In transaction work, discretion is not a comfort feature. A sale process that becomes visible too early loses negotiating position, unsettles staff and customers, and is sometimes irreparably damaged.
That is why introducing AI tools in this field runs differently than elsewhere. The question is rarely whether a model can solve the task. It is whether the data may go there.
The order decides
The usual route starts with capability: find the best model for the task, build a prototype, and settle the data protection questions afterwards. In an environment with confidentiality obligations that regularly produces a working prototype that is not allowed into service.
The reverse order works better. Data classes first: what is public, what is confidential, what carries a confidentiality obligation, what is subject to professional privilege? Only then the question of which processing is permissible for which class.
The result is rarely a prohibition. It is usually a split: some tasks run through a provider with a data processing agreement and European endpoints, a smaller set through a model on your own infrastructure, and some stays deliberately manual.
Replaceability as a design decision
That split has an architectural consequence: which model sits behind a feature should be a configuration decision, not an architectural one.
There are two reasons. The obvious one is that the market moves faster than projects run. The more important one is that the assessment of a data class can change, through a new client, a new legal reading, or a requirement from the other side. Wire the integration in tightly and you rebuild. Build it to be replaceable and you change a setting.
When professional privilege applies
As soon as a law firm is in the process, a layer is added that data protection alone does not cover. Professional confidentiality concerns not only where data sits but who may see it at all, and in whose name a result goes out.
We have built platforms where the separation between operating company and mandate-holding firm reaches into the architecture: separate tenants, separate access paths, and logging that evidences who saw which document and when.
What it costs in practice
Less than most expect when it is settled at the start. Considerably more when it arrives afterwards.
The extra work sits almost entirely in preparation: naming the data classes, designing the access model, building the logging in from the beginning. Those are a few days in the concept phase that save a rebuild after sign-off, and in case of doubt they are the difference between a system that goes live and one that ends as a prototype.
