Over the past few months, I have been looking more closely at the growing number of artificial intelligence applications designed specifically for legal work. Spellbook was one of the products that first caught my attention because it operates directly within Microsoft Word and supports contract drafting, review and redlining. Once I started looking beyond Spellbook, however, I realised how quickly this market is expanding.

Harvey offers capabilities across legal research, drafting, document analysis and professional workflows. Thomson Reuters has developed CoCounsel Legal around research, analysis and drafting using recognised legal content such as Westlaw and Practical Law. LexisNexis offers Lexis+ with Protégé, while vLex has developed Vincent as an AI-supported legal research and analysis platform connected to a large international legal database. Luminance and Ironclad focus heavily on contracts, due diligence and contract lifecycle management, while Relativity aiR applies generative AI to litigation, investigations, document review and privilege-related workflows. Microsoft is also developing a Legal Agent within Word to support contract review and redlining, although, at the time of writing, it remains a preview capability within the Microsoft 365 Copilot Frontier programme rather than a mature, generally available legal platform.

It is not difficult to understand why legal departments and law firms are interested in these tools. Legal work involves enormous volumes of text, repeated comparison of clauses, review of precedents, examination of authorities and preparation of documents that often follow recognisable structures. A well-designed AI system can accelerate many of these activities and allow experienced professionals to concentrate more time on judgment, negotiation and strategy.

My position is not that legal AI should be resisted. I am a strong supporter of the responsible use of AI in business, and I believe these systems will increasingly become part of normal legal and commercial operations. My concern is different. I look at this subject as a leader who wants to introduce AI safely, obtain the operational benefits and, at the same time, protect the organisation from liabilities that could have been anticipated and controlled.

This article therefore reflects my interpretation of the governance and legal issues that leaders should consider before implementing legal AI. It is not intended as legal advice, and the exact legal position will always depend on the jurisdiction, the nature of the organisation, the contractual arrangements and the way the technology is actually used. My purpose is to examine where I believe the main risks arise and what an organisation should be able to demonstrate if its use of AI is later challenged.

The distinction is important because there is a natural tendency to assume that a product designed specifically for lawyers must also be safe for lawyers. A specialist platform may certainly offer stronger controls, more appropriate data sources and better professional workflows than a public chatbot. It may connect to recognised legal databases, provide citations or operate under enterprise contractual terms. Those features are important, but they do not transfer the organisation’s responsibilities to the software provider.

The provider does not assume the lawyer’s professional duties simply because its system contributed to the work. It does not decide whether the user was authorised to upload a document, whether the information could lawfully be processed in another jurisdiction or whether the final contractual position was commercially acceptable. Those decisions remain with the organisation and the people acting on its behalf.

This is where I see the real governance gap.

The risk begins before the AI produces an answer

Most discussions about legal AI begin with hallucinations. We have all seen examples of systems inventing cases, producing inaccurate quotations or referring to authorities that do not exist. Those failures are serious, especially when the output is filed in court or relied upon in formal legal advice, but the legal exposure may begin much earlier, before the system has generated a single sentence.

Consider what happens when an employee uploads a contract containing pricing information, personal data, negotiation history, technical details and confidential correspondence. Before the AI responds, several questions already exist. Was that employee authorised to place the document into the system? Was the application approved for that category of information? Where will the data be processed? Will it be retained? Can the provider or one of its subprocessors access it? Is any part of the information used to improve the product or the underlying model? Can it be deleted completely? Does the processing involve a cross-border transfer, and is that transfer permitted under the applicable law and the organisation’s contractual obligations?

The AI may eventually produce an entirely accurate answer, but that accuracy does not correct an improper disclosure or an unlawful processing decision.

This is particularly relevant in legal work because the information may be protected not only by contractual confidentiality but also by professional duties and legal professional privilege. I would be cautious about making any universal statement that the use of an external AI platform automatically preserves or automatically destroys privilege. Privilege rules vary substantially between jurisdictions, and the outcome may depend on the circumstances of the disclosure, the purpose for which it was made, the contractual safeguards and the precautions adopted by the organisation.

My interpretation is that the issue should be examined before privileged material is uploaded, rather than after a dispute has arisen. An organisation should be able to explain why the processing was necessary, which safeguards were in place, who could access the information and how it was protected. If it cannot answer those questions, it may struggle to demonstrate that it exercised appropriate care.

The American Bar Association addressed many of these issues in Formal Opinion 512. That opinion is based on the US professional-conduct framework, so it should not be treated as universal law or as automatically binding in every US jurisdiction. Nevertheless, the underlying principles are highly relevant. Lawyers remain responsible for competence, confidentiality, communication, supervision, candour and the reasonableness of fees when using generative AI.

The same pattern appears in the guidance and judgments issued in England and Wales. The judiciary has warned that public AI tools may produce inaccurate or fabricated information and has advised against entering private material into publicly accessible systems. In the 2025 Ayinde judgment, the court reinforced that legal professionals remain responsible for verifying authorities and checking the material placed before the court.

For me, the important point is not that public and specialist tools should be treated as identical. They should not. A contracted enterprise platform may have controls that a public chatbot does not have. However, the description “enterprise,” “legal-grade” or “secure” should not replace proper due diligence. These are provider descriptions. They may be supported by contractual commitments, technical controls and certifications, but the organisation still needs to examine the actual arrangements.

A marketing statement is not a governance control.

Professional responsibility remains human

The second major area of risk concerns how the output is used.

Legal AI can generate language that is clear, confident and professionally structured. That is one of the reasons it is useful. It is also one of the reasons people may rely on it too easily. A poorly written answer naturally attracts scrutiny. A polished answer may pass through an organisation without receiving the same level of challenge.

Specialist legal products can reduce some of the weaknesses associated with general-purpose systems by grounding outputs in legal databases, company documents or recognised sources. They do not eliminate the need for verification. A Stanford and Yale study of legal research products, based on versions tested in 2024 and later published in the Journal of Empirical Legal Studies, found that specialist systems could still produce unsupported or inaccurate answers.

I would not use the percentages from that study as an indication of the current performance of those products. The systems, models and retrieval methods have continued to develop. The more relevant conclusion is that access to authoritative sources can reduce error without removing it completely.

This is also reflected in recent court decisions. In June 2026, the US Court of Appeals for the Ninth Circuit issued its decision in LNU v. Blanche, involving nonexistent cases, misattributed quotations and inaccurate representations of genuine authorities. The court did not treat the mere use of AI as the misconduct. The problem arose because the lawyers signed and submitted material without properly reading and verifying it and then failed to deal candidly with the source of the errors.

This distinction matters. AI does not remove existing duties of competence, diligence and supervision. What it changes is the speed, volume and apparent credibility with which an error can be produced.

This is why I do not believe that the phrase “human in the loop” is, by itself, a sufficient governance control. It sounds reassuring, but it does not explain who the human is, what qualifications that person has or what they are expected to verify.

A meaningful review process should define the reviewer, the scope of the review and the authoritative sources against which the output must be checked. It should establish whether every citation must be opened, whether contractual summaries must be compared with the original document and who has final authority to approve the work.

A person who simply accepts the AI output is not exercising professional oversight. They are adding a human approval step to an automated process.

The required level of review should also reflect the potential consequence. Using AI to improve grammar in an internal communication is not equivalent to asking it to interpret a termination clause, prepare litigation submissions or advise on an employee’s legal rights. Governance should not treat these use cases as though they present the same level of exposure.

Data protection and contractual duties still apply

Legal AI may be new, but many of the liabilities associated with it arise from familiar legal obligations.

Legal documents frequently contain personal data. Employment disputes, investigations, claims, due diligence files, witness evidence and commercial agreements may all contain information about identifiable individuals. When these documents are processed through an AI application, the ordinary requirements of data-protection law remain relevant. The organisation may still need to establish a lawful basis, limit the information processed, define retention periods, implement appropriate security and understand the role of processors and subprocessors.

For organisations operating in the UAE, Federal Decree-Law №45 of 2021 concerning the Protection of Personal Data provides the federal framework, although its scope and exclusions must be considered carefully. DIFC and ADGM operate separate data-protection regimes. DIFC Regulation 10 is particularly relevant because it addresses personal data processed through autonomous and semi-autonomous systems.

The practical lesson, in my view, is that an organisation should not accept a general statement that information is stored “in the cloud.” It should understand where the relevant processing takes place, which legal regime applies, who has access to the data and whether international transfers are being made.

The use of a processor does not remove the accountability of the organisation that selected and deployed the system.

Client contracts and engagement terms may create even stricter conditions. A law firm or consultancy may approve an AI platform after a thorough technical and privacy review and still be prohibited from using that platform for a specific client. Outside-counsel guidelines, confidentiality agreements, procurement terms and engagement letters may restrict the use of generative AI, require prior consent or limit the use of subprocessors and international data transfers.

This means that approval of the tool is only one part of the process. The organisation may also need to assess whether it is approved for a particular client, matter or category of information.

A system may therefore be generally approved but prohibited for one transaction.

Unless those restrictions are visible to the people doing the work, a general corporate policy will not prevent a contractual breach.

Similar caution is needed with intellectual property and licensed content. Legal teams work with precedents, templates, subscription databases, expert reports and documents that may belong to clients or counterparties. Access to that content does not automatically mean that it may be used in any AI platform for any purpose.

The actual legal position will depend on licence terms, ownership, copyright law and contractual restrictions. My governance concern is that organisations should identify these issues rather than assume that the right to read a document includes the right to process it through AI.

Record retention creates another area where there is no single universal answer. Prompts, uploaded documents, generated drafts and user corrections may later become relevant in litigation, an investigation or a professional-negligence claim. Retaining everything indefinitely can increase privacy and discovery exposure, while deleting everything may make it impossible to reconstruct how an important decision was reached.

The appropriate approach will depend on the purpose of the system, the sensitivity of the information and the organisation’s broader retention and legal-hold obligations. What matters is that the decision is deliberate and documented.

A small error can become an organisational error

One of the characteristics that distinguishes AI risk from ordinary human error is scale.

A lawyer may misunderstand one agreement and affect one transaction. An AI system connected to a contract repository, negotiation playbook or automated workflow may apply the same misunderstanding repeatedly. It could recommend an incorrect liability position across hundreds of agreements, extract renewal dates incorrectly throughout a contract portfolio or fail to identify privileged material during a large disclosure exercise.

Consistency is one of the benefits of automation. It can also become a liability when the underlying instruction or model behaviour is wrong.

The same concern arises when legal AI is made available outside the legal department. Procurement, sales, human resources and operational teams may be given tools that can answer questions about contracts and legal obligations. This can reduce routine pressure on legal teams and improve access to information, but it can also create an informal legal-advice channel operating without sufficient supervision.

There is a significant difference between asking a system to locate a termination clause and asking it whether the organisation can terminate immediately without liability. The first is mainly a retrieval task. The second requires legal interpretation, factual context and professional judgment.

I am not suggesting that legal AI should be restricted to lawyers in every circumstance. I am suggesting that organisations should define where information retrieval ends and legal interpretation begins, who is permitted to cross that boundary and when escalation to the legal function is required.

Governance cannot be reduced to a short AI policy

Many organisations respond to AI risk by publishing an acceptable-use policy. This is a useful starting point, but it does not amount to a complete governance system.

A general statement telling employees not to upload confidential information into public tools will not explain which legal applications are approved, what information may be processed, which users may access them or what level of review is required before output is relied upon.

A workable legal-AI governance framework should begin with the intended use of each system. Spellbook used to suggest contract language presents a different risk from Relativity aiR being used during discovery, even though both operate in the legal domain. Each application has different users, data, consequences and verification requirements.

The organisation should identify the approved platforms and configurations, the categories of data that may be processed and the activities that remain prohibited. It should define system ownership, access rights, training requirements and the review standard applicable to different categories of work.

Vendor due diligence should extend beyond a review of cybersecurity certification. It should consider data use, retention, deletion, subprocessors, data location, model improvement, auditability, incident notification and what happens to the information when the contract ends. It should also address product changes. A system approved for document drafting may later introduce integrations, agents or autonomous actions that materially alter its risk profile.

The governance framework should also address client restrictions, records management, incident escalation and periodic testing. If a system repeatedly produces an incorrect interpretation or fails to identify a particular type of clause, the organisation should be capable of detecting the pattern before it spreads.

Training should reflect the role of the user and the risk of the use case. A lawyer verifying legal authorities requires different training from a procurement manager using AI to locate contractual obligations. Within the scope of the EU AI Act, Article 4 also requires providers and deployers to take measures to support an appropriate level of AI literacy among relevant staff and others operating AI systems on their behalf.

It is equally important not to overstate the classification of legal AI under the EU AI Act. The Act does not classify every product used by a lawyer as a high-risk system. Annex III identifies certain systems used by, or on behalf of, judicial authorities to assist with researching and interpreting facts and law and applying the law to specific facts. Comparable use in alternative dispute resolution may also fall within scope, while purely ancillary administrative activities are distinguished from those functions.

My interpretation is that a commercial contract assistant used by an in-house legal team would not fall into that Annex III category merely because it performs legal work. Its classification would depend on its intended purpose, functionality and actual deployment. Other obligations, including data protection, contractual duties, professional responsibilities and AI literacy, may still apply.

The question is not simply whether the tool works

When leaders consider a legal-AI platform, the first discussion will naturally focus on capability. Can the system review contracts accurately? Can it save time? Can it connect to the existing document-management environment? Can it reduce the cost of external support?

Those are important questions, but they are not sufficient.

As a leader considering implementation, I would also want to know whether the organisation can explain what information enters the system, where it is processed and how long it remains there. I would want to know whether users understand the difference between an AI-generated answer and a verified legal conclusion. I would ask how client restrictions are communicated, how outputs are reviewed and how the organisation would reconstruct an important decision if it were challenged months later.

Most importantly, I would ask who remains accountable.

My position is not anti-technology, nor is it based on the belief that every use of AI creates unacceptable legal exposure. In fact, I believe the opposite: AI can deliver significant value, but the organisation must remain in control of the information, the process and the final decision.

I have also used AI in preparing this article. It supported my research, helped me express my thoughts more clearly in English, challenged the consistency of some of my arguments and assisted me in checking factual statements against relevant sources. However, the conclusions, interpretations and opinions expressed here are my own. AI has supported and strengthened the way I have presented them, but it has not replaced my judgment or my responsibility for what I have written.

Legal AI is becoming more capable and more deeply integrated into the applications people already use. As the technology moves into Word, contract repositories, research databases and document-review platforms, employees may no longer experience it as a separate AI system. It will simply appear as another function in their normal workflow.

That makes governance more important, not less.

I want organisations to use AI, but I also want leaders to be able to demonstrate that the technology was selected carefully, configured appropriately, used by competent people and monitored throughout its lifecycle. Productivity alone cannot be the measure of successful implementation. A system that saves time while creating unmanaged confidentiality, data-protection or professional-liability exposure has not been implemented successfully.

The legal profession has never operated on the principle that a useful tool can replace professional responsibility. AI should not become the exception.

My concern is therefore not simply that legal AI may make mistakes. Lawyers, legal departments and business leaders have always faced that possibility. The real concern is that, without clear governance, an organisation can make the same mistake faster, apply it more widely and discover it only after the liability has become much harder to contain.

This article is also published on Medium as part of my public research and writing.