Own AI or rented AI: when does it pay off?

The short answer

Own AI or rented AI: when does it pay off?

Whether own AI pays off depends less on the price of the model than on the time people spend checking and correcting. That time almost always decides what a good result costs.

By own AI I mean a model you run yourself, on your own servers or with a European provider. Sometimes you fine-tune it on your own examples. Rented AI is a model as a service, such as ChatGPT, Copilot or Claude, paid per subscription or per use.

The calculation

What does a good result really cost?

A good result costs the sum of model costs, checking, correcting, missed errors, building and upkeep, divided by the number of good results.

At GainWorks we calculate per task with that sum. The price of the model is usually the smallest item in it; checking by a person weighs more.

Our cost example at GainWorks is invented, but it shows how it works out. An installation company with 60 people receives 1,500 customer emails a month, and AI sorts them by type, urgency and who picks them up. With a rented model, good instructions and examples, a correctly sorted email costs about 88 cents. With an own, fine-tuned model on a European server it is 1.95 euros.

The own model makes fewer errors, but the server, the building and the upkeep weigh more. In our example at GainWorks the own model only becomes cheaper from around 20,000 to 50,000 emails a month.

The thresholds

When does own AI pay off?

Own AI pays off when three things come together: large volumes of one fixed task, a rented model that misses the bar after an honest try, and errors with a fixed pattern.

A fixed pattern is, for example, the same mix-up every time, the wrong tone or the wrong format. You can teach a model that.

Missing or changing knowledge is not something you solve with fine-tuning: you supply it from your own documents. Ovadia and colleagues compared both routes in a 2024 study, and supplying documents consistently did better than fine-tuning.

A hard requirement can turn the decision around. May your data not leave the building, must the answer come within a fraction of a second, or must the system work offline? Then you compare an own model with not doing the task at all, and you price that honestly.

My reading

What did I see in San Francisco?

In San Francisco I saw that an own, fine-tuned model has become far more attainable than I thought. For larger organizations it really starts to pay off at a certain volume of one task.

In early October Coinbase described how it fine-tuned a small open model of 9 billion parameters on fraud cases. On its own fraud test that model scored 9.6 percentage points higher than the large rented model Coinbase started with, and it answered faster.

According to Coinbase the training run cost less than 100 dollars in compute. Preparing the data, building, testing and upkeep are not in that figure.

Two developments weigh against each other. According to Epoch AI, the best open models have trailed the best closed models by only four months on average since January 2026. At the same time rented AI is getting cheaper fast: Epoch AI calculated that the price for the same level has fallen about thirteen times a year since 2023.

I expect fine-tuning to become a large part of what we see next year. Meanwhile OpenAI is stopping new training jobs on its own platform, so anyone who wants to fine-tune looks to open models sooner.

In practice

What does this mean for your company?

Start with one task instead of with the choice of a model: how often does it come up, what does it cost now, and when is a result good enough?

• Count the volumes: how often a month does the task come up, and how many hours does it cost now?
• Let the owner set the bar: whoever owns the task decides in advance when a result is good enough, not the vendor.
• Build the rented route properly first: instructions, your own documents, examples and a route to a person when in doubt.
• Calculate per good result: with checking, correcting, upkeep and the errors that slip through.

If the rented route meets the bar, you are done. If not, a fair comparison with an own model is the next step, and when in doubt we at GainWorks choose the simplest route.

In the article on the five levels of AI you can read where that choice belongs: the task is usually at level 2 or 3, the choice of a model is a level 4 question. Where AI pays off most in your company is what we find out with the BORS.