Does Your Business Need Its Own AI Model, or Is Claude or ChatGPT Enough?

The question comes up in almost every boardroom conversation about AI now, usually phrased as a worry: should we be building our own? Someone has read about a company training a model, or a client has asked where their data goes, and the room starts to wonder whether using Claude or ChatGPT is the shortcut that will cost them later.

For most businesses, the honest answer is no. A hosted model on the right plan is enough, and it is the most capable option available. But there are real cases where it is not, and it is worth knowing which side of the line you are on.

The three options, plainly

A hosted model. Claude, ChatGPT and similar, used through a business or enterprise plan or through the provider's API. The provider runs the model; you send it work. The most capable models available, with nothing to maintain.

An open-weight model. A model whose weights are published, so you can run it on hardware you choose: your own, a New Zealand data centre, or a cloud region you pick. You control where the data goes. You also take on running it, or pay someone to.

A custom or fine-tuned model. A model, often a smaller open-weight one, trained further on your own examples so it does one task in your way. Powerful for the right task, and the most work to build and keep current.

When a hosted model is enough

Start here. A hosted model is the right answer when:

• The work is varied: drafting, analysis, research, the everyday jobs across a team.
• Your data can be processed offshore under the provider's business terms, and nothing in your contracts or regulations says otherwise.
• Volume is modest or uneven, so paying per use beats paying for hardware that sits idle.

Most of the businesses we work with sit here, and the bigger gain for them is getting the team using it well, not changing the model.

When an open-weight model earns its place

The data must stay under your control or in a named country. Some health, legal, government and financial work carries rules or client contracts about where data is processed. In New Zealand, the Privacy Act 2020 sets conditions on sending personal information overseas, and sector codes such as the Health Information Privacy Code add their own. When the rule is real, an open-weight model on infrastructure in the right place is how you meet it.

The volume is high and steady. When the same kind of request runs thousands of times a day, every day, owning the compute can cost less than paying per request. The sums depend on your volume and the model size, so model them before you buy anything.

The work cannot depend on an outside provider. If an outage, a price change or a policy change at a provider would stop something critical, running your own model removes that dependency.

When a custom model is worth it

Rarely first. Fine-tuning suits a narrow task that runs at high volume, where you have plenty of good examples of the right output and a smaller tuned model can match a larger general one. Before fine-tuning, try two cheaper things: better instructions, and giving the model your own documents to read at the time it answers. Most of the time those close the gap.

How to decide without guessing

Take twenty real examples of the job, with the answers you would accept. Run them through each candidate: a hosted model, and an open-weight model if your constraints point that way. Compare the outputs side by side, on quality and on what each would cost to run at your volume. That afternoon of testing is worth more than any benchmark table, because it is your work.

If you want a structured way to think it through first, our free model choice skill has Claude ask about your job, data and volume and write a one-page recommendation. And Open-weight models is where we publish what we learn from running them.