Local AI for branches with unreliable internet

The question can be translated, the status has to be checked

The branch has a working internal network and a prepared computer with a model, and the internet has gone down. A customer shows a question: „Is my order ready for pickup?”. The employee types the text into the local app and gets it in their own language, in this example Polish: „Czy moje zamówienie jest gotowe do odbioru?”. They understand what the visit is about.

The order system runs in the cloud and cannot be reached. The employee writes, in Polish, that they have no access to the order system right now and so cannot confirm whether the order is ready. The assistant proposes the English wording: „I can't access the order system right now, so I can't confirm whether your order is ready”. The employee checks the text before showing it to the customer. The assistant does not guess the order status.

What local AI in a branch costs and when it pays back

The count covers the time spent searching, reading and compiling the material. What stays is the time to check the output against the source, 25% of today's work in the scenario. We measure both times in a trial on 20 to 100 cases the team has already handled. The scenario assumes 600 conversations a month across a network of 10 branches. With fewer conversations the payback gets longer, and saved phrases may be enough.

What must stay available in the branch

While the internal network works, the workstations connect to a shared computer in the branch, so access to the model depends on that connection. Ollama describes a local mode without cloud models or web search, and sharing a model on the local network. Several workstations working at once depend on memory and queue handling. The app must open at the start of a shift without internet, together with the sign-in and the saved phrases. The choice between a shared computer and an install at the workstation is covered in local AI on a workstation or a server.

When to start with a backup internet connection

If most of the work at the counter needs live data from head office, a local assistant solves a small part of the problem. Meraki describes failover to a backup connection. The behaviour depends on the operating mode and the version, and switching over can come with interruptions. If conversations during an outage come down to a few messages, approved translations at the counter or an offline translator are a simpler start. A model is worth testing when customers' questions vary.

Where to start

The work starts with one branch and chosen languages. The branch manager brings examples of conversations, and a person who knows both languages judges whether the proposals keep the meaning. We roll out local models as part of private LLMs, and we build the view for working with sources as an AI app. For the free process review, a description of one case and the places where the material sits is enough. Confidential files stay in the company, and you get a written plan within 2 working days.