depending on the task at hand, not stuck with a single vendor’s roadmap.
The quality of support shouldn’t change overnight just because a company you’ve never worked with updated a model somewhere else in the stack. That’s the whole point of this approach: the level of AI interacting with your customers is under your control, not subject to someone else’s pricing decisions or product changes.
Want to see how Echomind works with your support setup?
Run a pilot on your chat
We pick one conversation segment together and fix success criteria up front. No rip-and-replace.