What Cross-Border M&A Teaches About the Limits of Legal AI
For two years, the legal technology conversation has turned on a single premise, that artificial intelligence can read a contract. It can.
In due diligence, where the review of hundreds or thousands of agreements once set the cost and the calendar, a well-pointed model now does in an afternoon what a team of associates needed two weeks to finish. The most labor-intensive phase of diligence has become, for practical purposes, nearly free.
The temptation is to treat that as the whole victory. In my experience it is not. Compressing the middle of a process does not eliminate the work that surrounds it. It moves that work to the two ends, and those ends are where legal AI is least useful. Cross-border technology M&A makes the point as well as any practice I know.
A Process That Resembles a Software Build
It helps to look at diligence the way an engineer looks at building software. Software moves through three phases, architecture, engineering and deployment. Due diligence moves through three of its own, scoping, issue identification and remediation. In both, the middle phase was historically the constraint. Engineering consumed the calendar, and issue identification consumed the associates who read every contract and flagged every clause.
Coding assistants compressed the first of those phases. Legal AI is compressing the second. The consequence is the same in each case. When the middle of a process is no longer the bottleneck, the work that determines the outcome shifts to the front and the back. Engineers spend their judgment on architecture and deployment. Diligence lawyers should spend theirs on scoping, which decides what to examine, and on remediation, which decides what the findings mean. Those phases turn on judgment rather than volume, and judgment is the thing the model does not provide.
The Familiar Problem of Garbage In, Garbage Out
A fast engine will carry out a poor instruction faithfully. Point legal AI at a loosely scoped review and it returns 500 findings, of which perhaps 50 bear on the transaction. The fee a team thought it had saved comes back as the labor of sorting the pile, and a deal-defining issue can sit in that pile receiving the same brief attention as the boilerplate around it.
This is the part of the story that tends to go unmentioned. The model surfaces everything, which means a person still has to decide what everything means. A tool that makes identification cheap raises the value of the scoping that comes before it and the remediation that follows. It does not lower it.
Where the Judgment Lives, In Practice
The point is clearest in the diligence of a modern software company, which is global whether or not anyone planned it that way. Code is written wherever the engineers live. Intellectual property is created across jurisdictions. A target with modest revenue can carry employees, data, and tax exposure in a dozen countries without a single foreign office. The central scoping question becomes geographic. Which jurisdictions deserve real scrutiny, and which do not.
That question rests on a distinction a model will not draw reliably on its own, the difference between a few foreign employees and a genuine foreign operating footprint. A handful of remote engineers in three countries presents real but bounded risk, usually handled with focused local employment advice. Running a full review in each of those countries spends money to disprove a risk that was never material. A company that genuinely operates abroad is a different matter, and two questions reward careful attention. The first is whether the company owns its intellectual property at all. Many jurisdictions, unlike the United States, do not vest an employee’s work product in the employer by default, and if the chain of title fails under local law, the buyer may not own the asset it is paying for. The second is tax. The assumption that a structuring review has answered the tax question can obscure whether a dispersed workforce has created a taxable presence in a country where nothing was ever filed.
A model can flag an assignment clause. It will not tell you that the clause is unenforceable under Portuguese law, or that the missing clause for a contractor in Bangalore is the one that matters. Recognizing that is scoping and remediation, the work at the two ends, and it remains human.
The Discipline the Tools Require
The lesson reaches past M&A, and it is the one worth carrying into any workflow that AI now touches. An accelerated middle pays off only when the ends are handled with discipline. In diligence that means scoping tightly enough that the machine looks in the right places, and remediating carefully enough to separate the exposures that threaten a deal from the ones that only look alarming on a printout. It also means scoping for the parties who rely on the work after signing. The representations and warranties insurer decides what it needs to see before it will cover a risk, and it penalizes too little diligence and too much in equal measure. The lenders have their own requirements for guarantees, share pledges, and audited financial statements, and those requirements surface at financing rather than at signing unless counsel has anticipated them.
Legal AI can read every document in a data room faster than any team a firm could assemble. It cannot tell a buyer which jurisdiction holds a real operation and which holds a single laptop, and it cannot tell the buyer what the deal’s eventual insurers and lenders will require. That judgment is the work the technology has made more valuable rather than less. The firms that see this will use AI to clear the middle and put the time it frees back into judgment at the ends. The firms that treat a faster review as a finished one will reach the wrong answer sooner than before.
Reprinted with permission from the June 4, 2026 edition of the Legaltech news ©2026 ALM Global Properties, LLC. All rights reserved. Further duplication without permission is prohibited, contact 877-256-2472 or [email protected].
