Questions to ask a private LLM consultant in Singapore

4 Nov 2026 · This note is general information, not legal advice.

We think most enterprises should run their own models. That's our bias, so here is how to test anyone who pitches you a private LLM, including us. Each question below says what a checkable answer looks like.

About the model

1. Where will the model run, and do we get the weights?

"Private" can mean your server, your cloud tenant, or a vendor's endpoint with a contract. Ask which, and ask for the model name and its licence. Open-weight licences vary. DeepSeek-R1's code and weights are MIT-licensed, and its card says the series "support commercial use, allow for any modifications and derivative works" (DeepSeek-R1 model card). Qwen3-8B is listed under Apache 2.0 (Qwen3-8B model card). A checkable answer names a model and a licence you can read.

2. Can it be fine-tuned on our own workflows, and who keeps the result?

With weights in hand you can adapt the model to your documents, formats and decisions. Methods such as LoRA adapt a model by fine-tuning "a small number of (extra) model parameters" instead of all of them (Hugging Face PEFT), which keeps the hardware bill sane. Ask who owns the adapter files and the training data afterwards. The answer should be: you, on your storage.

3. Can our staff read how the model reached an answer?

Hosted reasoning models often hide this. OpenAI's API docs say "we don't expose the raw reasoning tokens emitted by the model". You can ask for a summary instead (OpenAI, Reasoning models). An open-weight model such as Qwen3 in thinking mode will "generate think content wrapped in a <think>...</think> block, followed by the final response" (Qwen3-8B model card). You can log that block next to the answer and the documents it used. Treat it as a record to review, not as proof the answer is right.

4. Who could ever train on our data?

Read the hosted option fairly. OpenAI says it does not train on business data by default, unless you opt in. It also says API inputs and outputs may be retained for up to 30 days to provide the service and identify abuse, with zero data retention available for eligible use cases (OpenAI enterprise privacy). That is a policy you rely on. On your own hardware, there is no model vendor in the data path to make or change that promise. Ask the consultant to show you where prompts, outputs and logs are stored.

About the consultant

5. Will you handle our personal data, and how will you track it?

PDPC's 2024 advisory guidelines say third-party developers of bespoke AI systems "are data intermediaries who have obligations under the PDPA i.e., Protection and Retention Obligations". They call it good practice to use data mapping and labelling, and to "maintain a provenance record" of where training data came from and how it was transformed (PDPC Advisory Guidelines, paras 1.4 and 11.2). Ask to see a sample provenance record before you sign.

6. Will you help us write the PDPA notification?

The same guidelines encourage organisations to tell people what the AI feature does, what personal data it uses, why that data is relevant, and which features of the data are most likely to influence the result (para 9.5). They also encourage service providers to support clients in meeting their notification, consent and accountability obligations (para 11.4). A good consultant hands you that information, not just a model.

7. How will you test it before and after launch?

PDPC suggests technical tools such as AI Verify, and using explainability testing to identify the data features most likely to influence a decision (para 10.10). For generative AI, IMDA's Model AI Governance Framework, released on 30 May 2024, sets out 9 dimensions for a trusted environment (MGF for GenAI). Ask which tests will run on your use case, on what data, and who repeats them after an update.

8. When you leave, can our people explain a result?

"The AI said so" is not an answer you can give a customer, an auditor or a regulator. Ask what your staff will be able to show for any output: the source documents, the reasoning record, the model version and the prompt. Also ask who patches the server and the model, and what they need to know to do it.

Score the answers

Want these eight answers for your own use case? We'll map the workload, name the model and licence, and show you where every log lives. Start here →