Is a private model less capable than ChatGPT?

For narrow, well-scoped tasks — summarizing your documents, answering from your knowledge base, classifying tickets — the gap is small enough that users do not notice. For broad general knowledge and complex reasoning, frontier public models still lead. Most business deployments only need the narrow part.

Does GDPR force us to use private AI?

No. GDPR requires you to control the processing, define the legal basis and manage subprocessors — all possible with public APIs under the right agreements. What GDPR does forbid is the careless habit: pasting customer data into consumer tools with unknown retention.

Can we start public and go private later?

Yes, and it is the standard path: prototype against a public API with synthetic data, prove the value, then re-architect the parts that touch sensitive data onto a private deployment. The mistake is the reverse — building everything private before knowing if anyone wants the product.

// ai

Private AI vs public models: when your data cannot leave the building

3 min read · all guides

"Can we just use ChatGPT for this?" is usually the first question in an AI project, and it is the right one — public models are absurdly capable and cheap. The follow-up question is the one that matters: what happens to the data? If transcripts contain patient records, contracts contain client names, or your product itself is the model's output, "the vendor's privacy policy" stops being a sufficient answer.

This guide separates the real constraints from the folklore: what actually leaves your building when you call a public API, when that is fine, when it is not, and what "private AI" costs in 2026.

What actually leaves the building

When you call a public LLM API, three things travel: the prompt, the conversation context, and — on most business tiers — nothing else, because the major providers stopped training on API data years ago. The realistic risk is not the model vendor training on your secrets; it is everything around the call: logs, retention windows, subprocessors, and the thirty other services your prompt passes through on the way. Most "data leak" incidents we investigate are not model breaches — they are a developer piping a production database into a prompt for a quick test.

When a public model is completely fine

Drafting marketing copy, summarizing public documents, generating boilerplate code, translating internal notes with no personal data — for these, a public API on a business tier is the right tool, full stop. The productivity gain is real and the exposure is close to zero. The discipline that makes this safe is boring: a written list of what may never go into a prompt (customer identifiers, credentials, health data, unpublished financials), plus a review before any new data source gets connected. Policy beats architecture for low-sensitivity work.

When you need a private model

Three situations genuinely justify it. Regulated data: health, legal and financial records where your own compliance team — not a vendor FAQ — must answer for every byte. Trade secrets: manufacturing processes, security research, unreleased products where even a retention policy is a concession. Contractual: enterprise clients who simply forbid third-party subprocessors, and no argument about API tiers will change their paper. In these cases "private" means the model runs in infrastructure you control — your cloud account or on-premise — with no external calls at inference time.

What private AI costs in 2026

The sticker price dropped hard. A capable open-weight model serving a few hundred users fits on a single GPU server or a dedicated cloud instance; expect infrastructure in the hundreds of euros per month, plus engineering to run it well. The honest total is not the GPU — it is the operations: keeping it patched, evaluated, backed up and actually good at your task, because nobody else will tune it for you. Our rule of thumb: below roughly fifty regular users, the fixed engineering cost usually outweighs any data-sensitivity savings, and a well-prompted public model with strict data hygiene wins on economics.

Frequently asked questions

Is a private model less capable than ChatGPT?: For narrow, well-scoped tasks — summarizing your documents, answering from your knowledge base, classifying tickets — the gap is small enough that users do not notice. For broad general knowledge and complex reasoning, frontier public models still lead. Most business deployments only need the narrow part.
Does GDPR force us to use private AI?: No. GDPR requires you to control the processing, define the legal basis and manage subprocessors — all possible with public APIs under the right agreements. What GDPR does forbid is the careless habit: pasting customer data into consumer tools with unknown retention.
Can we start public and go private later?: Yes, and it is the standard path: prototype against a public API with synthetic data, prove the value, then re-architect the parts that touch sensitive data onto a private deployment. The mistake is the reverse — building everything private before knowing if anyone wants the product.

// next steps

Working on something adjacent? Tell us about it — or read the services we deliver today.