# Your own model? For most companies, a standard model with your materials is enough.

Jaro Rais · 6 min read · 2026-10-26 · AI and language models
Source: https://itashu.co/en/know-how/your-own-ai-model

“Shouldn't we train our own AI on our data?” We hear this question often. It sounds logical: your own model that knows your field better than anyone. Two years ago, it often made sense. Today, standard models catch up before you finish your own. But there's always a big BUT.

![Stairs: brief, search, skills and an outlined last step for training](https://itashu.co/img/art-13-hero-1280.webp)

## What training your own model means

The technical term is **fine-tuning**. You take a finished model that learned from a huge amount of text from the internet and continue training it on your data. For example, on thousands of contracts or a history of customer requests. The result is a new model that should be better in a narrow area.

**For a sense of scale:** meaningful fine-tuning needs thousands of high-quality, checked examples from your own practice. Large projects involve millions of pages of text, as the story below shows. Without data like that, training has nothing to build on.

- **With a person:** You send an employee on a two-year specialist course. They come back as an expert, but in those two years the field has moved on.
- **With AI:** Training is that course. Expensive, slow, and the result goes stale when the next generation of standard models comes out.

## A story that turned around

The legal AI company Harvey had OpenAI train a custom model on US case law. They added the equivalent of 10 billion tokens of legal data. **Lawyers preferred its answers over the GPT-4 of the time 97% of the time**, and the model gave 83% more factual answers. A clear win.

But in May 2025, Harvey wrote that in less than a year, **seven standard models** had outperformed its originally benchmarked system on its own legal test, BigLaw Bench, three of them from vendors other than OpenAI. Today Harvey combines standard models from several vendors and improves what surrounds them: workflows, firm knowledge and collaboration with people.

![Two timelines: a custom model ahead in 2023 and standard models overtaking it by 2025](https://itashu.co/img/art-13-timeline-1280.webp)

## Why standard models are catching up

There are three reasons:

- **Models now read hundreds of pages at once.** Why bake documents permanently into a model when you can give them to it right with the question?
- **Models think at the moment of the question.** Answer quality depends more and more on how the model reasons now, not on what it learned months ago.
- Standard models get cheaper and better every few months. So a custom model keeps playing catch-up. By the time you finish it, the next version may have overtaken it.

## What works instead of training

A standard model can become a specialist without training. Layers you add step by step are enough:

- **A good brief and materials.** Clear instructions, the right data and the answer format in one package.
- **Search in your documents** (technically RAG). For every question, AI pulls up current materials. More in our [article on company knowledge](https://itashu.co/en/know-how/language-model-frozen-in-time).
- **Skills.** Written procedures for how things are done at your company. AI loads them when it needs them. For example, the exact procedure for preparing a quote.

- **With a person:** Instead of a two-year course, you give a capable colleague a valid handbook, access to the files and a written procedure. Within a week, they work like an expert.
- **With AI:** A brief, search and skills work the same way. And when a better model comes out, it takes over the handbook and procedures without new training.

> **In practice:** “An accounting office wants AI that answers clients exactly in the office's style and by its procedures. Instead of training, a collection of sample answers, current guides and a skill with the procedure for VAT questions is enough. When the law changes, they update one document. With a custom model, they'd have to train again.”

## Training costs more than it seems

Training itself is only part of the cost. **You have to collect and clean examples, verify results, make sure the model doesn't get worse at something, and maintain it** when conditions change. For a small or medium-sized company, it's almost always an unnecessary investment.

**A concrete example:** in 2023, Bloomberg trained its own financial model, BloombergGPT, with 50 billion parameters. Training alone took almost 1.3 million hours on Nvidia A100 GPUs. Most companies have neither the budget nor the team for that.

**And the options are shrinking.** Since May 2026, OpenAI has been gradually shutting down self-serve training of its models. New organizations can no longer start it, and from January 6, 2027, existing customers won't be able to create new training jobs either. According to its own documentation, Anthropic doesn't currently offer fine-tuning of Claude models through its API and asks interested companies to contact it directly. The vendors themselves are moving toward adapting standard models through briefs, materials and tools.

## Three cases where training makes sense

Training isn't dead. It has just narrowed to specific cases:

- **When fractions of a second matter.** For example, a voice agent that answers the phone. The most powerful models think too long, and a small trained model can be the better choice.
- **When you want to turn a large model into a small, cheap one.** The large model generates high-quality answers, and a small one is trained on them. This is called distillation.
- **When correctness can be checked automatically.** If there's a clearly correct answer, a model can be trained to reach it more often.

![Three cases where training makes sense: a fast answer on the phone, distillation and a verifiable result](https://itashu.co/img/art-13-cases-1280.webp)

Today you usually don't train the whole model either. You train a small add-on that attaches to a finished model (technically LoRA). According to the original 2021 study, that means training up to 10,000 times fewer parameters with three times less GPU memory. It's cheaper, but still a last resort.

| ORDER | WHAT TO TRY | WHEN |
| --- | --- | --- |
| 1 | Brief and materials | Always |
| 2 | Search in your documents | When AI needs current or company knowledge |
| 3 | Skills | When AI doesn't know how something is done at your company |
| 4 | Training | Only when the previous steps aren't enough and you can say why |

**Ask yourself:** What exactly should a custom model do better than a standard model with your documents and procedures? If you can't say it in one sentence, you don't need training yet.

When choosing, we always start from the cheapest layer. Most companies stop at the second or third step and get a specialist without paying for their own model.

---

Tell us what your AI needs to know. We'll show you the simplest way to get there. https://itashu.co/en/consultation