AI & Automation

How to Choose an LLM for Business Use: Hosted API, Private Cloud or Self-Hosted

November 11, 2026 · 4 min read · Etzal Group Team

A year ago the question was whether to use a language model. Now it is which one, and the list changes every few months. A sensible way to choose an LLM for a company does not start from a leaderboard. It starts from three questions about your own situation: where is the data allowed to go, what does the task actually need, and who will keep the system running.

The three deployment options

Hosted API. You send requests to a provider's service and pay per use. You get the most capable models, no infrastructure to run, and new versions without effort. Your data leaves your environment for processing, under the provider's terms.

Private cloud deployment. The same or similar models, offered inside a major cloud platform under your own account and region. You gain control over location and network access, and often simpler procurement. The price is usually similar or slightly higher.

Self-hosted open model. You run an openly available model on your own servers or rented GPUs. Data stays where you put it, and there is no per-request fee. You take on hardware cost, performance tuning, security and upgrades, and the models are generally less capable than the best hosted ones at demanding tasks.

Question 1: where may the data go?

Start here, because it removes options quickly.

  • What kinds of data will pass through the model: public text, internal documents, personal data, health or financial records?
  • Do your contracts or regulators require data to stay in a particular country?
  • Does the provider keep request data, for how long, and is it used for training?
  • Can you sign a data processing agreement that your legal team accepts?

For most business data, a hosted API with clear retention terms and no training on your data is acceptable. For regulated data with strict residency rules, a private cloud region or self-hosting may be the only route. Decide this with your compliance function before comparing quality.

Question 2: what does the task need?

Models differ, and so do tasks. Describe the task concretely:

  • Reasoning depth. Multi-step analysis and agent work benefit from the strongest models. Classification and extraction usually do not.
  • Language. Test in the languages your customers use. Quality in Indonesian, including informal chat style, varies between models more than quality in English.
  • Input length. Long contracts or whole conversation histories need a large context window.
  • Tool use. Agents depend on reliable tool calling. This is a capability to test, not assume.
  • Speed. A live chat needs an answer in a few seconds. A nightly batch does not.
  • Images and documents. Reading scans, photos of receipts or charts requires a model with vision.

Question 3: who will operate it?

A hosted API needs almost no operations. A self-hosted model needs people who can manage GPU capacity, monitor latency, apply updates and respond when it fails at night. If you do not have that team, the apparent saving on usage fees disappears into salaries and downtime.

Comparing cost honestly

Compare cost per completed task, not price per token.

  • For hosted models, multiply your measured tokens per task by the price, and include retries.
  • For self-hosting, include hardware or rental running all day, engineering time, and the idle capacity you pay for when traffic is low.

Self-hosting tends to make financial sense at high, steady volume on a narrow task. At low or uneven volume, paying per use is cheaper.

Run your own test

Public benchmarks measure general ability. Your decision depends on your task. Take fifty to one hundred real examples with known correct results and run each candidate model on them. Record:

  • Accuracy against your expected results.
  • Cost per task.
  • Response time.
  • Behaviour on the cases it should refuse.

The model that wins on your set is the right one for that task, whatever the rankings say.

Avoid lock-in

Models and prices will change. Keep the ability to move:

  • Call models through a small internal layer, not directly from every part of the application.
  • Store prompts and evaluation sets in version control.
  • Avoid provider-specific features unless they give a clear benefit.
  • Use different models for different steps where it helps.

A typical outcome

Many companies settle on a mix: a strong hosted model for complex reasoning and customer-facing replies, a smaller cheaper model for classification and routing, and a private deployment only for the workflows that handle regulated data. That mix is the result of answering the three questions task by task.

Model selection with a documented test is part of our AI Consulting engagements, and our Enterprise Software team builds the abstraction layer that keeps you free to switch. Write to us with the task and the data constraints.

Summary

Decide where data may go, describe what the task needs, and be realistic about who will operate the system. Then test the candidates on your own examples and compare accuracy, cost per task and speed. Keep the design flexible enough to choose again next year.

Frequently asked questions

Is a self-hosted model safer than a hosted API?

It keeps data on infrastructure you control, which some policies require. It also makes you responsible for securing, patching and monitoring that infrastructure. For many companies a hosted API with a suitable data processing agreement and no training on customer data is the lower-risk choice.

Should we standardise on one model for everything?

Standardise the way you call models, not the model. Put a thin layer between your application and the provider so that each task can use the model that suits it and you can switch when prices or quality change.

How often should we review our choice of model?

Every six months, or when a provider releases a major version. Rerun your evaluation set on the candidates and compare quality, cost per task and speed. Without an evaluation set the review becomes a matter of opinion.

Author

Etzal Group Team

Etzal Group Team, helping businesses grow with technology. Contact us.

Related service

WhatsApp AI Chatbot

We build, train, and run a WhatsApp AI agent that answers customers 24/7, qualifies leads, and follows up.