Which Smart Home Automations Need AI in 2026?
In our reading of the documentation, most smart home routines need no AI. Motion, schedule and state rules are ordinary Home Assistant automations, and an AI conversation agent is a separate, optional layer. Assist can work locally, and a language model is optional for more conversational requests. Home Assistant says Assist can run fully on your own hardware, so your voice commands stay private. But its simplest documented path uses cloud speech voices, and an Ollama model whose name ends in :cloud runs on Ollama’s servers. The useful question is where each step processes data.
This guide is based on official Home Assistant, Ollama, Android, ESPHome and Frigate documentation. We did not run these setups hands-on.
Which Automations Need AI at All?
Routine automations do not need a language model. Motion, schedule and state-based automations are rules that trigger, check conditions and act. The Home Assistant Ollama integration is a separate conversation agent, and it does not integrate with sentence triggers. Add an LLM only where open-ended conversation earns its place.
Model sizes and hardware fit are covered in our guide to running LLMs on smart home devices. This post is about the data path, which matters more than the hardware once you add a voice layer.
Where Voice Data Goes: The Two Documented Assist Paths
Home Assistant documents two voice paths. The local path is an Assist pipeline on your own hardware, where your spoken commands never leave your home. The other is the Home Assistant Cloud speech voices, which the local setup guide calls the simplest path.
The Assist overview page labels the choices “I plan to process my voice locally” and “I plan to use Home Assistant Cloud (recommended as it is the simplest)”. That recommendation is about effort, not privacy. Choose the pipeline by reading its speech stages, not by the assistant’s name.
For the local route, the documented options are Speech-to-Phrase or Whisper for speech-to-text, and a local text-to-speech engine such as Piper. Our privacy-first build log covers the server side of that setup.
Wake Words: Detection Is Local, but Speech Is Not Automatically
ESPHome implements an on-device wake word detection framework, so the wake word itself can be detected on the satellite. Once the wake word fires, ESPHome devices with a microphone stream the audio to Home Assistant to be processed there by Assist.
So the wake word and the speech processing are separate questions. The wake word can stay on the device while the recognized speech is handled by a cloud pipeline. Home Assistant lists the Voice Preview Edition as its recommended voice hardware, but for any other device, check the vendor’s privacy page before assuming where audio goes.
The Ollama Conversation Layer: What the AI Can Touch
The Ollama integration adds a conversation agent in Home Assistant powered by a local Ollama server. It needs an external Ollama server reachable over the network. Control is experimental. The “Device control” tools reach only the entities you expose to voice assistants, while the separate “Home Assistant” tool accesses everything and only works for administrators.
The documentation recommends exposing fewer than 25 entities, because smaller models are more likely to make mistakes. Only models that support Tools can control Home Assistant. The integration’s default context window is 8k, against Ollama Server’s 2k default. Its keep-alive default is -1, which keeps the model in memory after a message, so a loaded model stays resident on the Ollama host.
Ollama :cloud Models Run on Ollama’s Servers
Ollama describes its cloud models as running on Ollama’s servers, so they are fast on any computer and need no download. In the CLI, a cloud model name ends in :cloud, for example gemma4:cloud, and those models need usage credits or a paid plan.
Installing the Ollama server on your own computer does not make every model it serves local. The model name decides where inference runs. The Home Assistant documentation does not discuss cloud models, so the sources used here do not establish whether the integration will serve a :cloud model. Check the exact model name you configure, and treat a :cloud suffix as a cloud destination.
The “Most Private” Analog-Phone Tutorial Depends on Its Pipeline
Home Assistant’s analog-phone tutorial, the world’s most private voice assistant, relies on a handset that will only listen when you pick up the horn. Its required material lists either a Cloud assistant pipeline or a manually configured local assistant pipeline.
The handset controls when listening starts. The pipeline you choose decides where the speech is processed, so the privacy result depends on that choice, not on the name of the tutorial. The tutorial leaves that choice to you, which is why the checklist below asks you to read the pipeline’s speech stages. A later step gives the assistant a personality through the OpenAI integration, which is a cloud AI path.
Your Phone’s On-Device AI: Gemini Nano and Hybrid Routing
Android’s Gemini Nano works without needing a network connection or sending data to the cloud. AICore does not have direct internet access, and it doesn’t store any record of the input data or the resulting outputs after processing them.
Apps can still choose cloud routing. Android’s own guidance describes apps that automatically route requests to Gemini Flash or Gemini Pro models in the cloud when Gemini Nano is unavailable. ML Kit GenAI inference is also limited to the app in the foreground: background use returns BACKGROUND_USE_BLOCKED. For a companion app, read its privacy policy before assuming a request stays on the phone.
Camera Detection: Where Inference Runs
Frigate is designed to run object detection on a dedicated AI accelerator or GPU rather than the CPU. That describes where inference runs. It does not describe where video goes, and the Frigate documentation page we read does not address vendor cloud features. If a camera uses a vendor’s detection service, check that vendor’s privacy page. For broader hardening, see our smart home privacy and security guide.
The Data-Flow Checklist: Verify Each Layer
Use the named setting for each layer rather than a general “local” or “private” label.
| Layer | Documented local path | Cloud or :cloud destination |
How to confirm |
|---|---|---|---|
| Wake word | On-device framework, micro wake word | Audio after wake streams to Home Assistant, where the pipeline decides the processing (ESPHome voice assistant) | Check the device config and the pipeline’s speech stages |
| Speech-to-text | Speech-to-Phrase or Whisper, run locally | Home Assistant Cloud speech voices (local setup guide) | Check the speech-to-text engine in the Assist pipeline |
| Conversation agent | Ollama server you run, with a model name without :cloud |
:cloud model runs on Ollama’s servers (Ollama) |
Read the exact model name you configured |
| Text-to-speech | Local engine such as Piper | Home Assistant Cloud text-to-speech voices (local setup guide) | Check the text-to-speech engine in the same pipeline |
| Phone-side AI | Gemini Nano via AICore | App-chosen cloud routing when Nano is unavailable | Read the app’s privacy policy; ML Kit GenAI requires the app in the foreground |
| Camera detection | Designed for a dedicated accelerator or GPU (Frigate) | Not covered by these docs; vendor cloud paths vary | Check the vendor’s privacy page |
FAQ
Does Home Assistant Cloud mean my voice commands leave my house?
On the Cloud path, speech-to-text and text-to-speech use the voices included with Home Assistant Cloud, so those stages run on cloud services. Home Assistant also documents Assist that can run fully on your own hardware. The Cloud option is labelled the simplest path, which is a statement about setup effort rather than privacy.
How can I tell whether my Ollama model is running locally?
Read the model name. Ollama says a cloud model name ends in :cloud, for example gemma4:cloud, and cloud models run on its servers. Models without that suffix run on the computer hosting Ollama, where speed depends on hardware. The Home Assistant documentation does not cover cloud models, so verify the exact name you configure.
Do motion lights and schedules need AI to work?
No, in our reading. Motion, schedule and state automations are rules with triggers, conditions and actions. The Ollama integration is a separate conversation agent and does not integrate with sentence triggers. Add an LLM agent only for open-ended, conversational requests, such as asking a question in plain language.
The Bottom Line
Keep routine automations as rules, and add an LLM only where open-ended conversation is worth the extra data path. Assist can run fully on your own hardware, but the simplest documented path, Home Assistant Cloud speech, and Ollama :cloud models send data to cloud services. Run the checklist after setup and again after any change to a pipeline, model or companion app.
How This Guide Was Built
This guide is based on official Home Assistant, Ollama, Android, ESPHome and Frigate documentation. The main pages were fetched in full on 2026-11-12. Some linked repositories and pages, such as the Speech-to-Phrase, Whisper and Piper repositories, were checked for HTTP status only. We did not run these setups hands-on, quote no prices the sources do not publish, and omit performance figures that vendors do not publish. Model and hardware detail lives in our guide to running LLMs on smart home devices and in our Home Assistant build log.
← Back to guides