Legal Engineers Are Rewiring Law Firms for AI. Someone Has to Govern the Client Data They Move

For most of the last two decades, the people who knew how a law firm’s software worked and the people who knew the law sat on different floors. That arrangement is ending. Big firms, legal technology vendors and in-house legal departments are now hiring “legal engineers,” lawyers and legal professionals who understand both the law and AI tools well enough to rebuild how legal work gets done and to teach colleagues how to use those tools.

The New York Times DealBook section reported on October 31 that the job title, along with cousins like “knowledge and innovation counsel,” “innovation attorney” and “legal applications engineer,” is showing up at Latham & Watkins, Kirkland & Ellis, Gibson Dunn, Ropes & Gray, Microsoft, Nvidia, Harvey, Legora, OpenAI, Thomson Reuters and Google. Advertised salaries in the Times’ sample ran from $124,000 for an Nvidia legal operations engineer to $320,000 at the top of Harvey’s range.

Most coverage treats this as a career story. For privacy, security and risk teams it is also a data story. A legal engineer’s job is to move client documents, matter files and privileged analysis into AI-assisted workflows. Every workflow they design decides where confidential information goes, who can see it, how long it stays, and whether it can still be called privileged later. Firms and legal departments that hire for the role without giving it a governance mandate are building a fast pipeline with no valves.

What a legal engineer actually does

The role is still being defined, and it looks different depending on who is paying. According to the Times report:

• At AI vendors, legal engineers sometimes help build product, but more often they help customers fit the tool into their work, run trials and onboard staff. Harvey said its legal engineering team grew from 180 to 215 people in two weeks, all of them former practicing lawyers. Legora’s North America head of legal engineering said he had hired about 70 people in the past year, part of a global team of about 200 built since 2024.
• At law firms, the work is similar but wider: redesigning workflows across practice groups and getting lawyers to adopt them. Ropes & Gray hired the first of its five “practice innovation attorneys” at the end of last year. Latham told the paper it now has more than 100 people dedicated to AI initiatives, some building proprietary technology and some focused on AI-enabled workflows.
• In-house, companies such as Microsoft and Nvidia are hiring legal engineers to automate legal operations inside the business.

Stanford Law School’s AI initiative described the job to the Times as “law plus technology plus business process.” Its planned Legal Engineering Academy, a two-week executive program announced in September, is not yet taking applications, but the school says about 1,400 senior leaders have asked to be notified, roughly half from in-house teams. Recruiting platform Firm Prospects told the paper that almost 50 lawyers left the 200 largest U.S. firms for AI companies in the first half of 2026, 15 of them into jobs with legal engineering titles.

Very few schools offer formal training. Most legal engineers are former practitioners who taught themselves, people from legal operations or paralegal work, or consultants. The rare lawyer who was first a software engineer was described by one hiring manager as “the unicorn.” That matters for governance: the people wiring confidential data into AI systems are often learning the systems on the job.

Why this is a confidentiality problem before it is a productivity win

Legal work runs on information that the profession is required to protect. When a legal engineer builds an AI workflow for contract review, due diligence, litigation prep or regulatory research, the inputs are usually client confidential, often personal data, and sometimes privileged. Three bodies of rules now bear directly on how those workflows are designed.

The ethics rules already apply

The American Bar Association’s Formal Opinion 512, issued July 29, 2024, was the ABA’s first formal ethics guidance on generative AI. It did not create new rules. It explained how existing Model Rules apply: competence (Rule 1.1, including understanding the benefits and risks of the technology used), confidentiality (Rule 1.6), communication with clients (Rule 1.4), reasonable fees (Rule 1.5) and supervision of lawyers and nonlawyers (Rules 5.1 and 5.3). The opinion says a lawyer generally needs a client’s informed consent before inputting information relating to the representation into a self-learning generative AI tool, and that boilerplate consent buried in an engagement letter is not enough.

Model Rules are not binding until a state adopts them, and state bars have issued their own AI guidance that varies in detail. But the direction is consistent. The lawyer, not the tool and not the legal engineer, carries the duty. A workflow that quietly routes client files through a consumer chatbot is an ethics issue even if nothing leaks. Captain Compliance has covered how those duties translate into daily practice in our working guide to AI ethics and governance for lawyers.

Privilege can be lost in the prompt box

In February, Judge Jed Rakoff of the Southern District of New York decided what he called a question of first impression nationwide. In United States v. Heppner, a criminal defendant who knew he was a target had used Anthropic’s Claude to generate written analyses of possible defense strategies, then shared them with his lawyers. The court held on February 17, 2026 that the documents were protected by neither attorney-client privilege nor the work-product doctrine, as the ABA Litigation Section’s summary of the ruling explains. Claude is not a lawyer. The platform’s privacy policy allowed collection of inputs and outputs, training and disclosure to third parties, so there was no reasonable expectation of confidentiality. Sending the output to counsel afterward did not make it privileged. And because the defendant created the material on his own, not at counsel’s direction, work product did not apply.

Heppner is one trial-court ruling, and its facts involved a client acting alone on a public consumer tool. It left open whether the answer changes for an enterprise tool with contractual confidentiality, or for AI use directed by counsel. Those open questions are exactly the territory legal engineers work in. The way a firm configures its tools, the contract terms it negotiates with vendors, and whether AI use is documented as directed by counsel may all become evidence in a future privilege fight.

Data protection law treats firms as controllers

Law firms and legal departments process large amounts of personal data about clients, counterparties, employees, witnesses and litigants. Under the GDPR, U.S. state privacy laws and sector rules, putting that data into an AI system is a processing activity like any other. It needs a lawful basis or a permitted purpose, a vendor contract with the right terms, a retention rule and a record. European regulators have started to say so directly; Spain’s data protection authority, for example, has taken a strict view of law firms’ AI use, as we covered in our analysis of the AEPD’s position on law-firm AI.

Security pressure adds to the stakes. Law firms are attractive targets because they hold concentrated, sensitive data on many clients at once; CrowdStrike’s latest legal-sector findings are summarized in our breakdown of the 2026 law firm threat report. Every new AI integration is another set of credentials, connectors and stored outputs an attacker can go after.

Follow one matter file through an AI workflow

Consider a hypothetical: a mid-size firm’s new legal engineer builds a due diligence workflow for an acquisition. The data room holds 4,000 vendor contracts, employee records and customer lists. Here is what governance has to answer at each step, and what usually goes unanswered.

• Collection. Which documents are loaded into the AI tool, and does the set include personal data the review does not need, such as HR files or health information?
• Recipients. Is the tool hosted by the vendor, a cloud provider, or a subprocessor? Does the contract bar training on client data, and does it cover the model provider underneath the vendor’s interface?
• Purpose. Is the data used only for this matter, or are prompts, outputs and feedback reused to improve firm-wide templates or the vendor’s product?
• Retention. When the deal closes, are uploads, embeddings, chat logs and generated memos deleted on the same schedule as the underlying documents? A closed matter in the document management system is not proof that the AI tool’s copies are gone.
• Access controls. Can lawyers on other matters, or on the other side of an ethical wall, query the same workspace?
• Evidence. If the client, a regulator or a court asks later, can the firm show which tool processed what, under which terms, and on whose instruction?

A legal engineer is often the only person who knows the answers to all six questions. That makes the role a natural owner of the evidence, but not the right owner of the risk decision.

A seven-step governance mandate for legal engineering teams

Firms and legal departments do not need a new bureaucracy to manage this. They need to give the people building AI workflows a clear set of guardrails and a place to record their decisions. A practical starting point:

1. Inventory every AI tool and workflow. List the tool, vendor, model provider, hosting location, data categories allowed, and the practice groups using it. Include pilots and the “just testing” tools individual lawyers signed up for.
2. Classify the data each workflow touches. Separate public material, client confidential information, privileged communications and special-category personal data, and set which classes each tool may receive.
3. Read the vendor terms before the demo, not after. Confirm no training on client data, defined retention and deletion, subprocessor lists, breach notice timing and where data is stored. Our practical program for third-party AI tool risk walks through the questions.
4. Document counsel’s direction. After Heppner, record when AI use in a matter is directed by a supervising lawyer and for what litigation purpose. Whether that preserves work product is untested, but its absence will not help.
5. Get informed client consent where the rules require it. Explain specifically which tools are used and how, rather than relying on general engagement-letter language.
6. Set retention for AI artifacts. Treat prompts, outputs, embeddings and logs as records with their own deletion rules tied to the matter lifecycle and any legal hold.
7. Review outputs and test the controls. Human review of AI work is required by the competence and candor rules. Testing whether access restrictions and deletion settings actually work is a separate job, because a written policy is not proof the controls function.

Who signs off on each step is a structural question many organizations have not settled. It mirrors a wider debate about whether AI governance belongs to legal, privacy, security or a new function, which we explored in a look at who should own AI governance. For legal engineering, a workable answer is that the engineer builds and documents, the general counsel or risk partner approves, and privacy and security review the vendor and the data flows.

Where software helps, and where it does not

Tools can carry part of this load. Captain Compliance’s privacy compliance platform includes an AI governance module for inventorying AI systems, assigning guardrails and keeping evidence, alongside data mapping, assessments and vendor risk workflows that can record which AI vendors process which categories of client and personal data. That gives a firm one record of what was approved and why.

Software does not make the judgment calls. It cannot decide whether a particular client’s consent covers a new tool, whether an output is accurate, or whether a court will treat a workflow as privileged. And an inventory only reflects what people enter into it. The value of a legal engineer with a governance mandate is that the person building the workflow is also the person keeping the record current.

Is the job here to stay?

The people quoted by the Times disagree. Ropes & Gray’s head of practice AI, Howard Glazer, said the role “is absolutely growing” but asked whether it is “a transition position.” Mary O’Carroll, a longtime legal operations executive who now runs LegalEng Consulting Group, said she expects legal engineers to be needed to maintain systems as technology and work keep changing. Seyfarth Shaw’s chief data and AI officer, Byong Kim, said it was too early to say how it will play out.

From a governance view the title matters less than the function. Someone has to understand both the legal duties attached to client information and the technical paths that information now travels. If that knowledge sits with a single legal engineer who leaves for a vendor, as dozens of Big Law lawyers did in the first half of the year, the firm loses its map of its own AI workflows. Documentation is how that knowledge survives turnover.

Frequently asked questions

Does a legal engineer need a law degree?

Not always. Many are former practicing lawyers, and Harvey says all of its legal engineers are. But the Times reported that others come from legal operations, paralegal work or consulting. Where a legal engineer is not a lawyer, Model Rule 5.3 on supervising nonlawyer assistants becomes relevant to how the firm oversees their work.

Does using an enterprise AI tool protect privilege?

No court has said so yet. Heppner involved a public consumer tool used by a client acting alone, and the court relied in part on the platform’s privacy policy permitting training and third-party disclosure. Enterprise terms that bar training and restrict disclosure may help a confidentiality argument, but they are not a guarantee, and the question remains open.

Is ABA Formal Opinion 512 binding?

ABA formal opinions are advisory. They interpret the Model Rules, which become binding only as adopted by each state, and state bars may issue different or more specific guidance. Lawyers should check the rules and opinions in every jurisdiction where they are licensed.

What should a firm ask an AI vendor about client data?

At minimum: whether inputs or outputs are used for training, how long prompts and files are retained, who the subprocessors and model providers are, where data is stored, how deletion is confirmed and how quickly the vendor reports a breach. Ask for the answers in the contract, not the sales deck.

Who should own AI governance in a law firm?

Usually a shared model works best: legal engineers build and document workflows, a designated partner or general counsel approves risk decisions, and privacy and security teams review vendors and data flows. What matters is that ownership is written down.

The record that outlasts the engineer

Law firms spent years learning that a document management system is only as good as the matter numbers people attach to files. AI workflows are heading for the same lesson at higher speed. The firms hiring legal engineers by the dozen are betting that the role will make lawyers faster. The ones that also ask those engineers to keep a defensible record of every tool, data flow and consent may find that the more valuable product is not the workflow, but the proof of how it was built when a client, regulator or judge comes asking.