Confidentiality Is Not a Feature: What In-House Teams Must Demand from Legal AI

The Duty Has Not Changed. The Risk Profile Has.

Legal professional privilege and the duty of confidentiality are not new concepts. What is new is the surface area of exposure. When a lawyer or in-house counsel feeds a contract, a board memo, or a dispute chronology into a generative AI tool, that information travels somewhere. It may be used to train a model. It may sit on a third-party server in a jurisdiction with different data protection standards. It may, in edge cases, surface in a response given to a completely unrelated user.

The Singapore Law Gazette recently revisited these questions with reference to tools like Claude, noting how thoroughly AI assistants have worked their way into legal practice. The underlying concern is one that any serious practitioner should already be sitting with: when you use a commercial large language model to assist with legal work, who else is in the room?

The answer shapes everything about how AI should be evaluated and deployed in legal settings.

What 'Reading From Your Side' Actually Means

Most general-purpose AI tools are built for breadth. They are trained on vast corpora, tuned for general helpfulness, and operated by vendors whose primary obligation runs to their own commercial interests, not to your client or your organisation.

Legal AI that is genuinely fit for purpose works differently. It reads contracts and correspondence from the perspective of the party using it. It understands that the question 'is this indemnity clause acceptable?' has a very different answer depending on whether you are the buyer, the seller, the lender, or the borrower. And it handles the information you give it with the same care a trusted adviser would.

This is not a marketing distinction. It is a structural one. When Adira processes a contract, the analysis is anchored to your position, your negotiating posture, and the governing law of the jurisdiction you are actually working in. The tool does not treat your commercial terms as training data for someone else's benefit.

In-house teams in particular need to press vendors on this point with specificity. Vague assurances about 'enterprise-grade security' are not a substitute for clear contractual commitments about data use, retention, and model training.

The Jurisdiction Problem That Most Tools Ignore

Confidentiality obligations in Singapore are shaped by the Legal Profession Act, the Personal Data Protection Act, and the professional conduct rules administered by the Law Society. They are not identical to the obligations that govern a London firm, a New York partner, or a Sydney GC. The nuances matter.

A generic AI tool calibrated on Anglo-American legal materials will not reliably flag that a particular disclosure obligation or privilege carve-out operates differently under Singapore law. It may not even know to look. The result is analysis that feels authoritative but is subtly wrong for the jurisdiction in which the advice will actually be relied upon.

This is why jurisdiction-awareness is not a nice-to-have. It is the difference between a tool that adds genuine value and one that creates quiet liability. Adira is built to know the law of the jurisdiction it is working in, which means it can apply the right confidentiality framework rather than defaulting to the nearest common-law approximation.

Practical Steps for Legal Teams Evaluating AI Today

If your firm or in-house function is currently using or considering any AI tool for contract work, the following questions should be non-negotiable before you proceed.

First, ask the vendor directly: is client or matter data used to train or fine-tune the underlying model? Get the answer in writing, in the contract, not in a sales presentation.

Second, understand where data is processed and stored. Cross-border data transfers carry regulatory risk that does not disappear because the tool is convenient.

Third, assess whether the tool has any concept of which party it is acting for. An AI that simply summarises a contract without understanding your position is not giving you legal analysis. It is giving you a synopsis.

Fourth, consider your own professional obligations. Lawyers in Singapore, like those in most common-law jurisdictions, remain personally responsible for the work product they deliver, regardless of the tools used to produce it. Delegation to an AI does not transfer the duty.

Confidentiality as a Design Principle

The broader lesson from the growing conversation around AI in legal practice is this: confidentiality should be a design principle, not a compliance checkbox applied after a product is already built.

Tools designed from the ground up for legal use, with data handling, jurisdiction specificity, and party-aware reasoning baked into the architecture, offer something that retrofitted general-purpose models cannot. They treat the protection of sensitive information as fundamental to what they do, rather than as a constraint imposed from outside.

In-house teams and law firms that treat confidentiality as a filter applied on top of AI, rather than as a property of the AI itself, are taking on risk that their clients and counterparties have not consented to. The question is not whether to use AI in legal work. The question is whether the AI you use was built with your obligations in mind.