Cuedis

AI on your own PC or in the cloud: privacy, cost and Singapore's PDPA

More and more business software comes with AI built in: tools that write summaries, answer questions about your data, or draft replies to customers. Sooner or later you will have to decide where that AI runs: on a computer in your own office, or on an online service such as OpenAI, Google Gemini or Anthropic Claude.

Both are reasonable choices. They differ in privacy, cost, speed and how much setting up they need. This post explains the trade-offs in plain terms, including what Singapore's Personal Data Protection Act (PDPA) means for the decision.

This is general information, not legal advice. If personal data is involved, check your own obligations with your data protection officer or a lawyer.

Option 1: AI on your own computer

A "local" AI model is a file of a few gigabytes that runs on an ordinary computer. Open models such as Qwen, Llama and Mistral can be used this way, and many are licensed for commercial use.

What it is good at

  • Privacy. Nothing leaves the computer. Your questions, your data and the answers all stay in your office.
  • No per-use cost. Once it is set up, it costs nothing extra however much you use it.
  • Works without internet, which matters for a shop PC with an unreliable connection.
  • Predictable. The model does not change unless you change it.

What to watch out for

  • Hardware. A useful model for business questions needs a reasonably modern computer. As a rough guide, about 16 GB of memory for the mid-sized models that write clear answers and decent SQL, and more if the same PC also runs your database. Apple silicon Macs and PCs with a recent graphics card are noticeably faster.
  • Quality. Small local models are good at wording answers from facts and writing straightforward queries. The largest online models are still better at complex reasoning.
  • Speed. On modest hardware, an answer can take several seconds or longer.

Option 2: An online AI service

Here, your software sends each question (and whatever information is needed to answer it) to a provider's servers, and gets the answer back.

What it is good at

  • No hardware worries. It works on any computer with an internet connection.
  • Quality. The large online models are excellent at reasoning, writing and complex queries.
  • Speed. Usually fast, regardless of how old your PC is.

What to watch out for

  • Data leaves your office. Whatever is sent with each question goes to the provider. Check what exactly your software sends.
  • Cost per use. Most services charge per amount of text processed. For occasional questions this is small; for heavy use it adds up. Set a spending limit in the provider's account.
  • Needs internet, and depends on the provider being available.
  • Provider terms. Read whether the provider keeps your data, for how long, and whether it is used to train their models. Business and API plans usually have stricter terms than free consumer apps.

Option 3: A server in your office

There is a middle way: run the AI model on one stronger computer in your office, and let other computers use it over your network. Tools like Ollama and LM Studio make this fairly easy. You get the privacy of a local model without needing a powerful PC at every counter.

What the PDPA says

If the data your AI sees includes personal data (customer names, phone numbers, member records, staff details), the PDPA applies. A few obligations are especially relevant:

  • Transfer Limitation Obligation. If personal data is transferred outside Singapore, you must make sure it gets a standard of protection comparable to the PDPA, for example through a contract with the recipient. Most large AI providers process data outside Singapore, so this matters if you send personal data to them.
  • You remain responsible. When an outside company processes personal data for you, as a data intermediary, your organisation is still accountable for that data as if you processed it yourself.
  • Protection Obligation. You must make reasonable security arrangements to protect personal data, which includes choosing and configuring tools sensibly.
  • Data breach notification. If a breach affects 500 or more people, or is likely to cause them significant harm, you must notify the PDPC no later than three calendar days after deciding the breach is notifiable.
  • Penalties. Financial penalties can go up to 10% of an organisation's annual turnover in Singapore if that turnover is above S$10 million, or S$1 million otherwise.

The simplest way to reduce PDPA risk is not to send personal data to the AI at all. For most business questions, such as which supplier is late, which category's margin is falling or what sold best last month, the AI does not need any names, phone numbers or IC numbers.

A practical way to decide

Your situation A sensible choice
Sensitive data, decent PC (16 GB or more) On your own computer
Several branches or counters, one strong PC available A server in your office
Older or small PCs, questions do not involve personal data An online service, with personal data switched off
You want the best possible answers and can control what is sent An online service, with a business or API account

Whatever you choose, ask your software vendor three questions: what exactly is sent to the AI with each question, can personal data be left out, and can I switch between local and online later?

How Cuedis handles this

Cuedis lets the admin choose, under Settings → AI model: models on the computer itself, an online service (OpenAI, Gemini, Claude, DeepSeek, OpenRouter or any OpenAI-compatible service), or a server in the office. The choice can be changed at any time without reinstalling.

With an online service, each question is sent with what Cuedis found to answer it, such as facts from its memory or the rows a query returned, and for new questions the names of tables and columns. Never the whole database. Personal details such as names, phone numbers, IC numbers and addresses are not read at all unless you switch them on for a particular business. So with that switch off, there is no personal data in what Cuedis sends. And the noticing and learning, the core of what Cuedis does, happens on your own computer whichever model you choose.

Local and online AI are both good options. The right one depends on your hardware, your data and how much you want to spend. What matters most is knowing exactly what leaves your office, and keeping personal data out of it unless you have a good reason.