Originally published in Spanish in El Derecho, Diario de Doctrina y Jurisprudencia, No. 16,260, Buenos Aires, September 4, 2026 (Cita Digital: ED-VI-CDIX-495).

I. Introduction

These are days of informational turbulence, and all that noise does little to help us take a calm, deep look at the reality we want to analyze.

There is plenty of talk and writing about «massive AI attacks,» «weaknesses in cybersecurity systems,» «the risk of generative AI hacking national defense systems,» «the dehumanization of the liberal professions,» «the dangers AI poses to the world of law,» and «the threat to confidentiality when information is handled with AI.» Yet little or none of it, from my personal standpoint, adds real value beyond describing specific situations or isolated events.

The incorporation of predictive and generative artificial intelligence tools into professional practice has stopped being a promise and become an everyday reality inside Argentine law firms.

Case law research, preliminary drafting of contracts and briefs, document translation, case file summaries, and risk analysis are all tasks where artificial intelligence (AI) now intervenes with growing intensity. And yet that adoption usually happens informally: no written internal rules, each lawyer using whatever tool is at hand, under the terms and conditions the platform provider imposes by default.

That scenario is incompatible with the obligations resting on the Argentine lawyer: professional secrecy (Section 156 of the Criminal Code and the ethics rules of the bar associations, such as Section 6(f) of Law 23,187 for the City of Buenos Aires), the personal data protection regime (Law 25,326 and the regulations issued by the Agency for Access to Public Information), civil liability for defective legal advice (Sections 1716 et seq. and 1768 of the Civil and Commercial Code), and the duties of diligence inherent to mandate and professional services agreements.

At this stage of the game, the answer for an organization such as a law firm should be the adoption of an internal, mandatory AI use protocol. My aim in this piece is to outline the essential contents that protocol should include.

II. Nature, Scope, and Principles of the Protocol

The AI use protocol of a law firm (let’s call it the «Protocol») should be an internal regulation of the firm, approved by its partners, that becomes part of the body of employment and contractual instructions applicable to lawyers, employees, interns, and outside collaborators.

It is advisable to serve the Protocol in a verifiable manner, have its receipt signed, and incorporate it by reference into employment agreements and professional services contracts, so that a violation amounts to breach of contract or serious labor misconduct, as the case may be.

As for scope, the Protocol should define which tools are permitted (a list of approved applications, with their versions and contracting arrangements), which are prohibited (particularly free mass-market versions, whose terms authorize the provider to use the data entered to train its models), and which uses are acceptable. The guiding principles should be, at a minimum, five:

  • effective human oversight of every output,
  • reinforced confidentiality,
  • minimization of the data fed to the AI,
  • transparency toward the client, and
  • traceability of every relevant use.

III. Protecting the Confidential Information Uploaded to the AI

This is the core chapter of the Protocol, because the gravest risk of using AI in a law firm is the leak of information protected by professional secrecy.

The starting rule must be categorical: it is forbidden to enter into any AI tool information that could identify a client, an opposing party, or a specific matter, unless the technical and contractual conditions that the Protocol itself establishes to secure confidentiality are met simultaneously.

Those conditions come in three orders.

First, contractual: only tools contracted under enterprise or API (Application Programming Interface) plans may be used, provided they guarantee, in writing, that the data entered will not be used to train models, that strict confidentiality applies, that a data processing agreement compliant with Law 25,326 is in place, and that the provider undertakes verifiable security obligations (encryption in transit and at rest, certifications such as ISO 27001 or SOC 2).

An API is a communications bridge that allows different programs to use AI models through shared rules and data. It is very commonly found in legal organizations that also run commercial or custom-built practice management software.

If the servers sit outside the country, the firm should verify that the international data transfer complies with Section 12 of Law 25,326 and the rules of Argentina’s Agency for Access to Public Information (AAIP), through adequate contractual clauses.

Second, technical: prior anonymization or pseudonymization of the data whenever the task allows it (replacing names with initials or generic labels), disabling training history when the tool offers that option, using corporate accounts with reinforced authentication, and an absolute ban on the use of personal accounts by the professionals and staff involved for handling firm matters.

Third, procedural: prior classification of information into levels (public, internal, confidential, secret), with the rule that information in the higher levels may only be processed with the express authorization of the partner responsible for the matter; and the creation and maintenance of an incident log requiring immediate reporting of any improper upload, in order to trigger the duties to notify the client and, where applicable, the supervisory authority.

IV. Ownership of the Information Uploaded to the AI and How to Secure It for the Firm

The Protocol must distinguish three categories of assets:

  • The first is client information (documents, data, case background): it belongs to the client, and the firm is merely its depositary and custodian, under a duty of secrecy.
  • The second is the firm’s own knowledge (contract models, standard briefs, systematized doctrine and case law databases, internal criteria): it constitutes confidential information and, to a large extent, a trade secret of the firm, protected by Trade Secrets Law 24,766, provided reasonable measures are taken to keep it reserved.
  • The third is the prompts, instructions, and consultation libraries developed by the firm, which form part of that same know-how.

To secure the firm’s ownership over its assets, the Protocol should provide for:

  • clauses in the contracts with AI providers expressly declaring that input and output belong to the user and that the provider acquires no rights over them beyond the minimum technical license needed to process the query;
  • assignment and confidentiality clauses with lawyers and employees, establishing that the prompts, templates, and developments created within the employment or services relationship belong to the firm (in line with Section 3 of Law 26,766 and Section 4 of Law 11,723 on collective works); and
  • systematic documentation of those measures, a requirement imposed by Law 24,766 itself in order to invoke trade secret protection.

V. Where, How, and When the Information Provided to the AI Should Be Stored

The Protocol should set a storage and retention policy.

As for the where: sensitive information must reside on the firm’s own servers or in the corporate cloud contracted by the firm, not on the AI provider’s platforms; zero or minimal retention of prompts and outputs should be agreed with the provider, along with certified deletion upon termination of the service.

As for the how: encryption, role-based access controls, an audit log (who queried what, when, and for which matter), and a copy, in the digital case file, of every AI output used as input for professional work, together with the final version reviewed by the responsible attorney.

As for the when: retention should be aligned with the statutes of limitations for professional liability and with the rules on preserving commercial books and supporting documentation (Section 328 of the Civil and Commercial Code). I suggest a minimum of ten years from the close of the matter, followed by secure destruction.

VI. Ownership of the Results: Does the Output Obtained with AI Belong to the Firm or to the Client?

The question calls for a nuanced answer, which the Protocol and, above all, the professional services agreement with the client should spell out in order to prevent disputes.

A sound criterion is the following: the final product delivered to the client (the legal opinion, the contract, the complaint) belongs to the client as far as its use and enjoyment are concerned, since it was the very object of the professional engagement and its price was included in the fees.

By contrast, the intermediate materials, methodologies, prompts, templates, and reusable argumentative structures remain part of the firm’s assets, as components of its know-how, and may be reused in other matters to the extent they do not reveal confidential client information.

One intellectual property caveat is in order: under the current state of Argentine law, Law 11,723 protects works featuring human creative contribution. Text generated entirely by an AI, without substantial input from the professional, will hardly qualify as a protected work. That reinforces the wisdom of settling the issue by contract rather than relying on any automatic authorship.

The Protocol should also require that the advice delivered always be the result of review, correction, and intellectual appropriation by at least one lawyer, which, at the same time, consolidates the document’s legal protection and the fulfillment of the duty of care.

VII. Training the Firm’s AI: Who Is Responsible and Who Owns It

When the firm moves toward customized tools (models fine-tuned with its own documentation, knowledge bases connected through retrieval-augmented techniques, internal agents), the question of «educating» or training the AI comes into play.

The Protocol should designate someone in charge: an AI Committee is recommended or, in smaller firms, a partner responsible for technology, with authority to approve the materials that feed the system, verify that they contain no unnecessary personal data or client information lacking a legal basis for that use, periodically validate the quality of the answers, and keep the system’s version log.

Regarding ownership, the rule must be twofold.

Inward: the training materials, knowledge bases, and model adjustments belong to the firm, as trade secret and as collective work, and this must be so agreed with the professionals who take part in building them.

Outward: the contract with the technology provider must stipulate that the fine-tuned model, its differential weights or parameters, the vector databases, and any improvement derived from the firm’s data are the exclusive property of the firm or, where that is not technically feasible, that the provider may neither use them for the benefit of third parties nor keep them once the relationship ends.

The outputs of the trained AI follow the regime described in the previous section: the final deliverable belongs to the client; the installed capacity that produces it, to the firm.

A delicate point the Protocol must expressly prohibit is training the system with client documents without their informed consent or without strong, secure anonymization, since the knowledge absorbed into the model could surface in answers given in other clients’ matters, in breach of professional secrecy and with potential conflicts of interest.

VIII. Other Necessary Contents

A complete Protocol should also address the following points:

  • Human oversight and responsibility: every AI-assisted product must be verified by a licensed attorney, who will answer professionally as if they had drafted it in its entirety. Case law or statutes suggested by the AI must always be checked against the official source, to avoid the well-known hallucinations.
  • Transparency with the client: the firm must define in which cases the use of AI is disclosed, adding a clause to the fee agreement or engagement letter, and what to do if the client forbids it.
  • Use before the courts: the firm must track and comply with the general laws and with the court rules and regulations issued on the disclosure of AI use in judicial filings.
  • Mandatory, periodic training for all personnel, with signed attendance records.
  • An internal disciplinary regime for noncompliance, graduated according to severity.
  • Review of the Protocol at least once a year, given the speed of technological and regulatory change.

IX. Conclusion

The AI Use Protocol is not a technology document; it is an instrument of professional governance: it translates the lawyer’s classic duties (secrecy, diligence, loyalty) into operating rules for a new tool.

Well drafted, it performs a threefold function: it protects the client, it shields the firm from liability, and it consolidates, as the firm’s own asset, the knowledge the firm builds while working with these technologies.

The law firms that adopt it early will not only reduce risk; they will turn the orderly use of AI into a competitive advantage and into an argument for trust before their clients.