Why an AI tool is a different case
A word processor or a case management system works on the firm’s own machines, and the firm knows where its data sits. With an AI tool, the whole point of the processing is that the content of the document reaches a model — and the decisive question is where that model runs and who can reach what it was given.
At a law firm the input is not arbitrary business data either. The contract received from the client, the correspondence around it and the internal note are material covered by professional secrecy, and handling it is a matter of professional obligation rather than IT preference. So it is not enough to look at how accurate an AI tool is: the route the file travels through it has to be established as well.
What happens when a contract goes into a public AI service
This is worth thinking through without indignation: the point is not that public AI services are bad products. The point is that they were not built for this use, and that shows up at three concrete places.
Where processing happens
The document is processed on an external provider’s infrastructure, typically outside the EU. The user does not choose which country that is, and usually cannot see it either. From that moment the file is no longer held by the firm.
Retention
How long the material is kept, and what further use is made of it, depends on the provider’s terms of the day, which they may amend unilaterally. The firm cannot reconstruct after the fact how long the material persisted or what purpose it served. That is uncomfortable even when no incident occurs at all.
Loss of control
The biggest practical problem is not some form of misuse but the absence of provability. If a client asks where their file was and who could reach it, the firm has to be able to answer precisely. With a public service that answer is not available — and its absence is on its own a sufficient reason for the route not to be on the table.
What private AI means in practice
Private AI is not a marketing adjective but an operating arrangement. It means the language models run on dedicated EU hardware, or entirely on the firm’s own server, and that the system calls no external AI API. The uploaded document stays in the same environment that stores it: there is no point in the process at which the file is handed to a different provider.
Three things belong with that, without which the claim would stay unverifiable. A named data processing agreement recording the place and purpose of processing. An audit trail from which it can be established what happened to a given document. And data export, because the firm has to be able to take its own material out at any time — without an exit there is no real control.
The own-hardware edition is the stricter version of this: there the entire system, models included, runs on the firm’s server, so privileged material does not even physically leave the building. Not every firm needs that, but for some it is the only acceptable answer.
Five questions worth putting to any vendor
- Where does the model run? A concrete answer is required: which country, whose infrastructure, and whether an external AI API sits anywhere in the chain.
- Who can read the file? Not only as against another customer, but as against the provider’s own staff.
- Is there a named data processing agreement? General terms of use are not a substitute for one.
- Is there an audit trail? Can it be established what happened to a given document, and how long is that record kept.
- Can the data be exported and deleted? Getting the firm’s own material out should not depend on the vendor’s goodwill.
If the answer to any of these is evasive, that is information in itself. It does not necessarily indicate bad intent — often it simply means the product was not designed for use of this sensitivity.
The two arrangements side by side
| Aspect | Public AI service | Private AI |
|---|---|---|
| Where the model runs | On an external provider’s infrastructure, typically outside the EU, whose location the user does not choose. | On dedicated EU hardware — or entirely on the firm’s own server. No external AI API is in the chain. |
| Where the file goes | It leaves the firm on upload. Retention period and any further use depend on the provider’s terms of the day. | It does not leave the system. The file is processed by the same environment that stores it. |
| Control | The user cannot see who accessed the file or when, and generally cannot demonstrate what happened to it. | An audit trail shows what happened to the document. Data can be exported and deleted at the firm’s decision. |
| Data processing agreement | General terms, unilaterally amendable, which the firm does not shape. | A named data processing agreement with the firm, recording the place and purpose of processing. |
The table compares the general characteristics of the two operating arrangements; the terms of individual services differ from these and change over time. It is worth asking the specific vendor for the answers in writing.
How this connects to AI Szerződéselemző
This is why the system was built as private AI: it runs on private EU hardware with no external AI API, or entirely on the firm’s own server. A data processing agreement, an audit trail and data export are part of the package rather than an add-on. Responsibility for the professional output — the clause extraction, the risk flagging and the drafting suggestions — still rests with the acting attorney: the system prepares, the lawyer reviews and finalises.
Confidentiality & GDPR — the detail →
This article is general information, not legal advice. Assessing the obligations that apply to handling privileged material is the firm’s own responsibility and judgement.