6 min read
Every time you paste a client contract, a patient note or a sales spreadsheet into a chatbot, that text leaves your computer. For most tasks that is fine. For some it is not, and in 2026 you finally have a realistic alternative: local AI, meaning an AI model that runs entirely on your own laptop with nothing sent to the internet. This guide compares local and cloud AI head to head for a one-person business, so you can decide which work belongs where.
Why this choice suddenly matters
Two years ago, running AI on your own machine was a hobby for engineers with expensive graphics cards. Today, free desktop apps make it close to point and click, and mainstream companies are adding local options to their products. Perplexity, for example, was reported in early November to have added a fully local mode for its Computer agent on Mac, recommending a recent Apple silicon Mac with 36 GB of memory.
Meanwhile, clients are asking harder questions. If you have ever had to explain your tools to a client, our guide on how to tell clients you use AI covers the paperwork side. This article covers the technical side: where the data actually goes.
The two options in one sentence each
- Cloud AI (ChatGPT, Claude, Gemini, Copilot): the model runs on the company’s servers. You get the most capable models, web search, file tools and connectors, but your text is processed by a third party under their terms.
- Local AI (apps such as Ollama or LM Studio running open models): the model runs on your computer. Nothing leaves the machine, it works offline and it is free to use, but the models are smaller and the setup is yours to manage.
Head to head: six factors that decide it
1. Privacy and confidentiality
Winner: local AI. When the model runs on your laptop, there is no server log, no retention policy to read and no chance your text is used for training. Business plans from the big cloud providers do promise not to train on your data, and that is a real protection, but it is still a promise from a third party rather than a physical fact. If a contract says client data must not leave your systems, local is the only option that clearly complies.
2. Quality of answers
Winner: cloud AI, by a wide margin. The models you can run on a normal laptop are good at summarizing, rewriting, extracting key points and drafting routine emails. They are noticeably weaker at long reasoning, tricky math, nuanced writing and anything that needs current information. Expect “capable junior assistant”, not “senior expert”.
3. Cost
Winner: it depends on your hardware. Local apps and open models cost nothing per month. But a model worth using wants plenty of memory. As a rough guide, 16 GB of memory runs small models comfortably, while 32 GB or more opens up much better ones. If you already own a recent machine, local is effectively free. If you would need to buy a new laptop just for this, a $20 a month cloud plan pays for itself for years.
4. Features
Winner: cloud AI. Web search, image generation, voice mode, file creation, connectors to your email and calendar: these are where cloud assistants pull far ahead. Local apps give you a chat window and, in some cases, the ability to ask questions about documents on your drive. That is useful, but it is a narrow toolkit.
5. Reliability and offline use
Winner: local AI. A local model works on a plane, in a client’s basement office with no signal, and during a cloud outage. It also never changes overnight. Cloud models get updated, renamed or moved between plans (Google’s AI Plus subscribers just lost Pro models), and your carefully tuned prompts can behave differently the next morning.
6. Security risks you might not expect
Mixed. Local AI removes the third-party risk, but you now own the risk of downloading software and model files. Only install apps from their official sites and download well-known models through the app itself. Cloud assistants with connectors carry a different risk: a malicious email or web page can try to steer the AI into leaking data, a trick explained in our beginner’s guide to prompt injection.
The scorecard
- Privacy: local
- Answer quality: cloud
- Monthly cost: local, if you already have the hardware
- Features: cloud
- Offline and stability: local
- Security: different risks, neither is risk free
The verdict: run a two-lane system
For almost every solo business, the right answer is not one or the other. It is both, with a clear rule for which lane each task goes in.
Local lane (sensitive, routine):
- Summarizing client contracts, medical, legal or financial documents
- Pulling action items from confidential meeting notes
- Rewriting emails that mention names, account numbers or health details
- Drafting from anything covered by an NDA
Cloud lane (public, complex):
- Marketing copy, blog posts and social content
- Research that needs the web
- Strategy questions where reasoning quality matters
- Images, slides, spreadsheets and anything that needs connectors
A simple test decides the lane: would I be comfortable if this exact text appeared in a vendor’s log? If yes, use the cloud. If no, use local, or remove the identifying details first and then use the cloud.
Setting up the local lane in about 30 minutes
- Check your memory. On a Mac, open “About This Mac”. On Windows, open Settings, then System, then About. Under 16 GB, local AI will be slow; consider sticking to the cloud with careful redaction instead.
- Install one app from its official site. LM Studio has the friendliest interface for non-technical users. Ollama is lighter and popular if you are comfortable with a few typed commands.
- Download one mid-sized model that the app recommends for your memory. Start small; you can always try a bigger one.
- Test it on a real but low-stakes task, such as summarizing an old contract you already understand, and compare the result with what you know is true.
- Write your lane rule down in one line and stick it on your monitor.
Keep your expectations in check. A local model can take several seconds per answer and occasionally miss details. Because it runs on your own hardware, it also has a smaller working memory, a limit explained in our guide to why AI forgets the start of a long chat. Feed it one document at a time.
Where to go from here
If you handle confidential client material even once a week, set up the local lane this month. It costs nothing if your computer is recent, and it gives you a clean answer the next time a client asks where their data goes. If your work is mostly public marketing and research, stay in the cloud and spend that half hour on something else. Do you already keep sensitive work off cloud AI? Share your rule in the comments.
Related reading
- How to Tell Clients You Use AI: Write a Disclosure Clause and Website Policy in One Hour
- What Is Prompt Injection? A Beginner’s Guide to the AI Trick Hiding in Emails and Web Pages
- What Is a Context Window? A Beginner’s Guide to Why AI Forgets the Start of a Long Chat
- Your Year-End AI Review: A 60-Minute Audit of Every Tool You Pay For Before January Renewals
