The plugin is rarely responsible for it; execution takes just a very few milliseconds on the plugin side. Most of the time goes into the AI model writing its answer, and into the distance between you and your server, and between your server and your AI provider (OpenAI, Anthropic, Google, etc.), plus how busy that provider is.

Make sure that your server is not slow either; if you have a lot of plugins installed and doing things, that will also slow down the time before AI Engine actually starts handling the request.

The request itself also matters. If you are using Knowledge (embeddings), function calls, MCP servers, web search, big instructions or long conversations, each of these adds time before the answer is ready.

Things that help

• Enable Streaming (Settings > AI). The reply then appears word by word as it is written, instead of all at once at the end. It doesn’t make the model faster, but it feels much faster.
• Pick a faster model for simple chatbots. Thinking (reasoning) models take time before writing the first word: if your chatbot uses one, lower its Reasoning setting in the chatbot’s AI Model section.
• Check for buffering. If, with Streaming on, the reply still appears all at once, very late, something between your server and the visitor (a proxy, a CDN, a cache or a security layer) holds the response until it’s complete. See When Your Host Blocks AI Engine.

Very long replies

Since AI Engine 3.8.4 and 3.8.5, streamed replies are no longer cut after 5 minutes while they are still arriving, and the connection is kept alive during long silent phases (like Code Interpreter or a slow function call), so Cloudflare or your host’s proxy doesn’t close it. If the visitor leaves the page, AI Engine stops the request to the AI provider, so you don’t pay for an answer nobody reads. If replies still stop after a minute or two, make sure AI Engine is up to date, then ask your host to allow requests of at least 120 seconds.