What Parts of an AI-Powered Home Can Keep Working During an Internet Outage?

An “AI-powered home” can mean two very different things.

One home sends speech, reasoning and device commands through outside services.

Another keeps devices, automations, voice processing and even the language model on local hardware.

During an internet outage, those architectures behave very differently.

Layer 1: local devices

Devices using local protocols can keep communicating when the WAN disappears, provided the local controller and network remain available.

Home Assistant can operate local devices using technologies such as Zigbee, Z-Wave, Matter, Thread and ESPHome depending on the hardware and integration.

Vendor-cloud devices are different.

If their normal control path requires the vendor's servers, losing internet can remove that path.

Layer 2: deterministic automations

Local Home Assistant automations can continue running without internet when their triggers, conditions and actions are local.

That includes useful work such as motion-driven lighting or local sensor responses.

These deterministic automations should remain the foundation for important predictable behavior.

An LLM is not necessary to turn a light on when a known sensor changes state.

Layer 3: local speech

Voice control does not automatically require the cloud.

Home Assistant supports fully local Assist pipelines using local speech-to-text and text-to-speech components.

If the microphone device, Home Assistant host and local speech services remain available, basic spoken control can continue during a WAN outage.

The actual setup must be tested, because individual voice devices or integrations can still depend on external services.

Layer 4: the local language model

Home Assistant has integrations for local conversation models including Ollama and llama.cpp.

A local model can keep generating responses while the internet is down if the model, integration and local AI host are all available.

Home control through a language model deserves caution.

Only expose the entities that need to be available, and keep high-consequence or safety-sensitive behavior in deterministic automation rather than relying on model interpretation.

Layer 5: outside tools and data

A local model does not magically create live internet data.

Web search, outside weather APIs, cloud AI providers, remote-access services and vendor-cloud integrations still need connectivity.

The LLM may remain able to reason about local context while losing the tools that normally provide fresh outside information.

That distinction should be visible to the user.

The LAN still matters

“No internet required” does not mean “no infrastructure required.”

The router or switches may still be carrying local traffic.

Home Assistant still needs its host.

The local AI model still needs its computer.

DNS or local name resolution may still matter depending on the setup.

A WAN outage is only one failure mode.

The AI server can fail independently

Design a graceful fallback.

If Ollama or another local model stops working, Home Assistant should still run ordinary automations.

If voice processing fails, wall switches, buttons, dashboards and normal controls should remain available.

AI should add capability without becoming the only door into the house.

Test by disconnecting the WAN

Do not infer offline behavior from architecture diagrams.

Disconnect the WAN connection while leaving local power and networking intact.

Test:

• local switches and sensors;
• deterministic automations;
• voice transcription;
• text-to-speech;
• the local conversation model;
• selected device control;
• cloud-dependent functions.

Write down what fails.

That gives you the real dependency map for your home.

For the non-AI version of this problem, see What Happens to a Smart Home When the Internet Goes Down?. For the local-AI side, see Can an AI Assistant Still Work When the Internet Is Down?.