13 min read

Can you have your own AI? What your company actually owns

Sébastien Nobour

Founder of DEVEDANOS

Solutions

“Your own AI.” The phrase turns up in plenty of YouTube videos: after dropshipping, influencers present selling AI integrations to small businesses as a line of work anyone can get into. The pitch they give the client fits in one sentence: an AI that knows your documents, your products, your customers, and that belongs entirely to you.

If someone has made you that offer, it raises the question: can you really have your own AI, meaning an AI you fully own?

As of 2026, training a model from scratch is out of reach for a small or mid-sized business. Fine-tuning an open model, one whose file you are handed, is possible. Most of the time, though, what you actually get under the name “your own AI” is software built around another company’s AI model.

That software can be excellent. But the model does not belong to you, and its developer decides its price, its terms and how long it lives. Depending on a supplier is nothing unusual. The risk is not knowing that you do.

Who can really build an AI?

A model like those behind ChatGPT or Claude is born from training on huge quantities of text. Researchers at Epoch AI, an institute that tracks how AI is evolving, have estimated what it costs: tens of millions of dollars in specialised chips, and tens of millions more in salaries. The training cost of the largest models has grown 2.4-fold every year since 2016. In June 2025, Epoch AI counted twelve organisations in the world that had trained models at the scale of GPT-4.

Even training presented as frugal stays out of reach for a small business. The Chinese company DeepSeek trained its DeepSeek-V3 model on 2,048 specialised chips and 14.8 trillion tokens of text. It reports a cost of 5.6 million dollars, and states that this figure does not include the research and trial runs that came before.

What is sold under the name “your own AI”

Four different things are sold with the same words, and what you own at the end changes each time.

Training a model from scratch. This is the case described above. Whoever pays for the training owns the model.

Fine-tuning an existing model. You specialise an already-trained model with your examples. In 2023, researchers showed this can be done on a single graphics card for a model with 65 billion parameters (the model’s internal settings). If the starting model is open, meaning its developer publishes the file, you end up with a file you can keep and use elsewhere. If it is closed, meaning the developer keeps the file to itself, the fine-tuning stays on the provider’s servers, and you receive nothing you can take with you.

Connecting your documents to a model, a technique known by the acronym RAG (retrieval-augmented generation). The model does not change: for each question, the software looks in your documents for the relevant passages and passes them to the model. You own your documents, their index (the table that lets the software find a passage quickly) and the code.

Writing instructions for a provider’s model. An interface and a block of instructions the user never sees (the prompt). You own the instructions and the interface.

The last three are legitimate integration work. Only the first, and the second on an open model whose file you are handed, leave you with a model that keeps working if the seller and the provider both disappear. The word “own” describes only those.

An offer that talks about “your own AI” without handing over a model file is selling you a fine-tune on a closed model, a connection to your documents, or a set of instructions. The phrase then promises ownership that the offer does not deliver.

Even fine-tuning a closed model depends on the provider’s calendar. At OpenAI, since 7 May 2026, an organisation that has never fine-tuned a model can no longer do so. From 6 January 2027, no customer will be able to start a new fine-tune. A model that has already been fine-tuned stays usable until the day OpenAI retires the model it started from.

What happens when someone else owns the model

A provider retires its models. Anthropic, the company behind Claude, commits to giving at least 60 days’ notice before retiring a public model; after that, requests to the model fail. On 14 April 2025, OpenAI announced the retirement of GPT-4.5 preview, which was gone on 14 July, three months later. Each time, the integration has to be rechecked with the replacement model.

Prices move, in both directions. In November 2024, Anthropic set the price of Claude 3.5 Haiku at four times that of the previous version. In June 2025, Google doubled the price of text sent to Gemini 2.5 Flash. On the other hand, a 2025 study shows that the price of a given level of performance falls 5 to 10 times a year.

So the problem is not only that prices go up: the price, the model and the terms are decided by someone else. And the service can stop: on 11 December 2024, according to OpenAI’s own write-up, all of its services were severely degraded or unavailable for more than four hours.

When the seller itself depends on a model

Some sellers of “your own AI” have no product other than software built on a provider’s model. Their whole business then rests on a decision they do not control. That risk becomes yours, since your software rests on their product.

This is what happened to Windsurf, the maker of a coding assistant, in June 2025: according to the company, Anthropic cut off almost all of its direct access to the Claude 3 models, with less than five days’ notice. Jared Kaplan, a co-founder of Anthropic, explained that it would be “odd” to sell Claude to OpenAI, which the press then reported was about to buy Windsurf. Neither company had confirmed it. A provider can also launch a product that competes with the seller’s: according to Sacra’s analysis, a firm that publishes research on technology companies, customers of the writing tools Jasper and Copy.ai left their 50-dollar-a-month plans for ChatGPT, free or at 20 dollars, after its launch in November 2022.

The providers’ terms of use point the same way. Under Anthropic’s contract, new pricing applies 30 days after it is published; under OpenAI’s, 14 days after. Both allow access to be suspended if the customer breaks the rules, and both forbid using the service to build a competing product.

So ask the seller three questions. Which model does its own product rest on? Has it tested a second provider? What would you be left with if its access were restricted overnight?

The same question applies to anyone thinking of launching their own “AI-native” start-up. A company whose entire product rests on someone else’s model depends, from day one, on that other company’s decisions.

Integrating AI with the dependency in plain sight

None of this is a reason to give up on AI. A dependency can be managed when it is known, written down and prepared for before the provider changes something.

In practice:

  • The model in use is named in writing: developer, name, version, hosting location.
  • The instructions, your documents, their index and the code are kept with you, not with the model’s developer.
  • A set of test questions, whose right answers are known, comes with the software. It is replayed each time the model changes, so you know, with figures, whether the new model does as well.
  • A second model, from another developer, takes the same test. The day the first one disappears, the fallback has already been tested.
  • Spending is tracked every month.

Open models and local AI: what it costs

Some developers release the file of their models. Mistral AI, a French company, has released Mistral Small 3.1 under the Apache 2.0 licence, which allows commercial use. According to Mistral, this model runs on a single RTX 4090 graphics card or on a Mac with 32 GB of memory. An AI running locally, without internet, on a workstation, is therefore possible.

To serve several users, you would rather rent a graphics card in a data centre. At Scaleway, a French host, a 24 GB L4 card costs €0.79 an hour, around €575 a month if it is left on around the clock. On top of that comes the time of someone who installs, updates and monitors it all.

A model that fits on one graphics card does not necessarily do as well as a large commercial model. The set of test questions checks that on your own cases.

Open does not mean free of conditions. The licence of Llama 3.1, Meta’s model, requires for instance that “Built with Llama” be displayed in the product.

Sovereign AI: what it changes for a small business

“Sovereign AI” is the term used when a country or a region seeks to keep control of AI models and of the chips needed to run them. Mistral AI presents Mistral Large 4, available since 6 October 2026, as “Forged in Europe. Built for AI sovereignty.”. Access to chips has already changed by political decision: on 15 January 2025 the United States adopted a rule limiting exports of AI chips to many countries, then rescinded it on 13 May 2025.

For a small business, the consequence is concrete: access to a model, or to the chips needed to run it, can change for reasons that have nothing to do with your company or with the quality of the model.

How to check what you are being sold

Asking the seller is not enough. A spoken answer proves nothing. What proves it is a document or a test.

Asking the AI itself who it is tells you nothing either. A block of instructions can make it give any name. And a study of 27 models showed that more than one in four gives a false identity when asked.

  • The model’s name in the contract. Developer, name, version, hosting location, and the obligation to warn you before changing it, in a signed annex.
  • The list of sub-processors. If the software passes personal data to the model (names, email addresses, customer files), the GDPR forbids the provider from using the model’s supplier without your written authorisation (Article 28). If the model’s supplier is on the list, you know in writing who the software entrusts your data to.
  • The invoices. In whose name are the accounts with the model’s supplier opened? An invoice from the supplier addressed to your company shows you which supplier is used, and what it costs. If the accounts are in the seller’s name, you will not receive any: that is already information.
  • The model file. If you are told a model was trained or fine-tuned for you, ask for the file and its licence, then have someone other than the seller run it. The licence names the starting model.
  • The cut connection. For an AI presented as local: on your own hardware, cut internet access and use the software. If it answers, the model is running on your side.
  • Getting your assets back, by trying it once. Have the seller spell out what you are handed at the end of the contract: instructions, set of test questions, index, code, model file if there is one. Then do the export once, before you need it. Only a successful export proves you can leave with your assets.

The European AI Act requires that people be told they are interacting with an AI (Article 50), not that the model be named. It does, however, require the developer of a general-purpose AI model to provide documentation of that model to the companies that integrate it (Article 53). Ask your provider for it: it covers one specific model, and so names it.

Your software, and the model of a provider you know

For a higher-education group, we integrated AI into its apprenticeship offers platform, reserved for its students. At registration, the group’s staff typed in each student’s profile by hand, from their CV. We connected the platform to an OpenAI model: the software reads the image of the CV and extracts the student’s profile, in a format validated with the group’s teams. This integration helped cut the staff’s manual entry time by 80%, and students no longer have to type everything in themselves.

At DEVEDANOS, this is how we integrate AI into the software we develop: model named in writing, instructions and test questions kept with you, a second model tested, spending tracked. We do not call that “your own AI”. It is your software, which uses a provider’s model, and you know which one.

For custom software development, we work to a fixed budget, invoiced at the end of each month for the work done. Your application is usable within a month, with a few essential features: your users start working with it straight away, and how they use it determines which features come next. The source code, infrastructure, deployment pipeline and documentation are your intellectual property from the first payment.

Further reading

Can an AI tool build the application itself? Our article on building an app with AI separates what the tool does on its own from what needs a professional. Are you budgeting for that application? Our article on budgeting an application explains why you should aim at a goal rather than a list of features.

Has someone offered you “your own AI”, and you want to know what you would really own? Shall we talk?

Sources

Read next