Guide
Local AI vs Cloud Assistants: What "Private" Really Means for Your Home
A cloud assistant sends your requests to a company's servers to be understood and answered; a local assistant does that work on a computer inside your home. Local is more private and keeps working without the internet, but it needs real hardware and makes more mistakes on complex requests. Here's how to judge which you need, and what to ask anyone who calls their product private.
The difference in one picture
Every AI assistant does the same three jobs: it receives your request, works out what you mean, and acts or answers. The question is where the middle step happens.
- Cloud assistant: your request, whether voice, text or both, travels over the internet to the provider's data center, where large models running on the provider's hardware interpret it. The answer or instruction comes back to your house. Alexa, Google Assistant and most AI chat apps work this way.
- Local assistant: the model that interprets your request runs on a computer in your home. The request is understood and handled there and doesn't need to leave the house.
That difference sounds technical, but it decides four things you'll care about: who can see your requests, what happens when the internet goes down, how capable the assistant is, and what you pay over time.
Side by side
| Cloud assistant | Local assistant | |
|---|---|---|
| Where requests are processed | The provider's servers | A computer in your house |
| Who can access request data | The provider, under its policies, which can change | You, and anyone you give access to |
| Works without the internet | No | Yes, for everything inside the house |
| General knowledge and reasoning | Very strong | Good and improving, but smaller models make more mistakes |
| Speed | Fast, but depends on your connection | Fast on good hardware; slow on weak hardware |
| Up-front cost | Low: a speaker or an app | Higher: dedicated hardware |
| Ongoing cost | Often free or a subscription, plus your data | Electricity and upkeep; no per-request AI fees |
| Who decides when it changes | The provider | You |
Why "who decides when it changes" matters
The last row in that table gets overlooked, and recent history shows why it matters. In March 2025, Amazon removed an Echo setting that let a few devices process requests without sending them to the cloud, citing the needs of its new generative AI features. Customers who'd chosen that option for privacy reasons didn't get a vote.
That's not a criticism unique to Amazon. With any cloud assistant, the provider decides what is processed where, what's kept, and what the product does next year. That's fine for asking about the weather. It matters more when the assistant knows your family's schedule, unlocks your doors and reads your email.
What a local AI assistant can do today
Local AI has moved quickly. Open models, freely downloadable models from companies and research groups, can now run on hardware that fits in a closet. Ollama, one of the most popular tools for running open models, lets a home computer host them and make them available to other software on the home network.
Home Assistant, the open-source smart-home platform built around local control, ties this into the house. Its Assist voice assistant can run fully on your own hardware, and its Ollama integration lets Assist use a local model as its brain. With control enabled, the model can operate the devices you choose to expose: lights, locks, thermostats, shades and so on.
In practice, that means a local assistant can handle requests like:
- "We're going to bed. Lock up and turn everything off downstairs."
- "Is anyone's door unlocked, and is the garage closed?"
- "What's on the family calendar this weekend?" (if you connect the calendar)
- "Make the guest room comfortable for Friday night."
The pieces of a local assistant
A fully local voice assistant is really several small systems working together, each of which can run in the house:
| Piece | Job | Where it runs in a local setup |
|---|---|---|
| Wake word | Notices when you're talking to it | The speaker or satellite device in the room |
| Speech-to-text | Turns your voice into words | The home computer |
| Language model | Works out what you mean and what to do | The home computer, through a tool such as Ollama |
| Home hub | Actually switches the lights, locks the door, reads the sensors | Home Assistant on the home network |
| Text-to-speech | Answers out loud | The home computer |
If you text the assistant instead of talking to it, you skip the voice pieces entirely, and no microphones need to listen.
What local AI still does less well
Being straight about the limits is the best way to decide.
- Smaller models make more mistakes. Home Assistant's own documentation says so: smaller models are more likely to make mistakes than larger ones, and may struggle to hold a conversation while controlling the house. It describes device control through local models as experimental and recommends exposing fewer than 25 entities for best results. A well-designed system respects that by carefully choosing what the AI can touch, and by requiring confirmation for locks, garage doors and alarms.
- It needs real hardware. A responsive household assistant needs a dedicated computer with a capable graphics processor and enough memory for the model. The more capable the model, the more memory it needs. A small single-board computer can run home automation well but won't run a large model quickly.
- It doesn't know today's news. A local model knows what it was trained on. For weather, traffic or scores, it needs an outside source, which reintroduces the internet for those requests only.
- Someone has to maintain it. Models, software and integrations need updates. Cloud providers do this invisibly; at home, you or your installer does.
The middle ground: hybrid setups
"Local" and "cloud" aren't all-or-nothing. A sensible design keeps the sensitive, high-frequency work local: understanding household requests, controlling devices, knowing who's home. Narrow lookups go outside only when needed, such as the weather forecast or a restaurant's opening hours. Home Assistant itself is built this way: it offers both local processing and a cloud option, and you choose per piece.
The test is simple. For any request, ask what leaves the house, and whether you'd be comfortable with that.
What "private" should mean: questions to ask
"Private" is easy to put on a box. Before you trust the word, ask:
- Where is my request processed? On hardware in my home, or on someone's servers?
- What is stored, where, and for how long? Can I see it and delete it?
- Who else can access it? The manufacturer, the installer, contractors, the model provider?
- Is my data used to train anything?
- What still works when the internet is down?
- How do messages reach it? If you text or chat with it, that message passes through the messaging service's network. iMessage and WhatsApp encrypt their messages, but they're still part of the chain. See our guide to texting your house.
- What about connected accounts? If the assistant reads your email or calendar, that data still lives with your email and calendar provider. A local assistant reading it doesn't move it out of the cloud.
- Who owns the hardware and accounts? If you stop paying, does it keep working?
A trustworthy answer to all eight is more meaningful than any marketing label.
Cost: up front vs forever
Cloud assistants look cheaper because the provider owns the expensive hardware. A smart speaker costs little, and Amazon's Alexa+ is $19.99 a month or free with Prime. A local assistant reverses that: you pay for a capable computer up front, then there's no per-request fee, no AI subscription and no change in terms you didn't agree to. We break down real numbers in what a private AI home assistant costs.
Which should you choose?
- Cloud is fine if you mostly want music, timers, general questions and simple device control, and you're comfortable with the provider's policies.
- Local makes sense if the assistant will know your family's routines, control locks and cameras, read personal accounts, or run a second home that needs to keep working through an internet outage.
- Hybrid is what most well-designed homes end up with: local for the house, outside lookups only when needed.
How Kotoh Home builds it
Our private AI home assistant runs on a dedicated computer in your home that you own, with Home Assistant as the hub for your devices. We choose what the AI can control, require confirmation for locks and doors, keep the system on its own secured network (see network and security), and tell you plainly which parts, like texting it from away, rely on outside services. If you're weighing this against a speaker in every room, our explainer on whether Alexa is always listening is a good place to start.
Common questions
Is a local AI assistant really private?
The AI processing is: your requests are understood and answered on hardware in your house rather than on a company's servers. But privacy depends on the whole chain. If you text it through a messaging app, the message travels through that app's network, and if it reads your email or calendar, that data still lives with your email provider.
Can a local AI control my smart home?
Yes. Home Assistant can connect its Assist voice assistant to a local AI model through its Ollama integration and let it control the devices you choose to expose. Home Assistant describes that control as experimental and recommends keeping the exposed devices to a manageable number, because smaller models make more mistakes.
What hardware do I need to run AI at home?
Small models run on modest computers, but a responsive household assistant generally needs a dedicated machine with a capable graphics processor and plenty of memory. The bigger the model you want to run, the more GPU memory you need.
Does a local assistant work without the internet?
The core does. Understanding requests, running automations and controlling devices inside the house all keep working. Anything that needs outside information, like weather, news or texting you while you're away, still needs a connection.
Is local AI cheaper than a cloud assistant?
It costs more up front because you buy hardware, but there's no per-request or per-month AI fee. Cloud assistants are often cheap or free to start because the provider runs the hardware, and you pay with data, subscriptions or both.