企業本地 AI 系統能處理甚麼、不能處理甚麼

A local ai system for business runs AI jobs on a model you host on hardware you control, and it fits routine language work: summaries, pulling fields out of free text, and sorting messages into queues. Hard, costly or ambiguous cases still belong with a larger cloud model or a person. Test the line on 50 of your own examples, and keep a person on anything a customer sees.

A local ai system for business is a set of AI jobs that run on a model you host, on hardware you control, instead of calling a cloud provider every time. The honest version is narrower than the pitch. A local model can do routine, repetitive language work well enough to be useful, and the hard cases still belong with a larger cloud model. This page sorts everyday service-business jobs into those two groups and shows how to test where your own line sits.

What "local" means here

A local model's files sit on a machine you own or rent, and requests are answered there. Ollama, a common way to run them, says that it does not see your prompts or data when you run it locally, and that by default it binds to 127.0.0.1, so nothing outside that machine can reach it until you change a setting.

That is the privacy argument, and it is real for the jobs you run locally. It does not cover jobs you send to a cloud model, and it does not make a badly set-up network safe. For the hardware and upkeep side, see our guide to self hosted AI for business. This page is about what to ask the model to do.

Jobs a local model handles well

Three kinds of work fit, because the input is text you already have, the output is short or structured, and a person or a rule checks the result.

Summaries. A night's worth of job notes, a long email thread, a day of call transcripts. Ollama's Llama 3.1 page lists long-form text summarization among its use cases, with a 128K-token context window on every size.

Extraction into fields. Pull a customer name, address, system type and urgency out of a voicemail transcript or a free-text note and put them in a form. Ollama can enforce a JSON schema on a model's reply so the output has the same fields every time. Its docs also say to validate the response against that same schema, and to use a low temperature for more predictable results. Treat the schema as a guardrail and the validation as the safety net.

Routing and sorting. Tag an inbound message as spam, new lead, existing customer, scheduling question or complaint, and send it to the right queue. The model's job is a label, not a decision, and a wrong label goes to a person.

A fourth category is fixed-pattern drafts: an appointment reminder, a "we received your request" note, a morning briefing in the same structure each day. The wording is set. The model fills the blanks. A person approves anything that goes to a customer until the pattern has proven itself.

Cortex runs this way at Detail Driven, where local models handle the overnight checks and the morning briefing on dedicated Linux hardware. It does not publish a performance number for that setup.

Jobs that still need a larger cloud model

Plan on sending these to a stronger model, or to a person:

• A messy customer email with three problems, a threat and a half-remembered invoice.
• Reasoning across many documents, such as comparing a contract, an estimate and a change order.
• Anything where a wrong answer costs money or makes a promise: pricing, refunds, warranty decisions.
• Unusual cases the model has not seen in your files.

Size is part of the reason. Ollama's Llama 3.1 comes in 8B at 4.9GB, 70B at 43GB and 405B at 243GB. The big one is a data-center model, not a desktop one, and the smaller versions are what most small businesses can host. We are not claiming a ranking between local and cloud models, because it changes every few months. Test it on your own files.

How to find your line in a week

1. Collect 50 real examples of one job: 50 job notes to summarize, 50 messages to sort.
2. Run them through a local model, with the schema or wording you plan to use.
3. Have the person who owns the work grade each result as fine, fixable or wrong.
4. Run the same 50 through a cloud model and grade those too.
5. Decide the line. If the local model is fine on nearly every case and the misses are caught by validation or review, run that job locally. If the misses are costly, route those cases to the cloud model or to a person.

A grading sheet beats an opinion. It also gives you something to rerun when a new model arrives.

How a hybrid system routes work

A practical setup uses three layers in order:

1. Rules first. If a plain rule doe